Nothing is ever as good as it first seems and nothing is ever as bad as it first seems.

— A best memory paraphrase of advice given to me by Vice Admiral Joe Dyer, former chief test pilot of the US Navy and former Commander of NAVAIR.

[You can follow me on social media: @rodneyabrooks.bsky.social and see my publications etc., at https://people.csail.mit.edu/brooks]

Table of contents

Introduction
   What I Nearly Got Wrong
   What Has Surprised Me, And That I Missed 8 Years Ago
   My Color Scheme and Past Analysis

My New Predictions
   Quantum Computers
   Self Driving Cars
   Humanoid Robots
   Neural Computation
   LLMs

Self Driving Cars
   A Brief Recap of what "Self Driving" Cars Means and Meant
   My Own Experiences with Waymo in 2025
   Self Driving Taxi Services
     __Cruise
     __Tesla
     __Waymo
     __Zoox
   Electric Cars
   Flying Cars

Robotics, AI, and Machine Learning
   Capabilities and Competences
   World Models
   Situatedness vs Embodiment
   Dexterous Hands

Human Space Flight
   Orbital Crewed Flights
   Suborbital Crewed Flights
   Boeing's Starliner
   SpaceX Falcon 9
   NASA, Artemis, and Returning to the Moon
   SpaceX Starship
   Blue Origin Gets to Orbit
   New Space Stations

Addendum

Introduction

This is my eighth annual update on how my dated predictions from January 1^st, 2018 concerning (1) self driving cars, (2) robotics, AI , and machine learning, and (3) human space travel, have held up. I promised then to review them at the start of the year every year until 2050 (right after my 95^th birthday), thirty two years in total. The idea was to hold myself accountable for those predictions. How right or wrong was I?

The summary is that my predictions held up pretty well, though overall I was a little too optimistic. That is a little ironic, as I think that many people who read my predictions back on  January 1^st, 2018 thought that I was very pessimistic compared to the then zeitgeist. I prefer to think of myself as being a realist.

And did I see LLMs coming? No and yes. Yes, I did say that something new and big that everyone accepted as the new and big thing in AI would come along no earlier than 2023, and that the key paper for its success had already been written by before I made my first predictions. And indeed LLMs were generally accepted as the next big thing in 2023 (I was lucky on that date), and the key paper, Attention Is All You Need, was indeed already written, and had first appeared in June of 2017. I wrote about this extensively in last year’s scorecard. But no, I had no idea it would be LLMs at the time of my correct prediction that something big would appear. And that lack of specificity on the details of exactly what will be invented and when is the case with all my predictions from the first day of 2018.

I did not claim to be clairvoyant about exactly what would happen, rather I was making predictions about the speed of new research ideas, the speed of hype generation, the speed of large scale deployments of new technologies, and the speed of fundamental changes propagating through the world’s economy. Those speeds are very different and driven by very different realities. I think that many people get confused by that and make the mistake of jumping between those domains of reality, thinking all the speeds will be the same.  In my case my estimates of those speeds are informed by watching AI and robotics professionally, for 42 years at the time of my predictions. I became a graduate student in Artificial Intelligence in January of 1976, just shy of 20 years after the initial public outing of the term Artificial Intelligence at the summer workshop in 1956 at Dartmouth. And now as of today I have been in that field for 50 years.

I promised to track my predictions made eight years ago today for 32 years. So I am one quarter of the way there. But the density of specific years of events or marking percentages of adoption that I predicted start to fall off right around now.

Sometime during 2026 I will bundle up all my comments over the eight years specifically mentioning years that have now passed, and put them in an archival mid-year post. Then I will get rid of the three big long tables that dominate the body of this annual post, and have short updates on the sparse dates for the next 24 years.

I will continue to summarize what has happened in self-driving cars generally, including electrification progress and the forever promised flying cars, along with AI and robotics, and human space flight. But early in 2025 I made five new predictions for the coming ten years, without specific dates, but which summarize what I think will happen.  I will track these predictions too.

What I Nearly Got Wrong

The day before my original prediction post in 2018 the price of Bitcoin had opened at $12,897.70 and topped out at $14,377.40 and 2017 had been the first year it had ever traded at over $1,000. The price seemed insane to me as Bitcoin wasn’t being used for the task for which it had been designed. The price seemed to me then, and now, to be purely about speculation. I almost predicted when it would be priced at $200, on the way down. But, fortunately, I checked myself as I realized that the then current state of the market made no sense to me and so any future state may not either. Besides, I had no experience or expertise in crypto pricing. So I left that prediction out. I had no basis to make a prediction. That was a wise decision, and I revisit that reasoning as I make new predictions now, and implore myself to only make predictions in fields where I know something.

What Has Surprised Me, And That I Missed 8 Years Ago

I made some predictions about the future of SpaceX although I didn’t always label them as being about SpaceX. A number of my predictions were in response to pronouncements by the CEO of SpaceX. My predictions were much more measured and some might say even pessimistic. Those predictions so far have turned out to be more optimistic than how reality has unfolded.

I had made no specific predictions about Falcon 9, though I did make predictions about the subsequent SpaceX launch family, now called Starship, but then known as BFR, which eight years later has not gotten into orbit.

In the meantime SpaceX has scaled the Falcon 9 launch rate at a phenomenal speed, and the magnitude of the growth is very surprising.

Eight years ago, Falcon 9 had been launched 46 times, all successful, over the previous eight years, and it had recently had a long run of successful landings of the booster whenever attempted. At that time five launches had been on a previously used booster, but there had been no attempts to launch Falcon Heavy with its three boosters strapped together.

Now we are eight years on from those first eight years of Falcon 9 launches. The scale and success rate of the launches has made each individual launch an unremarkable event, with humans being launched a handful of times per year. Now the Falcon 9 score card stands at 582 launches with only one failed booster, and there have been 11 launches of the three booster Falcon Heavy, all successful. That is a sustained growth rate of 38% year over year for eight years. And that it is a very high sustained deployment growth rate for any complex technology.

There is no other modern rocket with such a volume of launches that comes even close to the Falcon 9 record.  And I certainly did not foresee this volume of launches. About half the launches have had SpaceX itself as the customer, starting in February 2018, launching an enormous satellite constellation (about two thirds of all satellites ever orbited) to support Starlink bringing internet to everywhere on the surface of Earth.

But… there is one historical rocket, a suborbital one which has a much higher record of use than Falcon 9 over a much briefer period. The German V-2 was the first rocket to fly above the atmosphere and the first ballistic missile to be used to deliver bombs. It was fueled with ethanol and liquid oxygen, and was steered by an analog computer that also received inputs from radio guide signals–it was the first operational liquid fueled rocket. It was developed in Germany in the early 1940’s and after more than a thousand test launches was first put into operation on September 7^th, 1944, landing a bomb on Paris less than two weeks after the Allied liberation of that city. In the remaining 8 months of the war 3,172 armed V-2 rockets were launched at targets in five countries — 1,358 were targeted at London alone.

My Color Scheme and Past Analysis

The acronyms I used for predictions in my original post were as follows.

NET year means it will not happen before that year (No Earlier Than)
BY year means I predict that it will happen by that year.
NIML, Not In My Lifetime, i.e., not before 2050.

As time passes mentioned years I color then as accurate, too pessimistic, or too optimistic.

Last year I added hemming and hawing. This is for when something looks just like what I said would take a lot longer has happened, but the underlying achievement is not what everyone expected, and is not what was delivered. This is mostly for things that were talked about as being likely to happen with no human intervention and it now appears to happen that way, but in reality there are humans in the loop that the companies never disclose. So the technology that was promised to be delivered hasn’t actually been delivered but everyone thinks it has been.

When I quote myself I do so in orange, and when I quote others I do so in blue.

I have not changed any of the text of the first three columns of the prediction tables since their publication on the first day of 2018. I only change the text in the fourth column to say what actually happened.  This meant that by four years ago that fourth column was getting very long and skinny, so I removed them and started with fresh comments two years ago. I have kept the last two year’s comments and added new ones, with yellow backgrounds, for this year, removing the yellow backgrounds from 2025 comments that were there last year. If you want to see the previous five years of comments you can go back to  the 2023 scorecard.

My NEW PREDICTIONS

On March 26^th I skeeted out five technology predictions, talking about developments over the next ten years through January 1st, 2036. Three weeks later I included them in a blog post. Here they are again.

1. Quantum Computers. The successful ones will emulate physical systems directly for specialized classes of problems rather than translating conventional general computation into quantum hardware. Think of them as 21st century analog computers. Impact will be on materials and physics computations.

2. Self Driving Cars. In the US the players that will determine whether self driving cars are successful or abandoned are #1 Waymo (Google) and #2 Zoox (Amazon). No one else matters. The key metric will be human intervention rate as that will determine profitability.

3. Humanoid Robots. Deployable dexterity will remain pathetic compared to human hands beyond 2036. Without new types of mechanical systems walking humanoids will remain too unsafe to be in close proximity to real humans.

4. Neural Computation. There will be small and impactful academic forays into neuralish systems that are well beyond the linear threshold systems, developed by 1960, that are the foundation of recent successes. Clear winners will not yet emerge by 2036 but there will be multiple candidates.

5. LLMs. LLMs that can explain which data led to what outputs will be key to non annoying/dangerous/stupid deployments. They will be surrounded by lots of mechanism to keep them boxed in, and those mechanisms, not yet invented for most applications, will be where the arms races occur.

These five predictions are specifically about what will happen in these five fields during the ten years from 2026 through 2035, inclusive. They are not saying when particular things will happen, rather they are saying whether or not  certain things will happen in that decade. I will do my initial analysis of these five new predictions immediately below. For the next ten years I will expand on each of these reviews in this annual scorecard, along with reviews of my earlier predictions. The ten years for these predictions are up on January 1^st, 2036. I will have just turned 81 years old then, so let’s see if I am still coherent enough to do this.

Quantum Computers

The successful ones will emulate physical systems directly for specialized classes of problems rather than translating conventional general computation into quantum hardware. Think of them as 21st century analog computers. Impact will be on materials and physics computations.

The original excitement about quantum computers was stimulated by a paper by Peter Shor in 1994 which gave a digital quantum algorithm to factor large integers much faster than a conventional digital computer. Factoring integers is often referred to as “the IFP” for the integer factorization problem.

So what? The excitement around this was based on how modern cryptography, which provides our basic security for on-line commerce, works under the hood.

Much of the internet’s security is based on it being hard to factor a large number. For instance in the RSA algorithm Alice tells everyone a large number (in different practical versions it has 1024, 2048, or 4096 bits) for which she knows its prime factors. But she tells people only the number not its factors. In fact she chose that number by multiplying together some very large prime numbers — very large prime numbers are fairly easy to generate (using the Miller-Rabin test). Anyone, usually known as Bob, can then use that number to encrypt a message intended for Alice. No one, neither Tom, Dick, nor Harry, can decrypt that message unless they can find the prime factors of Alice’s public number. But Alice knows them and can read the message intended only for her eyes.

So… if you could find prime factors of large numbers easily then the backbone of digital security would be broken. Much excitement!

Shor produced his algorithm in 1994. By the year 2001 a group at IBM had managed to find the prime factors of the number 15 using a digital quantum computer as published in Nature. All the prime factors. Both 3 and 5. Notice that 15 has only four bits, which is a lot smaller than the number of bits used in commercial RSA implementations, namely 1024, 2048, or 4096.

Surely things got better fast.  By late 2024 the biggest numbers that had been factored by an actual digital quantum computer had 35 bits which allows for numbers no bigger than 34,359,738,367. That is way smaller than the size of the smallest numbers used in RSA applications. Nevertheless it does represent 31 doublings in magnitude of numbers factored in 23 years, so progress has been quite exponential. But it could take another 500 years of that particular version of exponential growth rate to get to conquering today’s smallest version of RSA digital security.

In the same report the authors say that a conventional, but very large computer (2,000 GPUs along with a JUWELS booster, which itself has 936 compute nodes each consisting of four NVIDIA A100 Tensor Core GPUs themselves each hosted by 48 dual threaded AMD EPYC Rome cores–that is quite a box of computing) simulating a quantum computer running Shor’s algorithm had factored a 39 bit number finding that 549,755,813,701 = 712,321 × 771,781, the product of two 20 bit prime numbers. That was its limit. Nevertheless, an actual digital quantum computer can still be outclassed by one simulated on conventional digital hardware.

The other early big excitement for digital quantum computers was Grover’s search algorithm, but work on that has not been as successful as for Shor’s IFP solution.

Digital quantum computation nirvana has not yet been demonstrated.

Digital quantum computers work a little like regular digital computers in that there is a control mechanism which drives the computer through a series of discrete steps. But today’s digital quantum computers suffer from accumulating errors in quantum bits. Shor’s algorithm assumes no such errors. There are techniques for correcting those errors but they slow things down and cause other problems. One way that digital quantum computers may get better is if new methods of error correction emerge. I am doubtful that something new will emerge, get fully tested, and then make it into production at scale all within the next ten years. So we may not see a quantum (ahem) leap in performance of quantum digital computers in the next decade.

Analog quantum computers are another matter. They are not switched, but instead are configured to directly simulate some physical system and the quantum evolution and interactions of components of that system. They are an embodied quantum model of that system. And they are ideally suited to solving these sorts of problems and cannot be emulated by conventional digital systems as they can be in the 39 bit number case above.

I find people working on quantum computers are often a little squirrelly about whether their computer acts more like a digital or analog computer, as they like to say they are “quantum” only.  The winners over the next 10 years will be ones solving real problems in materials science and other aspects of chemistry and physics.

Self Driving Cars

In the US the players that will determine whether self driving cars are successful or abandoned are #1 Waymo (Google) and #2 Zoox (Amazon). No one else matters. The key metric will be human intervention rate as that will determine profitability.

Originally the term “self driving car” was about any sort of car that could operate without a driver on board, and without a remote driver offering control inputs. Originally they were envisioned as an option for privately owned vehicles used by individuals, a family car where no person needed to drive, but simply communicated to the car where it should take them.

That conception is no longer what people think of when self driving cars are mentioned. Self driving cars today refer to taxi-services that feel like Uber or Lyft, but for which there is not a  human driver, just paying passengers.

In the US the companies that have led in this endeavor have changed over time.

The first leader was Cruise, owned by GM. They were the first to have a regular service in the downtown area of a major city (San Francisco), and then in a number of other cities, where there was an app that anyone could download and then use their service. They were not entirely forthcoming with operational and safety problems, including when they dragged a person, who had just been hit by a conventionally driven car, for tens of feet under one of their vehicles. GM suspended operations in late 2023 and completely disbanded it in December 2024.

Since then Waymo (owned by Google) has been the indisputable leading deployed service.

Zoox (owned by Amazon) has been a very distant, but operational, second place.

Tesla (owned by Tesla) has put on a facade of being operational, but it is not operational in the sense of the other two services, and faces regulatory headwinds that both Waymo and Zoox have long been able to satisfy. They are not on a path to becoming a real service.

See my traditional section on self driving cars below, as it explains in great detail the rationale for these evaluations. In short, Waymo looks to have a shot at succeeding and it is unlikely they will lose first place in this race. Zoox may also cross the finish line, and it is very unlikely that anyone will beat them.  So if both of Waymo and Zoox fail, for whatever reason, the whole endeavor will grind to a halt in the US.

BUT…

But what might go wrong that makes one of these companies fail. We got a little insight into that in the last two weeks of 2025.

On Saturday December 20^th of 2025 there was an extended power outage in San Francisco that started small in the late morning but by nightfall had spread to large swaths of the city.  And lots and lots of normally busy intersections were by that time blocked by many stationary Waymos.

Traffic regulations in San Francisco are that when there is an intersection which has traffic lights that are all dark, that intersection should be treated as though it has stop signs at every entrance. Human drivers who don’t know the actual regulation tend to fall back to that behavior in any case.

It seemed that Waymos were waiting indefinitely for green lights that never came, and at intersections through which many Waymos were routed there were soon enough waiting Waymos that the intersections were blocked.  Three days later, on December 23^rd, Waymo issued an explanation on their blog site, which includes the following:

Navigating an event of this magnitude presented a unique challenge for autonomous technology. While the Waymo Driver is designed to handle dark traffic signals as four-way stops, it may occasionally request a confirmation check to ensure it makes the safest choice. While we successfully traversed more than 7,000 dark signals on Saturday, the outage created a concentrated spike in these requests. This created a backlog that, in some cases, led to response delays contributing to congestion on already-overwhelmed streets.

We established these confirmation protocols out of an abundance of caution during our early deployment, and we are now refining them to match our current scale. While this strategy was effective during smaller outages, we are now implementing fleet-wide updates that provide the Driver with specific power outage context, allowing it to navigate more decisively.

As the outage persisted and City officials urged residents to stay off the streets to prioritize first responders, we temporarily paused our service in the area. We directed our fleet to pull over and park appropriately so we could return vehicles to our depots in waves. This ensured we did not further add to the congestion or obstruct emergency vehicles during the peak of the recovery effort.

The key phrase is that Waymos “request a confirmation check” at dark signals. This means that the cars were asking for a human to look at images from their cameras and manually tell them how to behave. With 7,000 dark signals and perhaps a 1,000 vehicles on the road, Waymo clearly did not have enough humans on duty to handle the volume of requests that were coming in. Waymo does not disclose whether any human noticed a rise in these incidents early in the day and more human staff were called in, or whether they simply did not have enough employees to make handling them all possible.

At a deeper level it looks like they had a debugging feature in their code, and not enough people to supply real time support to handle the implications of that debugging feature. And it looks like Waymo is going to remove that debugging safety feature as a way of solving the problem.  This is not an uncommon sort of engineering failure during early testing. Normally one would hope that the need for that debugging feature had been resolved before large scale deployment.

But, who are these human staff?  Besides those in Waymo control centers, it turns out there is a gig-work operation with an app named Honk (the headline of the story is When robot taxis get stuck, a secret army of humans comes to the rescue) whereby Waymo pays people around $20 to do minor fixups to stuck Waymos by, for instance, going and physically closing a door that q customer left open. Tow truck operators use the same app to find Waymos that need towing because of some more serious problem. It is not clear whether it was a shortage of those gig workers, or a shortage of people in the Waymo remote operations center that caused the large scale failures.  But it is worth noting that current generation Waymos need a lot of human help to operate as they do, from people in the remote operations center to intervene and provide human advice for when something goes wrong, to Honk gig-workers scampering around the city physically fixed problems with the vehicles, to people to clean the cars and plug them in to recharge when they return to their home base.

For human operated ride services, traditional taxi companies or gig services such as Uber and Lyft, do not need these external services. There is a human with the car at all times who takes care of these things.

The large scale failure on the 20^th did get people riled up about these robots causing large scale traffic snarls, and made them wonder about whether the same thing will happen when the next big earthquake hits San Francisco. Will the human support worker strategy be stymied by both other infrastructure failures (e.g., the cellular network necessary for Honk workers to communicate) or the self preservation needs of the human workers themselves?

The Waymo blog post revealed another piece of strategy. This is one of three things they said that they would do to alleviate the problems:

Expanding our first responder engagement: To date, we’ve trained more than 25,000 first responders in the U.S. and around the world on how to interact with Waymo. As we discover learnings from this and other widespread events, we’ll continue updating our first responder training.

The idea is to add more responsibility to police and fire fighters to fix the inadequacies of the partial-only autonomy strategy for Waymo’s business model. Those same first responders will have more than enough on their plates during any natural disasters.

Will it become a political issue where the self-driving taxi companies are taxed enough to provide more first responders? Will those costs ruin their business model? Will residents just get so angry that they take political action to shut down such ride services?

Humanoid Robots

Deployable dexterity will remain pathetic compared to human hands beyond 2036. Without new types of mechanical systems walking humanoids will remain too unsafe to be in close proximity to real humans.

Despite this prediction it is worth noting that there is a long distance between current deployed dexterity and dexterity that is still pathetic. In the next ten years deployable dexterity may improve markedly, but not in the way the current hype for humanoid robots suggests.  I talk about his below in my annual section scoring my 2018 predictions on robotics, AI, and machine learning, in a section titled Dexterous Hands.

Towards the end of 2025 I published a long blog post summarizing the status of, and problems remaining for humanoid robots.

I started building humanoid robots in my research group at MIT in 1992. My previous company, Rethink Robotics, founded in 2008, delivered thousands of upper body Baxter and Sawyer humanoid robots (built in the US) to factories between 2012 and 2018.  At the top of this blog page you can see a whole row of Baxter robots in China. A Sawyer robot that had operated in a factory in Oregon just got shut down in late 2025 with 35,236 hours on its operations clock. You can still find many of Rethink’s humanoids in use in teaching and research labs around the world. Here is the cover of Science Robotics from November 2025,

showing a Sawyer used in the research for  this article out of Imperial College, London.

Here is a slide from a 1998 powerpoint deck that I was using in my talks, six years after my graduate students and I had started building our first humanoid robot, Cog.

It is pretty much the sales pitch that today’s humanoid companies use.  You are seeing here my version from almost twenty eight years ago.

I point this out to demonstrate that I am not at all new to humanoid robotics and have worked on them for decades in both academia and in producing and selling humanoid robots that were deployed at scale (which no one else has done) doing real work.

My blog post from September, details why the current learning based approaches to getting dexterous manipulation will not get there anytime soon. I argue that the players are (a) collecting the wrong data and (b) trying to learn the wrong thing. I also give an argument (c) for why learning might not be the right approach. My argument for (c) may not hold up, but I am confident that I am right on both (a) and (b), at least for the next ten years.

I also outline in that blog post why the current (and indeed pretty much the only, for the last forty years) method of building bipeds and controlling them will remain unsafe for humans to be nearby. I pointed out that the danger is roughly cubicly proportional to the weight of the robot. Many humanoid robot manufacturers are introducing lightweight robots, so I think they have come to the same conclusion. But the side effect is that the robots can not carry much payload, and certainly can’t provide physical support to elderly humans, which is a thing that human carers do constantly — these small robots are just not strong enough. And elder care and in home care is one of the main arguments for having human shaped robots, adapted to the messy living environments of actual humans.

Given that careful analysis from September I do not share the hype that surrounds humanoid robotics today. Some of it is downright delusional across many different levels.

To believe the promises of many CEOs of humanoid companies you have to accept the following conjunction.

1. Their robots have not demonstrated any practical work (I don’t count dancing in a static environment doing exactly the same set of moves each time as practical work).
2. The demonstrated grasping, usually just a pinch grasp,  in the videos they show is at a rate which is painfully slow and not something that will be useful in practice.
3. They claim that their robots will learn human-like dexterity but they have not shown any videos of multi-fingered dexterity where humans can and do grasp things that are unseen, and grasp and simultaneously manipulate multiple small objects with one hand. And no demonstrations of using the body with the hands which is how humans routinely carry many small things or one or two heavy things.
4. They show videos of non tele-operated manipulation, but all in person demonstrations of manipulation are tele-operated.
5. Their current plans for robots working in customer homes all involve a remote person tele-operating the robot.
6. Their robots are currently unsafe for humans to be close to when they are walking.
7. Their robots have no recovery from falling and need human intervention to get back up.
8. Their robots have a battery life measured in minutes rather than hours.
9. Their robots cannot currently recharge themselves.
10. Unlike human carers for the elderly, humanoids are not able to provide any physical assistance to people that provides stabilizing support for the person walking, getting into and out of bed physical assistance, getting on to and off of a toilet, physical assistance, or indeed any touch based assistance at all.
11. The CEOs claim that there robots will be able to do everything, or many things, or a lot of things, that a human can do in just a few short years. They currently do none.
12. The CEOs claim a rate of adoption of these humanoid robots into homes and industries at a rate that is multiple orders of magnitude faster than any other technology in human history, including mainframe computers, and home computers and the mobile phones, and the internet. Many orders of magnitude faster. Here is a CEO of a humanoid robot company saying that they will be in 10% of US households by 2030. Absolutely no technology (even without the problems above) has ever come close to scaling at that rate.

The declarations being made about humanoid robots are just not plausible.

We’ll see what actually happens over the next ten years, but it does seem that the fever is starting to crack at the edges. Here are two news stories from the last few days of 2025.

From The Information on December 22^nd there is a story about how humanoid robot companies are wrestling with safety standards. All industrial and warehouse robots, whether stationary of mobile have a big red safety stop button, in order to comply with regulatory safety standards. The button cuts the power to the motors. But cutting power to the motors of a balancing robot might make them fall over and cause more danger and damage to people nearby.  For the upper torso humanoid robots Baxter and Sawyer from my company Rethink Robotics we too had a safety stop button that cut power to all the motors in the arms. It was a collaborative robot and often a person, or part of their limbs or body could be under an arm and it would have been dangerous for the arms to fall quickly on cutoff of power. To counter this we developed a unique circuit that required no active power, which made it so that the back current generated by a motor when powered off acted as a very strong brake. Perhaps there are similar possible solutions for humanoid robots and falling, but they need to be invented yet.

On December 25^th the Wall Street Journal had a story headlined “Even the Companies Making Humanoid Robots Think They’re Overhyped”, with a lede of “Despite billions in investment, startups say their androids mostly aren’t useful for industrial or domestic work yet”. Here are the first two paragraphs of the story:

Billions of dollars are flowing into humanoid robot startups, as investors bet that the industry will soon put humanlike machines in warehouses, factories and our living rooms.

Many leaders of those companies would like to temper those expectations. For all the recent advances in the field, humanoid robots, they say, have been overhyped and face daunting technical challenges before they move from science experiments to a replacement for human workers.

And then they go on to quote various company leaders:

“We’ve been trying to figure out how do we not just make a humanoid robot, but also make a humanoid robot that does useful work,” said Pras Velagapudi, chief technology officer at Agility Robotics.

Then talking about a recent humanoid robotics industry event the story says:

On stage at the summit, one startup founder after another sought to tamp down the hype around humanoid robots.

“There’s a lot of great technological work happening, a lot of great talent working on these, but they are not yet well defined products,” said Kaan Dogrusoz, a former Apple engineer and CEO of Weave Robotics.

Today’s humanoid robots are the right idea, but the technology isn’t up to the premise, Dogrusoz said. He compared it to Apple’s most infamous product failure, the Newton hand-held computer.

There are more quotes from other company leaders all pointing out the difficulties in making real products that do useful work. Reality seems to be setting in as promised delivery dates come and go by.

Meanwhile here is what I said at the end of my September blog post about humanoid robots and teaching them dexterity.  I am not at all negative about a great future for robots, and in the nearish term. It is just that I completely disagree with the hype arguing that building robots with humanoid form magically will make robots useful and deployable. These particular paragraphs followed where I had described there, as I do again in this blog post, how the meaning of self driving cars has drifted over time.

Following that pattern, what it means to be a humanoid robot will change over time.

Before too long (and we already start to see this) humanoid robots will get wheels for feet, at first two, and later maybe more, with nothing that any longer really resembles human legs in gross form.  But they will still be called humanoid robots.

Then there will be versions which variously have one, two, and three arms. Some of those arms will have five fingered hands, but a lot will have two fingered parallel jaw grippers. Some may have suction cups. But they will still be called humanoid robots.

Then there will be versions which have a lot of sensors that are not passive cameras, and so they will have eyes that see with active light, or in non-human frequency ranges, and they may have eyes in their hands, and even eyes looking down from near their crotch to see the ground so that they can locomote better over uneven surfaces. But they will still be called humanoid robots.

There will be many, many robots with different forms for different specialized jobs that humans can do. But they will all still be called humanoid robots.

As with self driving cars, most of the early players in humanoid robots, will quietly shut up shop and disappear. Those that remain will pivot and redefine what they are doing, without renaming it, to something more achievable and with, finally, plausible business cases. The world will slowly shift, but never fast enough to need a change of name from humanoid robots. But make no mistake, the successful humanoid robots of tomorrow will be very different from those being hyped today.

Neural Computation

There will be small and impactful academic forays into neuralish systems that are well beyond the linear threshold systems, developed by 1960, that are the foundation of recent successes. Clear winners will not yet emerge by 2036 but there will be multiple candidates.

Current machine learning techniques are largely based on having millions, and more recently tens (to hundreds?) of billions, of linear threshold units. They look like this.

Each of these units have a fixed number of inputs, where some numerical value comes in, and it is multiplied by a weight, usually a floating point number, and the results of all of the multiplications are summed, along with an adjustable threshold , which is usually negative, and then the sum goes through some sort of squishing function to produce a number between zero and one, or in this case minus one and plus one, as the output. In this diagram, which, by the way is taken from Bernie Widrow’s technical report from 1960, the output value is either minus one or plus one, but in modern systems it is often a number from anywhere in that, or another, continuous interval.

This was based on previous work, including that of Warren McCulloch and Walter Pitts’ 1943 formal model of a neuron, Marvin Minsky’s 1954 Ph.D. dissertation on using reinforcement for learning in a machine based on model neurons, and Frank Rosenblatt’s 1957 use of weights (see page 10) in an analog implementation of a neural model.

These are what current learning mechanisms have at their core. These! A model of  biological neurons that was developed in a brief moment of time from 83 to 65 years ago.  We use these today.  They are extraordinarily primitive models of neurons compared to what neuroscience has learned in the subsequent sixty five years.

Since the 1960s higher levels of organization have been wrapped around these units. In 1979 Kunihiko Fukushima published (at the International Joint Conference on Artificial Intelligence, IJCAI 1979, Tokyo — coincidentally the first place where I published in an international venue) his first English language description of convolutional neural networks (CNNs), which allowed for position invariant recognition of shapes (in his case, hand written digits), without having to learn about those shapes in every position within images.

Then came backpropagation, a method where a network can be told the correct output it should have produced, and by propagating the error backwards through the derivative of the quantizer in the diagram above (note that the quantizer shown there is not differentiable–a continuous differentiable quantizer function is needed to make the algorithm work), a network can be trained on examples of what it should produce. The details of this algorithm, are rooted in the chain rule of Gottfried Leibniz in 1676 through a series of modern workers from around 1970 through about 1982. Frank Rosenblatt (see above) had talked about a “back-propagating error correction” in 1962, but did not know how to implement it.

In any case, the linear threshold neurons, CNNs, and backpropagation are the basis of modern neural networks. After an additional 30 years of slow but steady progress they burst upon the scene as deep learning, and unexpectedly crushed many other approaches to computer vision — the research field of getting computers to interpret the contents of an image. Note that “deep” learning refers to there being lots of layers (around 12 layers in 2012) of linear threshold neurons rather than the smaller number of layers (typically two or three) that had been used previously.

Now LLMs are built on top of these sorts of networks with many more layers, and many subnetworks.  This is what got everyone excited about Artificial Intelligence, after 65 years of constant development of the field.

Despite their successes with language, LLMs come with some serious problems of a purely implementation nature.

First, the amount of examples that need to be shown to a network to learn to be facile in language takes up enormous amounts of computation, so the that costs of training new versions of such networks is now measured in the billions of dollars, consuming an amount of electrical power that requires major new investments in electrical generation, and the building of massive data centers full of millions of the most expensive CPU/GPU chips available.

Second, the number of adjustable weights shown in the figure are counted in the hundreds of billions meaning they occupy over a terabyte of storage. RAM that is that big is incredibly expensive, so the models can not be used on phones or even lower cost embedded chips in edge devices, such as point of sale terminals or robots.

These two drawbacks mean there is an incredible financial incentive to invent replacements for each of (1) our humble single neuron models that are close to seventy  years old, (2) the way they are organized into networks, and (3) the learning methods that are used.

That is why I predict that there will be lots of explorations of new methods to replace our current neural computing mechanisms. They have already started and next year I will summarize some of them. The economic argument for them is compelling. How long they will take to move from initial laboratory explorations to viable scalable solutions is much longer than everyone assumes. My prediction is there will be lots of interesting demonstrations but that ten years is too small a time period for a clear winner to emerge. And it will take much much longer for the current approaches to be displaced. But plenty of researchers will be hungry to do so.

LLMs

LLMs that can explain which data led to what outputs will be key to non annoying/dangerous/stupid deployments. They will be surrounded by lots of mechanism to keep them boxed in, and those mechanisms, not yet invented for most applications, will be where the arms races occur.

The one thing we have all learned, or should have learned, is that the underlying mechanism for Large Language Models does not answer questions directly. Instead, it gives something that sounds like an answer to the question. That is very different from saying something that is accurate. What they have learned is not facts about the world but instead a probability distribution of what word is most likely to come next given the question and the words so far produced in response. Thus the results of using them, uncaged, is lots and lots of confabulations that sound like real things, whether they are or not.

We have seen all sorts of stories about lawyers using LLMs to write their briefs, judges using them to write their opinions, where the LLMs have simply made up precedents and fake citations (that sound plausible) for those precedents.

And there are lesser offenses that are still annoying but time consuming. The first time I used ChatGPT was when I was retargeting the backend of a dynamic compiler that I had used on half a dozen architectures and operating systems over a thirty year period, and wanted to move it to the then new Apple M1 chips. The old methods of changing a chunk of freshly compiled binary from data as it was spit out by the compiler, into executable program, no longer worked, deliberately so as part of Apple’s improved security measures. ChatGPT gave me detailed instructions on what library calls to use, what their arguments were, etc. The names looked completely consistent with other calls I knew within the Apple OS interfaces. When I tried to use them from C, the C linker complained they didn’t exist. And then when I asked ChatGPT to show me the documentation it groveled that indeed they did not exist and apologized.

So we all know we need guard rails around LLMs to make them useful, and that is where there will be lot of action over the next ten years. They can not be simply released into the wild as they come straight from training.

This is where the real action is now. More training doesn’t make things better necessarily. Boxing things in does.

Already we see companies trying to add explainability to what LLMs say. Google’s Gemini now gives real citations with links, so that human users can oversee what they are being fed. Likewise, many companies are trying to box in what their LLMs can say and do. Those that can control their LLMs will be able to deliver useable product.

A great example of this is the rapid evolution of coding assistants over the last year or so. These are specialized LLMs that do not give the same sort of grief to coders that I experienced when I first tried to use generic ChatGPT to help me. Peter Norvig, former chief scientist of Google, has recently produced a great report on his explorations of the new offerings. Real progress has been made in this high impact, but narrow use field.

New companies will become specialists in providing this sort of boxing in and control of LLMs. I had seen an ad on a Muni bus in San Francisco for one such company, but it was too fleeting to get a photo. Then I stumbled upon this tweet that has three such photos of different ads from the same company, and here is one of them:

The four slogans on the three buses in the tweet are: Get your AI to behave, When your AI goes off leash, Get your AI to work, and Evaluate, monitor, and guardrail your AI. And “the AI” is depicted as a little devil of sorts that needs to be made to behave.

Self Driving Cars

This is one of my three traditional sections where I update one of my three initial tables of prediction from  my predictions exactly eight years ago today. In this section I talk about self driving cars, driverless taxi services, and what that means, my own use of driverless taxi services in the previous year, adoption of electric vehicles in the US, and flying cars and taxis, and what those terms mean.

No entries in the table specifically involve 2025 or 2026, and the status of  predictions that are further out in time remain the same. I have only put in one new comment, about how many cities in the US will have self-driving (sort of) taxi services in 2026 and that comment is highlighted,

A Brief Recap of what “Self Driving” Cars Means and Meant

This is a much abridged and updated version of what I wrote exactly one year ago today.

The definition, or common understanding, of what self driving cars really means has changed since my post on predictions eight years ago.  At that time self driving cars meant that the cars would drive themselves to wherever they were told to go with no further human control inputs. It was implicit that it meant level 4 driving. Note that there is also a higher level of autonomy, level 5, that is defined.

Note that in the second row of content, it says that there will be no need for a human to take over for either level 4 or level 5. For level 4 there may be pre-conditions on weather and within a supported geographic area. Level 5 eliminates pre-conditions and geographic constraints. So far no one is claiming to have level 5.

However the robot taxi services such as Cruise (now defunct), Waymo, currently operating in five US cities, and Zoox, currently operating in two cities with limited service (Las Vegas and San Francisco), all relied, or rely, on having remote humans who the car can call on to help get them out of situations they cannot handle. That is not what level 4 promises. To an outside observer it looks like level 4, but it is somewhat less than that in reality. This is not the same as a driver putting their hands back on the steering wheel in real time, but it does mean that there is sometimes a remote human giving high level commands to the car. The companies do not advertise how often this happens, but it is believed to be every few miles of driving. The Tesla self driving taxis in Austin have a human in the passenger seat to intervene when there is a safety concern.

One of the motivations for self driving cars was that the economics of taxis, cars that people hire at any time for a short ride of a few miles from where they are to somewhere else of their choosing, would be radically different as there would be no driver. Systems which do require remote operations assistance to get full reliability cut into that economic advantage and have a higher burden on their ROI calculations to make a business case for their adoption and therefore their time horizon to scaling across geographies.

Actual self-driving is now generally accepted to be much harder than every one believed.

As a reminder of how strong the hype was and the certainty of promises that it was just around the corner here is a snapshot of a whole bunch of predictions by major executives from 2017.

I have shown this many times before but there are three new annotations here for 2025 in the lines marked by a little red car. The years in parentheses are when the predictions were made. The years in blue are the predicted years of achievement. When a blue year is shaded pink it means that it did not come to pass by then. The predictions with orange arrows are those that I had noticed had later been retracted.

It is important to note that every prediction that said something would happen by a year up to and including 2025 did not come to pass by that year.  In fact none of those have even come to pass by today. NONE. Eighteen of the twenty predictions were about things that were supposed to have happened by now, some as long as seven years ago. NONE of them have happened yet.

My Own Experiences with Waymo in 2025

I took two dozen rides with Waymo in San Francisco this year. There is still a longer wait than for an Uber at most times, at least for where I want to go. My continued gripe with Waymo is that it selects where to pick me up, and it rarely drops me right at my house — but without any indication of when it is going to choose some other drop off location for me.

The other interaction I had was in early November when I felt like I was playing bull fighter, on foot, to a Waymo vehicle.  My house is on a very steep hill in San Francisco, with parallel parking on one side and ninety degree parking on the other side. It is rare that two cars can pass each other traveling in opposite directions without one having to pull over into some empty space somewhere.

In this incident I was having a multi-hundred pound pallet of material deliverd to my home. There was a very big Fedex truck parked right in front of my house, facing uphill, and the driver/operator was using a manual pallet jack to get it onto the back lift gate, but the load was nine feet long so it hung out past the boundary of the truck. An unoccupied Waymo came down the hill and was about to try to squeeze past the truck on that side. Perhaps it would have made it through if there was no hanging load. So I ran up to just above the truck on the slope and tried to get the Waymo to back up by walking straight at it. Eventually it backed up and pulled in a little bit and sat still. Within a minute it tried again. I pushed it back with my presence again. Then a third time. Let’s be clear it would have been a dangerous situation if it had done what it was trying to do and could have injured the Fedex driver who it had not seen at all. But any human driver would have figured out what was going on and that the Fedex truck would never go down the hill backwards but would eventually drive up the hill. Any human driver would have replanned and turned around. After the third encounter the Waymo stayed still for a while. Then it came to life and turned towards the upwards direction, and when it was at about a 45 degree angle to the upward line of travel it stopped for a few seconds. Then it started again and headed up and away.  I infer that eventually the car had called for human help, and when the human got to it, they directed it where on the road to go to (probably with a mouse click interface) and once it got there it paused and replanned and then headed in the appropriate direction that the human had made it already face.

Self Driving Taxi Services

There have been three self driving taxi services in the US in various stages of play over the last handful of years, though it turns out, as pointed out above that all of them have remote operators. They are Waymo, Cruise, and Zoox.

___Cruise

Cruise died in both 2023 and 2024, and is now dead, deceased, an ex self driving taxi service. Gone. I see its old cars driving around the SF Bay Area, with their orange paint removed, and with humans in the driver seat. On the left below are two photos I took on May 30th at a recharge station. “Birdie” looked just like an old Cruise self driving taxi, but without an orange paint. I hunted around around in online stories about Cruise and soon found another “Birdie”, with orange paint, and the same license plate. So GM are using them to gather data, perhaps for training their level 3 driving systems.

___Tesla

Tesla announced to much hoopla that they were starting a self driving taxi service this year, in Austin.  It requires a safety person to be sitting in the front passenger seat at all times. Under the certification with which they operate, on occasion that front seat person is required to move to the driver’s seat. Then it just becomes a regular Tesla with a person driving it and FSD enabled. The original fleet was just 30 vehicles, with at least seven accidents reported by Tesla by October, even with the front seat Tesla person. In October the CEO announced that the service would expand to 500 vehicles in Austin in 2025. By November he had changed to saying they would double the fleet.  That makes 60 vehicles. I have no information that it actually happened.

He also said he wanted to expand the “Robotaxi” service to Phoenix, San Francisco, Miami, Las Vegas, Dallas, and Houston by the end of 2025. It appears that Tesla can not get permits to run even supervised (mirroring the Austin deployment) in any of those cities. And no, they are not operating in any of those cities and now 2025 has reached its end.

In mid-December there were confusing reports saying that Tesla now had Model Y’s driving in Austin without a human safety monitor on board  but that the Robotaxi service for paying customers (who are still people vetted by Tesla) resumed their human safety monitors. So that is about three or four years behind Waymo in San Francisco, and not at all at scale.

The CEO of Tesla has also announced (there are lots of announcements and they are often very inconsistent…) that actually the self driving taxis will be a new model with no steering wheel nor other driver controls. So they are years away from any realistic deployment. I will not be surprised if it never happens as the lure of humanoids completely distracts the CEO.  If driving with three controls, (1) steering angle of the front wheels, (2) engine torque (on a plus minus continuum), and (3) brake pedal pressure, are too hard to make actually work safely for real, how hard can it be to have a program control a heavy unstable balancing platform with around 80 joints in hips and waist, two legs, two arms and five articulated fingers on each hand?

___Waymo

Meanwhile Waymo had raised $5.6B to expand to new cities in 2025. It already operated in parts of San Francisco, Los Angeles, and Phoenix. During 2025 it expanded to Austin and Atlanta, the cities it had promised. It also increased its geographic reach in its existing cities and surrounding metropolitan areas.  In the original three cities users have a Waymo app on their phone and specifically summon a Waymo. In the new cities however they used a slightly different playbook. In both Austin and Atlanta people use their standard Uber app.  They can update their preference to say that they prefer to get a Waymo rather than a human driven car, but there is no guarantee that a Waymo is what they will get. And any regular user of the Uber app in those cities may be offered a Waymo, but they do get an option to decline and to continue to wait for a human driven offer.

In the San Francisco area, beyond the city itself, Waymo first expanded by operating in Palo Alto, in a geographically separate area. Throughout the year one could see human operated Waymos driving in locations all along the peninsula from San Francisco to Palo Alto and further south to San Jose. By November Waymo had announced driverless operations throughout that complete corridor, an area of 260 square miles, but not quite yet on the freeways–the Waymos are operating on specific stretches of both 101 and 280, but only for customers who have specifically signed up for that possibility. Waymo is now also promising to operate at the two airports, San Jose and San Francisco. The San Jose airport came first, and San Francisco airport is operating in an experimental mode with a human in the front seat.

Waymo has announced that it will expand to five more cities in the US during 2026; Miami, Dallas, Houston, San Antonio, and Orlando. It seems likely, given their step by step process, and their track record of meeting their promises that Waymo has a good shot at getting operations running in these five cities, doubling their total number of US cities to 10.

Note that although it does very occasionally snow in five of these ten cities (Atlanta, Austin, Houston, San Antonio, and Orlando) it is usually only a dusting. It is not yet clear whether Waymo will operate when it does snow. It does not snow in the other five cities, and in San Francisco Waymo is building to be a critical part of the transportation infrastructure. How well that would work if a self driving taxi service was subject to tighter restrictions than human driven services due to weather could turn into a logistical nightmare for the cities themselves. In the early days of Cruise they did shut down whenever there was a hint of fog in San Francisco, and that is a common occurrence. It was annoying for me, but Cruise never reached the footprint size in San Francisco that Waymo now enjoys.

No promises yet from Waymo about when it might start operating in cities that do commonly have significant snow accumulations.

In May of 2025 Waymo announced a bunch of things in one press release. First, that they had 1,500 Jaguar-based vehicles at that time, operating in San Francisco, Los Angeles, Phoenix, and Austin. Second, that they were no longer taking deliveries of any more Jaguars from Jaguar, but that they were now building two thousand  of their own Jaguars in conjunction with Magna (a tier one auto supplier that also builds small run models of big brands — e.g., they build all the Mini Coopers that BMW sells) in Mesa, Arizona. Third, that they would also start building, in late 2025, versions of the Zeekr RT, a vehicle that they co-designed with Chinese company Geely, that can be built with no steering wheel or other controls for humans, but with sensor systems that are self-cleaning.

It is hard to track exactly how many Waymos are deployed, but in August 2025, this website, citing various public disclosures by Waymo, put together the following estimates for the five cities in which Waymo was operating.

Phoenix            400
San Francisco      800
Los Angeles        500
Austin             100
Atlanta             36

No doubt those numbers have increased by now.  Meanwhile Waymo has annualized revenues of about $350M and is considering an IPO with a valuation of around $100B.  With numbers like those it can probably raise significant growth capital independently from its parent company.

___Zoox

The other self driving taxi system deployed in the US is Zoox which is currently operating only in small geographical locations within Las Vegas and San Francisco. Their deployment vehicles have no steering wheel or other driver controls–they have been in production for many years. I do notice, by direct observation as I drive and walk around San Francisco, that Zoox has recently enlarged the geographic areas where its driverful vehicles operate, collecting data across all neighborhoods. So far the rides are free on Zoox, but only for people who have gone through an application process with the company. Zoox is following a pattern established by both Cruise and Waymo. It is roughly four years behind Cruise and two years behind Waymo, though it is not clear that it has the capital available to scale as quickly as either of them.

All three companies that have deployed actual uncrewed self driving taxi services in the US have been partially or fully owned by large corporations. GM owned Cruise, Waymo is partially spun out of Google/Alphabet, and Zoox is owned by Amazon.

Cruise failed. If any other company wants to compete with Waymo or Zoox, even in cities where they do not operate, it is going to need a lot of capital. Waymo and Zoox are out in front. If one or both of them fail, or lose traction and fail to grow, and grow very fast, it will be near to impossible for other companies to raise the necessary capital.

So it is up to Waymo and Zoox.  Otherwise, no matter how well the technology works, the dream of  driverless taxis is going to be shelved for many years.

Electric Cars

In my original predictions I said that electric car (and I meant battery electric, not hybrids) sales would reach 30% of the US total no earlier than 2027.  A bunch of people on twitter claimed I was a pessimist. Now it looks like I was an extreme optimist as it is going to take a real growth spurt to reach even 10% in 2026, i.e., earlier than 2027.

Here is the report  that I use to track EV sales — it is updated every few weeks. In this table I have collected the quarterly numbers that are finalized. The bottom row is the percentage of new car sales that were battery electric.

Although late in 2024 EV sales were pushing up into the high eight percentage points they have dropped back into the sevens this year in the first half of the year. Then they picked up to 10.5% in the third quarter of 2025, but that jump was expected as the Federal electric vehicle (EV) tax credits ended for all new and used vehicles purchased after September 30, 2025, as part of the “One Big Beautiful Bill Act”.  People bought earlier than they might have in order to get that tax credit, so the industry is expecting quite a slump in the fourth quarter, but it will be a couple more months before the sales figures are all in.  YTD 2025 is still under 8.5%, and is likely to end at under 8%.

The trends just do not look like we will get to EVs reaching 12% of US cars being sold in 2027, even with a huge uptick. 30% is just not going to happen.

As for which brands are doing better than others, Tesla’s sales dropped a lot more than the rest of the market. Brand winners were GM, Hyundai, and Volkswagen.

The US experience is not necessarily the experience across the world. For instance Norway reached 89% fully electric vehicles of all sold in 2024, largely due to taxes on gasoline powered car purchases. But that is a social choice of the people of Norway, not at all driven by oil availability. With a population of 5.6 million compared to the US with 348 million, and domestic oil production of 2.1 million barrels per day, compared to the US with 13.4 million b/d, Norway has a per capita advantage of almost ten times as much oil per person (9.7 to be more precise).

Electrification levels of cars is a choice that a country makes.

Flying Cars

The next two paragraphs are reproduced from last’s years scorecard.

Flying cars are another category where the definitions have changed. Back when I made my predictions it meant a vehicle that could both drive on roads and fly through the air.  Now it has come to mean an electric multi-rotor helicopter than can operate like a taxi between various fixed landing locations. Often touted are versions that have no human pilot. These are known as eVTOLs, for “electric vertical take off & landing”.

Large valuations have been given to start ups who make nice videos of their electric air taxis flying about. But on inspection one sees that they don’t have people in them. Often, you might notice, even those flights are completely over water rather than land. I wrote about the lack of videos of viable prototypes back in November 2022.

The 2022 post referred to in the last sentence was trying to make sense of a story about a German company, Volocoptor, receiving a $352M Series E investment. The report from pitchbook predicted world wide $1.5B in revenue in the eVTOL taxi service market for 2025.  I was bewildered as I could not find a single video, as of the end of 2022, of a demo of an actual flight profile with actual people in an actual eVTOL of the sort of flights that the story claimed would be generating that revenue in just 3 years.

I still can’t find such a video. And the actual revenue for actual flights in 2025 turned out to be $0.0B (and there are no rounding errors there — it was $0) and Volocoptor has gone into receivership, with a “reorganization success” in March 2025.

In my November 2022 blog post above I talked about another company, Lilium, which came the closest to having a video of a real flight, but it was far short of carrying people and it did not fly as high as is needed for air taxi service. At the time Lilium had 800 employees.  Since then Lilium has declared bankruptcy not once (December 2024), but twice (February 2025), after the employees had been working for some time without pay.

But do not fear. There are other companies on the very edge of succeeding. Oh, and an edge means that sometimes you might fall off of it.

Here is an interesting report on the two leading US eVTOL companies, Archer and Joby Aviation, both aiming at the uncrewed taxi service market; both with valuations in the billions, and both missing just one thing. A for real live working prototype.

The story focuses on a pivotal point, the moment when an eVTOL craft has risen vertically, and now needs to transition to forward motion. In particular it points out that Archer has never demonstrated that transition, even with a pilot onboard, and during 2025 they cancelled three scheduled demonstrations at three different air shows. They did get some revenue in 2025 by selling a service forward to the city of Abu Dhabi, but zero revenue for actual operations–they have no actual operations.  They promise that for this year, 2026, with revenue producing flights in the second half of the year.

Joby Aviation did manage to demonstrate the transition maneuver in April of 2025. And in November they made a point to point flight in Dubai, i.e., their test vehicle managed to take off somewhere and land at a different place. The fact that there were press releases for these two human piloted pretty basic capabilities for an air taxi service suggests to me that they are still years away from doing anything that is an actual taxi service (and with three announced designated place to land and take off from it seems more like a rail network with three stations rather than a taxi service–again slippery definitions do indeed slip and slide). And many more years away from a profitable service. But perhaps it is naive of me to think that a profitable business is the goal.

As with many such technology demonstrators the actual business model seems to be getting cities to spend lots of money on a Kabuki theater technology show, to give credit to the city as being technology forward. Investors, meanwhile invest in the air taxi company thinking it is going to be a real transportation business.

But what about personal transport that you own, not an eVTOL taxi service at all,but an eVTOL that you can individually own, hop into whenever you want and fly it anywhere? In October there was a story in the Wall Street Journal: “I Test Drove a Flying Car. Get Ready, They’re Here.” The author of the story spent three days training to be the safety person in a one seat Pivotal Helix (taking orders at  $190,000 a piece, though not yet actually delivering them; also take a look at how the vehicles lurch as they go through the pilot commanded transition maneuver). It is a one seater so the only person in the vehicle has to be the safety person in case something fails. He reports:

After three hellish days in a drooling, Dramamine-induced coma, I failed my check ride.

The next month he tried again. This time he had a prescription for the anti-emetic Zofran and a surplus-store flight suit. The flight suit was to collect his vomit and save his clothes.  After four more days of training (that is seven total days of training), he qualified and finally took his first flight, and mercifully did not live up to his call sign of “Upchuck Yeager”.  $\190,000 to buy the plane, train for seven days, vomit wildly, have to dress in a flight suit, and be restricted to take off and landing and only fly over privately owned agricultural land or water. This is not a consumer product, and this is not a flying car that is here, despite the true believer headline.

Two years ago I ended my review of flying cars with:

Don’t hold your breath. They are not here. They are not coming soon.

Last year I ended my review with:

Nothing has changed. Billions of dollars have been spent on this fantasy of personal flying cars.  It is just that, a fantasy, largely fueled by spending by billionaires.

There are a lot of people spending money from all the investments in these companies, and it is a real dream that they want to succeed for many of them. But it is not happening, even at a tiny scale, anytime soon.

Robotics, AI, and Machine Learning

We are peak popular hype in all of robotics, AI, and machine learning. In January 1976, exactly fifty years ago, I started work on a Masters in machine learning. I have seen a lot of hype and crash cycles in all aspects of AI and robotics, but this time around is the craziest.  Perhaps it is the algorithms themselves that are running all our social media that have contributed to this.

But it does not mean that the hype is justified, or that the results over the next decade will pay back the massive investments that are going in to AI and robotics right now.

The current hype is about two particular technologies, with the assumption that these particular technologies are going to deliver on all the competencies we might ever want.  This has been the mode of all the hype cycles that I have witnessed in these last fifty years.

One of the current darling technologies is large X models for many values of X (including VLMs and VLAs), largely, at the moment, using massive data sets, and transformers as their context and sequencing method. The other, isn’t even really a technology, but just a dream of a form of a technology and that  is robots with humanoid form.

I have now put these two things in my five topics of my new predictions shared at the beginning of this post and will talk about them explicitly for each of the next ten years.

Back in 2018 I did not talk about either of these technologies in my predictions, but rather talked about competences and capabilities.  I fear that I may have been overly optimistic about many of these and in the table below I point out that my predicted time of arrival has now come, but the capabilities or competencies have not.  I’m sure that many true believers in the two technologies mentioned above will have very short time scales on when they say this will be achieved. I pre-emptively disagree with them.

Capabilities and Competences

The predictions that are commented upon in the table above are all about when we would see robots and AI systems doing some things that simple creatures can do and others that any child of age nine or perhaps less can do without any difficulty.

Even children aged three or four can navigate around cluttered houses without damaging them (that is different from when they may want to damage them). They can get up and down single stairs, and even full stair cases on two legs without stumbling (or resorting to four limb walking as a two year old might). By age four they can open doors with door handles and mechanisms they have never seen before, and safely close those doors behind them. They can do this when they enter a particular house for the first time. They can wander around and up and down and find their way.

One of the many promises about humanoid robots is that they too will be able to do this. But that is not what they can do today. But wait, you say, “I’ve seem them dance and somersault, and even bounce off walls.” Yes, you have seen humanoid robot theater. All those things are done on hard surfaces, and anything specific beyond walking has been practiced and optimized by reinforcement learning, for exactly the situation of the floors and walls as they are. There is no real-time sensing and no ability to wander in previously unseen environments, especially not those with slipping hazards such as towels or sheets of cardboard on the floor.  Children can do so easily.  While four legged robots are much better at it than humanoid robots, they are wider than people, and still have significant foot slipping problems, and cannot open random doors themselves as children can.

A nine year old child can pretty much do any task (but with less weighty packages) than any delivery driver can do.  That includes climbing out of a van, walking up and down slopes, going up and down previously unseen external staircases, sometimes ending in a dark porch or vestibule area, then putting the package on the ground, or putting it into a drop bin after grasping and pulling on the handle — again never having encountered that particular design of bin and handle. All this can be done immediately upon seeing the scene for the first time. We have not seen anything remotely like that in a lab demo for robots, despite my hope from eight years ago that by now such would have been demonstrated. And again, here a four legged robot might be able to do the walking and stair climbing, but it won’t be able to manipulate the package. Also note that humans doing these tasks don’t just carry single packages out in front of them with two outstretched arms, but often use their elbows, their hips, and their bellies to support multiple packages as they locomote.

Elder care is a commonly quote target market for robots, and with good reason given the current and growing demographic inversions in much of the world. There are far fewer younger people relative to the number of older people than there have been historically, and so less people to provide elder care.  In providing care to the very elderly, there is a need to support those people physically, both passively, providing compliant support for them to lean on, and actively, getting people into and out of bed, into and out of bathtubs or shower enclosures, and getting people onto and off of toilets. And sometimes wiping their bums. There are no force sensing and control capabilities on any of today’s robots which are remotely capable of doing any of these sorts of things safely and comfortably. And machine learning is not going to provide those capabilities. There are many fundamental design, materials, and engineering problems to solve to make these things possible.  A bitter lesson, perhaps, for those who think that more data will solve everything.

But the other unresolved capability that I have in my predictions table above is an agent that understands the world in an ongoing way as we all understand it.  That includes knowing what to expect to be the same as it was yesterday, and will be tomorrow, and what has changed about the world since yesterday or is likely to change today or tomorrow. Such an understanding of the world will be important for any deployable systems that can take care of real and vulnerable humans, including the elderly. And the young. And the rest of us.

In summary, I thought that more progress would be made on many of these problems than has been achieved over the last eight years. That lack of progress is going to have real, and negative, impact on the quality of life of the newly elderly for the next couple of decades. Ouch!

VCs, please take note: there are real pulls on having technologies that can help the elderly, and being in there first with something that can actually deliver value in the next three to five years will be a come with a very large upside.

World Models

Lots of people are talking about world models and their importance, as add ons to LLMs, as mechanisms for agentic AI to exploit, and for allowing robots to do real tasks. These aspirations are probably reasonable to have, and successfully working on them can have real impacts.

Unfortunately the talkers are not the doers, and not the deployers, and not the people who have to solve real problems. And so they all have different, and convenient for themselves, understandings of what world models are.  That, along with the worship of big data and the belief that machine learning will solve all problems means we have a big mess, with people jumping to “solutions” before they understand the problems.

Some people are even claiming that they will build world models by learning them from having agents play video games.  But how do those video games work? They have a coded geometry-based world model, with a little physics engine. It is already built!  Using machine learning (and tens of millions of dollars) to extract it rather than just looking at the source code (and perhaps buying or licensing that code) is just wacky.

Expect more confusion and lots and lots of reinvention. This fever has quite a ways to go before today’s memes and slogans get replaced by the next generation of memes and slogans, with perhaps some good work coming out in a rational interregnum. We can hope.

Situatedness vs Embodiment

One of the new things that people are getting excited about is Embodied Intelligence.  I agree that it is worth being excited about, as it is what I have spent the last forty years work on.  It is certainly about robots being in the world.

But since 1991 I have made a distinction between two concepts where a machine, or creature can be either, neither, or both situated and embodied. Here are the exact definitions that I wrote for these back then:

[Situatedness] The robots are situated in the world—they do not deal with abstract descriptions, but with the here and now of the world directly in-fluencing the behavior of the system. 

[Embodiment] The robots have bodies and experience the world directly—their actions are part of a dynamic with the world and have immediate feed-back on their own sensations. 

At first glance they might seem very similar.  And they are, but they are also importantly different. And, spoiler alert, I think much of the work at companies, large and small, right now, is trying abstract out the embodiment of a robot, turning it into a machine that is merely situated.

An algorithm, written as code, to find the greatest common divisor of two numbers, when running, is neither situated nor embodied.

A robot that is thrown into the air with just an inertial measurement unit (IMU) as its sensor that moves its limbs about to zero out rotations and then is caught by a net is embodied but not situated.

A robot that has a physical face that can make expressions with it, a voice synthesizer, cameras, and microphones and that can talk to a person giving appropriate responses both with its choice of words and with appropriate prosody and facial expressions, to some purpose and in response to how the person talks and moves, is situated but not really embodied. Embodied in its presence yes, but not embodied in any physical interactions with its environment.

A robot that can roll around without hitting stationary objects, wherever they are, nor hitting moving people or other vehicles, that can go to a location specified by a warehouse management system, that responds safely to people grabbing it anywhere, and can give a person who grabs its control handle agency over it going wherever the person pushes it with a light touch no matter how much weight it is currently carrying, is both embodied and situated. [And yes, this is what our Carter robots do at Robust.AI.]

These are just point examples of the four classes of entities that come from having or not having the two properties of situatedness and embodiment.

Real robots that do real work in dynamic human occupied environments must be both situated and embodied. For instance, a robot that is to help with in home elder care needs to be aware of the situation in the world in order to know what to do to help the person.  It needs to be able to open doors with different handles and latching mechanisms, and then control the inertia of the closing door so that the environment is both safe and quiet for the person. The robot needs to be able to accommodate the person reaching for it dynamically, looking for support that so that they don’t fall. The robot needs to able to take things handed to it by the person, and pass things to the person in a way which is both safe and makes it easy for the person to grasp. Etc., etc.

In short the robot needs to control forces and inertias in the world and to be responsive to them, at that same time as it is acting in a way that can be understood as sentient.

Being both situated and embodied is still a challenge to robots in the world.  [[Now here is the most important sentence of this whole blog post.]] I think the training regimes that being used for both locomotion and dexterity are either ignoring or trying to zero out the embodiment of physical robots, their inertias and forces, reducing them to merely being situated, just apps with legs and arms, characters in video games, not the reality of real physical beings that the tasks we want them to do requires.

Dexterous Hands

I talked about the challenges for dexterity earlier this year. In the table above I have a new comment this year saying that there has been improvement in the dexterity of suction based grippers but not for articulated grippers.

Suction grippers have plastic suction cups which themselves are compliant. Under the force of the suction they can change shape, to a degree, to accommodate unknown shapes in the thing being grasped (sucked up to).  They also allow for a little bit of torsional rotation about the axis of sucking and a bit of rocking of the suction cup in the two degrees of freedom in the plane orthogonal to the suction axis.

While suction cups have evolved to better pick things up and so are common for handling packaged goods, the companies that package materials to be shipped through automated systems choose versions of plastics for bags that won’t be sheared open by the suction pulling against outer parts of such cups.

The result is that the control of the embodied action of grasping can become much more a simply situated action. Once the pick orientation and vacuum gripper selection has been made it is really an open loop as all the work is done by the indiscriminate force of suction and the mutual compliance of the gripper and the grippee.

Above I had argued against do this with a general purpose humanoid hand. It makes no sense there as the adaptability of the hand is its most precious attribute. But here in a special purpose hand, a suction gripper, it actually simplifies things within the specialization of task, and here a purely situated hand may make sense. And it may be possible to train it with purely visual data.

So what does this tell us?  It says that there is plenty of room for mechanical design, and simpler computational embodied control for all sorts of grippers and things in the world that need to be gripped.

The end of Moore’s Law, at least the version that said we could reduce feature size on silicon by a factor two every year, opened up a new golden era of chip design. The winners (through early luck and then dogged determination), matched untraditional designs to new problems (machine learning) and achieved speedups (and corporate valuations) that were unheard of. In the last 10 years we have moved from general purpose silicon to special purpose silicon for our most high volume computations.  That was not on most people’s expectation list twenty years ago.

So too today, with stalled capabilities from full human hand emulation efforts through machine learning from visual observation, there is a rich array of more specialized manipulation tasks where special purpose grippers, clever interplay of materials and force applications, geometric planning, specialized sensing, and maybe even some machine learning may lead to enormous application markets.

For instance, a specialized robot body, hands (of some sort), arms, and support limbs or wheels that can safely manipulate an elderly human could have enormous impact on elder care around the world. A single human care-giver along with one human-manipulator robot could provide a lot more care for a frail elderly person than the care-giver alone could do.

Special purpose manipulators for fruits, or for some range of small mechanical parts, or clothing, could each open enormous markets for automation in particular handling tasks for each of them. And countless other specialities.

Economic pull is out there.  Being the smart academic researcher, entrepreneur, or technology investor, may lead to enormous new types of deployable automation.

The new dexterity may turn out to be special purpose. And eventually we may come to understand that just because the hands we know best happen to be our own, does not mean that our own hands are the best for the majority of tasks in our human world.

Humanoid romanticism may not be our future after all.

HUMAN SpaceFLIGHT

Looking at the missions and numbers over the last three years it appears that human spaceflight is at a steady plateau, with, by the way, far fewer people going into orbit that in the time of the Space Shuttle.  Underneath though, there is a lot of churn, a likely new player, and the return of humans to lunar distances for the first time in 54 years.

Below is the updated scoring of my 2018 predictions for human spaceflight. There are six new comments in this table, but no new specific calling of predicted dates as right or wrong. It is now clear to me that I was way too optimistic in regard to my predictions for Mars, even though I was wildly out of step and much more pessimistic then the predictions coming out of SpaceX eight years ago. Given how slow things have turned out trying to land people on the Moon, the hoped for crewed colony on the Moon (think of it as ISS (International Space Station) on the lunar surface) may well slip to what I had predicted for Mars.  Mars is going to take much longer than the Moon.

Following the table there are the detailed numbers and trends on both orbital crewed flights, and suborbital crewed flights. Things will change from stasis in 2026.  A crewed flight to the Moon is scheduled to happen in a matter of weeks, with the vehicle already stacked, now.  And suborbital crewed flights may possibly have quite an uptick in 2026.  Following those two sections I have more on Boeing’s Starliner, SpaceX’ Starship, and Blue Origin’s New Glenn, NASA and the Moon, and what is going to happen with space stations given the scheduled end of life of the ISS in 2030.

Orbital Crewed Flights

In both 2024 and 2025 the US put 16 people into orbit and Russian and China put 6 people each into orbit; 28 people total went to orbit in each year. We have gone from a historical low of only eight people going to orbit in 2020 to a steady-ish state of roughly 28 people per year now. That may jump up to over 30 people in 2026 because of the additional Artemis II flight to the Moon, following checkout in LEO (Low Earth Orbit).  But even with that bump there may be other pressures which keep it from rising above the high twenties for 2026

We are certainly not seeing steady growth in the number of humans getting launched to orbit, and the numbers are significantly lower than the hey days of Shuttle launches in the nineties and early two thousands. There is no growth trend visible, and the long promised exponential growth of people going to orbital space has not even made a brief guest appearance.

Here is a more detailed history for the last six years where the first line in each box says how many crewed launches of the particular vehicle there were, and the second line, in square brackets says how many people, total, were onboard those flights. Wherever there are three numbers separated by forward slashes you have to sum the numbers to get the total.

The three countries with current crewed orbital launch capabilities are the US, Russia, and China.

All Chinese flights are state astronauts (or taikonauts) and all of them go to the Chinese space station. And there are no tourists, so far, on Chinese flights, so we just have single numbers for both launches and people.

All the state astronauts for both the US and Russia go to the International Space Station (ISS), but a state player (in Russia) and a non-state player in the US (SpaceX) have also launched tourist flights in the last six years. So for those two countries we have three numbers separated by slashes for both launches and people. The first of the three numbers refers to purely state launches to the ISS (note that the US and Russia both launch each others state astronauts to the ISS so that both countries have astronauts up-to-date trained on the other’s launch systems, in case of emergencies arising at some point). The second number in the triples is space tourists whose destinations have also been the ISS, while the third number (for both launches and people) is for tourist flights that have been independent of going to the ISS — there have been a total of three of these, all launched by SpaceX. Two of those three flights were purchased personally by Jared Issacman, who has now been sworn in as the NASA administrator just two weeks ago.

The one year in the last six where Russia has launched space tourists (after being the leaders in this endeavor early in the century) was 2021, where two flights of Soyuz to the ISS had one regular state cosmonaut and two space tourists. And, there was one slightly wobbly other launch of a Soyuz in 2024, not called out in the table, where a flight attendant from the state airline of Belarus was sent as a cosmonaut from that country to the ISS on a Russian Soyuz. That was most likely an event orchestrated by Russia to keep support from Belarus for their war against Ukraine. Ugly.

The term tourist needs some explanation. The people (as with suborbital Blue Origin flights) are a mixture of private people paying the experience (or having some other individual pay for them) or they are astronauts from countries that do not have their own launch capability. In the case of the three tourist flights to the ISS on a SpaceX Dragon, all have been paid for by the company Axiom, with a former NASA astronaut in command. The three others on each of those flights are people in the fledgling astronaut program of other countries who have paid Axiom for the seats. Axiom has commercial relationships with both SpaceX and NASA for the use of the Flacon 9 launch vehicle, the Dragon craft and use fee of the ISS.

Suborbital Crewed Flights

Virgin Galactic is on a multi-year hiatus on flights as they develop new flight vehicles, but they may well fly again in 2026. Thus, for the last year, only Blue Origin has been launching  tourists (again a mixture of private individuals and astronauts from other countries that have not yet developed their own crewed launch capability, but may be aiming at doing so) suborbital flights. Blue Origin also sells uncrewed launches for experiments that need to be exposed to the environment of space and/or operation in microgravity, if only for a few minutes.

In 2025 Blue Origin had seven launches each with six people on board. Previously they had had three crewed launches in each of 2021, 2022, and 2024, each with six people on board, with a hiatus in 2023.

Blue Origin has been quite careful with forward projections for both suborbital and orbital flights, so when they say what they intend to do and when, they are likely to come close to achieving that promise.

Recently they said that they are going to introduce three new flight vehicles starting in 2026 to run their suborbital flights, that they are looking at developing a second launch site, somewhere else than Texas, and that they believe they have the customer demand to support one flight per week. They do not disclose what they charge for the flights. Nor did they give any firm dates for reaching these goals. But I think it is likely that we will see a jump in the number of flights in 2026,

In December of 2025 I was at an event centered on solar system orbital dynamics and met a sub-orbital tourist there. He has already paid for and flown above the Kármán line on Virgin Galactic. Now he has paid for a Blue Origin sub-orbital flight and is waiting for a launch assignment. There is definitely a market for these flights, it remains to be seen whether the prices and demand combine in a way that makes it profitable for seat suppliers to keep doing it.

Boeing’s Starliner (not to be confused with the SpaceX Starship)

When it was first announced, in 2010, Boeing’s Starliner was originally scheduled to fly a human test crew in 2018. It was supposed send the crew to the ISS, then it would be under contract to launch six crews to the ISS, much as SpaceX has already launched 11 regular crews to the ISS.

In mid 2024 it delivered a human test crew to ISS,  Barry Wilmore and Sunita Williams, but after much analysis of anomalies it returned to Earth without them. NASA bumped two crew members from the next crew going on a SpaceX flight to the ISS to provide room for their return, on that SpaceX Dragon, which they did after an unexpected extra nine months on top of their originally scheduled week at the ISS.

Last year in my yearly update I said:

We do not know at this point, but I think it would not be a huge surprise if Starliner never flies again.

It turns out it is going to fly again! Including potentially twice in 2026. But there are some changes. The six missions which were contracted to take astronauts on regular assignment to the ISS were called Starliner-1 through Starliner-6. The contract with NASA has been modified to make the last two flights future options rather than sure things. And Starliner-1 scheduled for the first half of 2026 will be un-crewed again. Then the three remaining flights in the modified contract would each take four astronauts on regular rotations to the ISS. There is one little hiccup. Sunita Williams is the only active astronaut, not committed to other current or upcoming missions, who has trained to fly on Starliner. She now has over 600 days in space and another six month mission to the ISS would take her over radiation exposure limits.

SpaceX Falcon 9 

I gave the statistics for Falcon 9 in the introduction, talking about what has surprised me in the last 8 years. When I made my predictions Falcon 9 had been launched 46 times over 8 years. Only five of those launches re-used a previously flown first stage, and only in the previous year had successful landings of the first stage become reliable.

Now Falcon 9s are getting launched at a sustained rate of more than three per week, all attempts at landing boosters are successful, and typically each booster flies over 20 times.

Just phenomenal unmatched reliability and performance.

NASA, Artemis, and Returning to the Moon

I am jumping ahead of Starship (SpaceX) and New Glenn (Blue Origin) to talk about NASA’s plan to get people back to the lunar surface, and perhaps setting up a more or less permanent outpost there. This is how the ISS has been continuously occupied for 25 years, rotating crew members in and out twice a year. (China’s space station follows the same model, but with only 3 occupants compared to 7 for ISS).

2026 promises to be a big year for humanity and the Moon. No one has been beyond low Earth orbit (LEO) since the Apollo 17 mission had three people go to lunar orbit and two of them landed in December 1972, fifty three years ago.

In November 2022 the first launch of NASA’s SLS (Space Launch System) occurred taking its uncreewed Orion capsule in a looping orbit past the Moon and back. It approached the surface of the Moon in each direction, and then came back to Earth and splashed down. Note that this was the FIRST flight of both the multi-stage rocket, and the habitable capsule. It all worked FIRST time.  Everything was built by contractors, but it underwent NASA’s methodology to make sure things worked rather than failed.

The first stage consists of a liquid fueled rocket using four RS-25 engines, the same as the three engines on the Space Shuttle. It also has two solid fuel boosters strapped on, larger versions of the Space Shuttle solid fuel boosters. The second stage is essentially an off the shelf stage from the past Delta program. There will be a third stage added for the fourth and subsequent flights.  This is a derivative vehicle, with a long history of successful use of its components.

When Vice President Mike Pence announced the details of the program in 2019 the landing of people on the Moon was supposed to happen in 2024.  Things have slipped a little since then.

The first crewed mission to the vicinity of the Moon (no landing) Artemis II had slipped to April 2026, but now it has been pulled forward to February 2026 (next month!), when a crew of four will spend over ten days in space on Artemis II in a flight to the Moon approaching to within 4,600 miles, then in a free return manner (no need to have working engines) they will head back towards Earth.  All their energy will be removed by heat shields hitting the Earth’s atmosphere and then by the use of 11 parachutes, finally splashing down in the ocean. Note that on all 9 flights to the Moon of the Apollo program, the spacecraft came much closer to the Moon than this, and 8 of the flights went into orbit at around 60 to 70 miles above the surface. So this is a more conservative mission than those of Apollo.

Things at this stage are looking good for Artemis to fly in February 2026.

The next step of the Artemis is where things get wobbly. Rather than 2024, the first landing of astronauts on the Moon is currently scheduled for 2027. But that is not going to happen. Here is what the architecture of the mission currently looks like:

Here we see the problem with the schedule, even with it currently slipped to landing two astronauts on the Moon in 2027.

The architecture uses the SLS and Orion to get the astronauts to lunar orbit. Given there is a lunar flyby with astronauts onboard, scheduled for just two months from now (and the rocket is already stacked for that mission) that looks like a reasonable interpolation from existing progress.

The problem with the new plan is the landing vehicle and getting it to lunar orbit.  It is all based on SpaceX’s Starship. So far, Starship has had 11 flights, six of which have been successful in reaching their own goals, and 5 of which have been failures.  But there has not yet, in eleven flights, been a goal of getting anything into orbit. And just in 2025 two vehicles have been destroyed by failures on the ground when the tanks have been pressure tested. In the section on Starship below I will talk more about what I see as conflicting product requirements which together doom Starship to a very long development process.

For comparison, the Saturn V which took astronauts to the Moon nine times had a total of 13 flights, every one of which got payloads to Earth orbit. Two were uncrewed tests (and there were problems with the second and third stages on the second of these test flights). Its very first crewed flight (Apollo 8) took people to the Moon. and a total of 9 launches got people to the Moon. The other two flights were (Apollo 9) a crewed flight to test the Lunar Lander and orbital rendezvous in Earth orbit, and the uncrewed launch of the first space station, Skylab.

Now look again at the plan for the Artemis III mission.  It requires multiple (reported numbers range from 14 to somewhere into the twenties) launches of the Starship to orbit.

One of those launches uses the Super Heavy Booster and a special version of the second stage actual Starship, known as Starship HLS (Human Landing System).  That special version is expendable after it lands astronauts on the Moon, hosts them for perhaps two weeks, then brings them back to lunar orbit where they transfer to NASA’s Orion. Then it sends itself off into heliocentric orbit for all eternity. The HLS version is special in two ways. First it does not have to get back to Earth and so doesn’t need heat shields and does not need the three in-atmosphere Raptors for soft landing on Earth (see the section on Starship below).  That is good for all the mass equations. But it does, or might, have a second set of engines for landing on the Moon that are attached halfway up its body so that they cause less lunar dust to fly around as it lands. We have not yet seen a prototype of that version, not even a public rendering as far as I can tell. I have talked to people who are in regular communication with people inside SpaceX.  They report not a peep about what work has been done to design or build the lander.  That is not good for the current public schedule.

BUT the really, really bad thing is that the lunar lander stage will use up most its fuel getting into Earth orbit — it is the second stage of the rocket after all. So it cannot get to the Moon unless it is refueled.  That will be done by sending up regular versions of the Starship second stage, all on reusable Super Heavy Boosters. They too will use up most of their fuel getting to orbit, and will need to keep some to get back to Earth to be reused on another flight. But it will have a little margin and its extra fuel will be transferred to the lunar landing Starship in orbit.

No one has ever demonstrated transfer of liquid fuel in space. Because of the way the numbers work out it takes somewhere in the teens of these refueling operations, and depending on how quickly certified higher performance engines can be developed and tested for both the Super Heavy Booster and Starship itself, that number of refueling flights might range into the twenties.

As an engineer this architecture looks to me like trouble, and with an impossible future. I am sure it will not happen in 2027, and I have doubts that it ever will.

The acting administrator of NASA, Sean Duffy who is also the head of the US Department of Transportation, was worried about this too, and in October of 2025 he reopened bidding on the contract for a crewed lander for the Moon that collects and returns its crew from Orion in lunar orbit.

The day after this announcement SpaceX said they were working on a simplified architecture to land people in the Moon. They have given no details of what this architecture looks like, but here are some options proposed by the technical press.

A couple of weeks later the President announced the renomination of Jared Isaacman to be the NASA administrator, having withdrawn his nomination a few months before. Isaacman is a private citizen who personally paid for, and flew on, two of the three SpaceX crewed missions which have not flown to the ISS. He was confirmed to the NASA position on December 17^th, 2025, just two weeks ago.

At the very least expect turbulence, both political and technical, in getting astronauts landed on the Moon. And see a possible surprise development below.

SpaceX Starship (not to be confused with Boeing’s Starliner)

Starship is SpaceX’s superheavy two stage rocket, designed to put 150(?) tons of payload into orbit, with components having been under development since 2012, going through extensive redesigns along the way. There have also been three major designs, builds, and tests of the Raptor engines that power both stages.

This is how Wikipedia currently introduces them:

Raptor is a family of rocket engines developed and manufactured by SpaceX. It is the third rocket engine in history designed with a full-flow staged combustion fuel cycle, and the first such engine to power a vehicle in flight. The engine is powered by cryogenic liquid methane and liquid oxygen, a combination known as methalox.

SpaceX’s super-heavy-lift Starship uses Raptor engines in its Super Heavy booster and in the Starship second stage. Starship missions include lifting payloads to Earth orbit and is also planned for missions to the Moon and Mars. The engines are being designed for reuse with little maintenance.

Currently the Raptor 3 version is expected to be used for operational Starship launches, and it comes in two versions. There are 33 Raptors in the first stage designed to operate optimally in the atmosphere, along with three such engines in the second stage, which also houses three vacuum optimized Raptors. The first stage engines and the second stage vacuum engines are designed to get payloads to orbit. The vacuum engines on the second stage would also be used for further operations on the way to the Moon and descending towards the surface there. And for non-expendable second stages they would be used for the initial de-orbit burn for landing the second stage Starship back on Earth. After using the heat shields to burn off some more energy  as it enters the atmosphere the second set of engines, the atmosphere optimized Raptors, are used to slow it down to a soft landing.

Other systems for returning to Earth have used different tradeoffs. The Space Shuttle used its wings to slow down to very high horizontal landing speed, and then a combination of a drag parachute after touchdown and brakes on the wheels to get down to zero velocity. US capsules, such as Mercury, Gemini, Apollo, Orion, and Dragon have all used heat shields followed by parachutes during vertical fall, and lastly dropped into the sea for dampening the final residual velocity. (Soyuz, Starliner, and New Shepard all use last second retro rockets before hitting the ground, rather than water.)

This means that unlike all the other solutions Starship has to carry a complete set of engines into orbit just for use during landing, along with enough fuel and oxidant to land. This is a high performance price for the thing that flies in space, mostly. The engines on the Starship first stage, like those on Falcon 9 and Blue Origin’s New Glenn, do get to space but never get to more than a small fraction of orbital speed, so returning them to Earth is a much, much, lower performance price than Starship’s second stage return of engines and fuel.

The 2025 flights of Starship were, on average, better than the 2024 flights, but two vehicles destroyed themselves before getting to the flight stage, and still nothing got into orbit.

How close is it to working?  I don’t know.  But I do keep tabs on promises that have been made.

In November of 2024 the President of SpaceX  said “I would not be surprised if we fly 400 Starship launches in the next four years”.  A year ago today I said in response: “Looking at the success of Falcon 9 it is certainly plausible that I may live to see 400 Starship launches in a four year period, but I am quite confident that it will not happen in the next four years (2025 through 2028)”. We are a quarter of the way through her predicted time frame and we have gone from being 400 orbital launches away from her goal down to being a mere 400 away.

Blue Origin Gets to Orbit

The suborbital tourist flights that Blue Origin operates are not its main business. It has ambitions to compete head to head with SpaceX.

But it is almost 600 launches behind, how can it be competitive? In 2025 Blue Origin made clear that it is not to be dismissed. From zero orbital launches at the start of 2025 to having two orbiters on their way to Mars (SpaceX has not yet done that) and showing that it can land a booster that has very very close to the performance of Falcon Heavy’s three booster configuration when landing all three boosters. And it may well do a soft landing on the Moon in 2026 (SpaceX won’t come close to that goal for a number of years).

In February Blue Origin launched its first New Glenn rocket. It’s first stage is powered by seven BE-4 engines (“Blue Engine 4”), a methane burning engine that is more powerful than the Raptor 3 which will power new versions of SpaceX’s Starship. New Glenn reached orbit on its first attempt, and delivered a Blue Origin payload to space (a test version of their Blue Ring for in-space communications). The first stage attempted to land on Blue Origin’s Jacklyn landing platform at sea but failed.

The BE-4 had previously powered two United Launch Alliance Vulcan rockets to orbit under a deal where Blue Origin sells engines to ULA. The second stage of New Glenn is powered by two BE-3 engines, which are a variant of the single engine used on Blue Origin’s New Shepard.

In their second launch, in November, Blue Origin not only delivered three paid payloads to orbit (two of which are headed to Mars, where they will orbit the planet and carry out science experiments for UC Berkeley  on what happened to Mars’ atmosphere), but then the first stage (much larger than the first stage of a Falcon 9) landed on Jacklyn with an unrivaled level of precise control. Blue Origin plans to reduce the time spent hovering in future landings to reduce preserved fuel needs now that it has mastered return from orbit vertical landing. (Recall that they have landed dozens of New Shepard vertical landings on return from non-orbital flights.)

Soon after this impressive second outing for New Glenn, Blue Origin announced a number of upgrades.

They renamed the base vehicle that has now flown twice to be “New Glenn 7×2” where 7 and 2 refer to the number of first stage and second stage engines respectively.  They also announced that those flight engines would be upgraded to levels of thrust and duration that had already been demonstrated in ground tests. These are the new total thrust numbers, in pounds force.

             New Glenn 2025    New Glenn 7x2
1st stage    3,850,000     ==> 4,480,000
2nd stage      320,000     ==>   400,000

Additionally New Origin announced a new heavier, taller, and with larger payload faring, version, the “New Glenn 9×4” with two extra engines on each stage. Looking up from below the first stage the engine arrangement goes from the one on the left to the one on the right.

And here is who the two variants look compared to the Saturn V which took humans to the Moon in 1969.

The kicker to these successes is that the New Glenn 7×2 with a reusable first stage is very nearly equivalent to the Falcon Heavy when its three first stage boosters are reused. The reusable New Glenn 9×4 beats Falcon Heavy on all measures even when all three of Falcon Heavy are sacrificed and not recovered.  I can’t quite get all the numbers but this table makes the comparisons with the numbers I can find.

Note that a “tonne” is the British spelling for a metric ton, which is 1,000Kg. That is approximately 2,206 lbs, which is 206 lbs more than a US ton, and 34 lbs less than a British ton.

Meanwhile expectations are high for another launch of a New Glenn, the 7×2 version, sometime early in the new year. There has been no announcement from Blue Origin, nor any indication of the payload. But there is a general feeling that it may actually be a launch of Blue Origin’s Blue Moon Mark 1, an all up single launch mission to soft land on the Moon.  It was announced almost a year ago that Blue Origin has a deal to deliver a NASA payload to the Moon in the Blue Moon Pathfinder mission no earlier than 2026. The Mark 1 uses a BE-7 engine to soft land.

Here is where things get interesting for a re-appraisal of how NASA astronauts might first land on the Moon again.

Blue Origin is already under contract with NASA to land two astronauts on the Moon for a 30 day stay in 2030 using their much larger Blue Moon Mark 2.  The Mark 2 and Mark 1 share control systems and avionics, so a successful landing of Mark 1 will boost confidence in the Mark 2.  The architecture for the 2030 mission involves multiple launches. A NASA SLS launches a crewed Orion capsule to the vicinity of the Moon. A New Glenn gets a Mark 2 Blue Moon to an orbit that approaches the Moon. A “Cislunar Transporter” is launched separately and it gets fueled in LEO. Then it heads off to the same orbit as the Mark 2 and refuels it. The Mark 2 and the transporter both use three Blue Origin BE-7 engines which are now fully operational. Then the astronauts transfer to the Mark 2 to land on the Moon.  Note that this architecture uses in flight refueling as does the SpaceX version, though with far fewer launches involved.

BUT, soon after then NASA administrator Sean Duffy announced the re-opening of the contract for the lander for Artemis III, it appeared that he was considering having Blue Origin use their Mark 1 version for the crewed mission. Whether that enthusiasm survives the changing of the guard to Jared Isaacman, the new and current NASA administrator, remains to be seen.  And whether Blue Origin can pull off a rendezvous in lunar orbit, to pick up and return the crew members going to the lunar surface, from an orbiting Orion capsule is also an open question.  I think the key idea with this option is to remove the need for any in flight refueling for the first crewed landing.

There is going to be some stiff competition between SpaceX and Blue Origin. Either might win.

New space stations

The International Space Station will be at end of life in 2030 after continuous human habitation for almost thirty years. The other space station currently in orbit is the Chinese Tiangong station.

Expect to see a real pick up in the building of space stations over the next few years, in anticipation of the end of the ISS.

The Russian Orbital Service Station (ROS) is scheduled to begin construction, by Roscosmos, in orbit in 2027.  There is risk to this plan from the deterioration of the Russian economy.

India plans to start building their Bharatiya Antariksh Station (BAS) in 2028 and for it to be fully operational in 2035. India has had uncrewed orbital capability since 1980, and sent its first uncrewed mission to Mars in 2013. For BAS it is developing crewed launch capability. In 2025 India sent one of its own astronauts to the ISS on a SpaceX Dragon under an agreement with the company Axiom.

A consortium of countries (US, Canada, Japan, European Union, and the United Arab Emirates) are collaborating on building the Lunar Gateway, a space station orbiting the Moon. Launch of the first module is scheduled for 2027 on a SpaceX Falcon Heavy. Blue Origin is competing for additional components and launches for the Gateway.

A host of private companies plan on launching smaller private space stations in the near term, with one claiming it will do so in May 2026.

This is going to be an active frontier, and may lead to more humans going on orbital flights than the current status quo of about 28 per year.

Addendum

I have had a long warm personal relationship with Jonathan Hurst, Co-founder and Chief Robot Officer at Agility Robotics, back from pre-Agility days when he was first a junior faculty member at Oregon State University.

On January 2^nd I received an email from him responding to my twelve points about what one has to believe for deployment of dexterous humanoid robots that can replace any physical human capability, all in the context of what Agility says and delivers. I have included them below, as he wrote them. Clearly Agility and its Digit robot is not in the hype class that I am pushing back on — they are promising much less than general purpose human replacement. At the same time Jonathan believes that they are following a much longer road to general purpose humanoid robots that starts with much simpler applications. I do not think that they will end up at this lofty goal in even Jonathan’s lifetime (and he is much younger than me!). The two linked videos below are from Jonathan and accurately show the current capabilities, including that the people and robots are separated for safety, and that they have a use case and solution that does not rely on dexterous hands. I think this fits, directionally at least, with the last sentences of my September post on humanoid dexterity and safety, and included above in the new predictions segment, where I said: There will be many, many robots with different forms for different specialized jobs that humans can do. But they will all still be called humanoid robots.

From Jonathan Hurst:

1. Digit has moved over 140,000 totes to date, and has been paid to do it with an ongoing commercial contract. Our customers include GXO, Amazon, Shaeffler, and growing. The numbers are still small, limited by the safety case, and we will have a safety solution for our initial applications deploying in 2026.
2. Digit has grippers that are sometimes criticized as “not dexterous hands,” yet they are the only manipulator I’m aware of that can power-grasp and lift 25kg totes, among other things.  In 2026, Digit will have tool-change capabilities, so we can use the right tool for the right application, and deploy increasingly capable and general manipulators.
3. We never, ever claim that we will learn human-like dexterity anytime soon. I claim that there are no manipulators in existence that are even physically capable of it, assuming we even know how to control them. I have opinions on how to get there, and we have work in progress, but it’s a long road, and I think Ken Salisbury was farther along that road in 1985 than most people are today.
4. Agility regularly shows fully autonomous operation, including live on stage, including live at trade shows, including deployed with customers. Here’s a 6-hour video from Modex.
5. We have no near-term plans for working in customer homes because that is not a realistic goal. I have stated publicly that humanoids aren’t going into homes for a decade or more, because the home is the highest bar for complexity, safety, and cost sensitivity. But that doesn’t mean there aren’t a lot of other markets to enter first, on the long road into the home.
6. We have a clear safety strategy which specifies that the robots will stop moving and sit down before a human can physically touch them. They are still useful in human workflows and human spaces. It’ll be a long road to operating safely in close proximity to humans. We are collaborating to develop a new international safety standard, ISO 25785-1, which will specify how dynamically stable (i.e. balancing) robots can safely operate.
7. Digit has long been able to get up after falling.
8. Digit’s battery life is 90-120 minutes while doing heavy work, easily hitting 4hrs if it’s not working too hard. And Digit V5 will charge in 7–10 minutes.
9. Digit has been able to recharge itself for some time now on its own charging dock.
10. Agreed: humans cannot touch the humanoid safely for some time.
11. We have long said we build “Human-centric, multi-purpose” robots, and that the long road to generality is achieved by doing one thing well, then two, then five, and working our way to clear, metric-driven performance on many different tasks.
12. We’re being real. Jennifer Hunter, our COO, deployed Kiva robots at Amazon. Pras Velagapudi, our CTO, deployed many hundreds of robots with Berkshire Grey. Melonee Wise was a helpful influence, sharing her experience deploying Fetch AMRs safely around people. We know what it takes to build and deploy robots.

There is a lot of hype around humanoid robots, much like there is around AI — and also like AI, there is a lot of near-term value to gain, for those who are strategic about it.

On March 26^th I skeeted out five technology predictions, talking about developments over the next ten years through January 1st, 2036. I’ll incorporate these new predictions into my scorecard posts (every Jan 1st) and see how accurate they have been over that time period. Note that none of these predictions are saying what might happen after the 10 years are up.

1. Quantum computers. The successful ones will emulate physical systems directly for specialized classes of problems rather than translating conventional general computation into quantum hardware. Think of them as 21st century analog computers. Impact will be on materials and physics computations.

2. Self driving cars. In the US the players that will determine whether self driving cars are successful or abandoned are #1 Waymo (Google) and #2 Zoox (Amazon). No one else matters. The key metric will be human intervention rate as that will determine profitability.

3. Humanoid Robots. Deployable dexterity will remain pathetic compared to human hands beyond 2036. Without new types of mechanical systems walking humanoids will remain too unsafe to be in close proximity to real humans.

4. Neural computation. There will be small and impactful academic forays into neuralish systems that are well beyond the linear threshold systems, developed by 1960, that are the foundation of recent successes. Clear winners will not yet emerge by 2036 but there will be multiple candidates.

5. LLMs that can explain which data led to what outputs will be key to non annoying/dangerous/stupid deployments. They will be surrounded by lots of mechanism to keep them boxed in, and those mechanisms, not yet invented for most applications, will be where the arms races occur.

[You can follow me on social media: @rodneyabrooks.bsky.social and see my publications etc., at https://people.csail.mit.edu/brooks]

This is my seventh annual update on how my dated predictions from January 1^st, 2018 concerning (1) self driving cars, (2) robotics, AI , and machine learning, and (3) human space travel, have held up. I promised then to review them at the start of the year every year until 2050 (right after my 95^th birthday), thirty two years in total. The idea is to hold myself accountable for those predictions. How right or wrong was I?

I have decided to change my rules for myself a little bit after this year, in response to the many many people who have said how much they enjoy seeing my updates.

My predictions were mostly for the first few years, and by next year the density of due dates will be very low.  So, on the eight anniversary of my first set of predictions, i.e., a year from today, I will be making a new set of predictions centered on the period January 1^st 2026 to January 1^st 2036, and that will give a new density of predictions where there will be real meat to see how accurately they turned out.

What I Want to Achieve and a Changing Hype-driven Landscape

The level of hype about AI, Machine Learning and Robotics completely distorts people’s understanding of reality. It distorts where VC money goes, always to something that promises impossibly large payoffs–it seems it is better to have an untested idea that would have an enormous payoff than a tested idea which can get to a sustainable business, but does not change the world for ever. It distorts what young researchers work on as they do not want to be seen as old fashioned even when the current hyped topic is sort of dumb–soon the dumbness is forgotten and the heat of the chase becomes all. It distorts what people think they need to get a degree in at college in order to have good career prospects.

I want people to use rational thought processes when they hear about hyped ideas and be able to assess what is really going on, and what is just plain (to use the technical term) bullshit.

My Color Scheme and Past Analysis

The acronyms I used for predictions in my original post were as follows.

NET year means it will not happen before that year (No Earlier Than)
BY year means I predict that it will happen by that year.
NIML, Not In My Lifetime, i.e., not before 2050.

As time passes mentioned years I color then as accurate, too pessimistic, or too optimistic.

This year I have added hemming and hawing. This is for when something looks just like what I said would take a lot longer has happened, but the underlying achievement is not what everyone expected, and is not what was delivered. This is mostly for things that were talked about as being likely to happen with no human intervention and it now appears to happen that way, but in reality there are humans in the loop that the companies never disclose. So the technology that was promised to be delivered hasn’t actually been delivered but everyone thinks it has been.

I have not changed any of the text of the first three columns of the prediction tables since their publication on the first day of 2018. I only change the text in the fourth column to say what actually happened.  This meant that by two years ago that fourth column was getting very long and skinny, so I removed them and started with fresh comments last year. I have kept last year’s comments and added new ones, with yellow backgrounds, for this year. If you want to see the previous five years of comments you can go back to  the 2023 scorecard.

Overview of changes this year

There has been a lot of activity in both self driving cars (the demise of Cruise a big push by Waymo to scale human assisted deployments, and lots of smoke and mirrors from an electric car company) and in AI, where robotics has been pulled into the ultra hyposphere while in generative AI the end of scaling and the introduction of inference mechanisms (!!) have been hotly announced and disputed.  The human spaceflight endeavor, as it did last year, has crawled along and again has stretched out dates that were probably too optimistic in the first place.

But First.

<rant>

We all know about FOMO, Fear Of Missing Out. In late 2023, for a talk on generative AI that I gave at MIT, I coined another acronym,  FOBAWTPALSL, Fear Of Being A Wimpy Techno-Pessimist And Looking Stupid Later. Perhaps that one is a little bit too much of a mouthful to catch on. These two human insecurities lead people to herd-like behavior in establishing and propagating the zeitgeist on almost any topic.

They lead to people piling on the hype fiestas, rushing to invest (money, effort, or hope) in marginal ideas once they have become a little bit popular, or believing our airspace is being invaded by foreign drones.

“Mounting evidence, and lack thereof, suggests that perhaps the whole craze has been a sort of communal fever dream fueled by crowd mentality, confirmation bias and a general distrust in all things official.”

That quote is from the drone story linked to above, but it could well as been about the hype that we are moving towards AGI (Artificial General Intelligence).

I want to be clear, as there has been for almost seventy years now, there has been significant progress in Artificial Intelligence over the last decade. There are new tools and they are being applied widely in science and technology, and are changing the way we think about ourselves, and how to make further progress.

That being said, we are not on the verge of replacing and eliminating humans in either white collar jobs or blue collar jobs. Their tasks may shift in both styles of jobs, but the jobs are not going away. We are not on the verge of a revolution in medicine and the role of human doctors. We are not on the verge of the elimination of coding as a job. We are not on the verge of replacing humans with humanoid robots to do jobs that involve physical interactions in the world. We are not on the verge of replacing human automobile and truck drivers world wide. We are not on the verge of replacing scientists with AI programs.

Breathless predictions such as these have happened for seven decades in a row, and each time people have thought the end is in sight and that it is all over for humans, that we have figured out the secrets of intelligence and it will all just scale.  The only difference this time is that these expectations have leaked out into the world at large. I’ll analyze why this continues to happen below in the section on AI and ML.

Here is a list of some of those hype cycles that I, personally, have perceived and lived through, as taken from my presentation at MIT in late 2023 that I referenced above re FOBAWTPALSL.

Really, was there really hype about all these things?  Yes, there was, within the circles that cared. Those circles have gotten wider and wider and when reigning world chess champion Garry Kasparov was beaten by I.B.M.’s Deep Blue computer under tournament conditions in 1997 it was widely reported in the popular press, And it was declared that it was all over for humans.

Back in February 2011 a computer program named Watson played on the television game show Jeopardy against all time human champions. John Markoff, legendary technology reporter at the New York Times, wrote stories about this the day before the competition, and the day after, when Watson had indeed beaten the humans, with the same questions (fed as text to it as the same time as the humans heard the questions) all running on a cluster of machines not connected to an outside network. Here are three successive paragraphs from the second of those stories.

For I.B.M., the future will happen very quickly, company executives said. On Thursday it plans to announce that it will collaborate with Columbia University and the University of Maryland to create a physician’s assistant service that will allow doctors to query a cybernetic assistant. The company also plans to work with Nuance Communications Inc. to add voice recognition to the physician’s assistant, possibly making the service available in as little as 18 months.

“I have been in medical education for 40 years and we’re still a very memory-based curriculum,” said Dr. Herbert Chase, a professor of clinical medicine at Columbia University who is working with I.B.M. on the physician’s assistant. “The power of Watson- like tools will cause us to reconsider what it is we want students to do.”

I.B.M. executives also said they are in discussions with a major consumer electronics retailer to develop a version of Watson, named after I.B.M.’s founder, Thomas J. Watson, that would be able to interact with consumers on a variety of subjects like buying decisions and technical support.

My personal experience at that time was people I did not know, but who had heard about my role at MIT (as director of the MIT AI Lab, and then founding director of MIT CSAIL, the Computer Science and Artificial Intelligence Lab) would come up to me and ask about the future of medicine. The people were variously doctors or health industry executives. I reassured them that medicine as we knew it then would stay much the same and was not about to be rendered obsolete.

And then in 2016 Geoff Hinton, one of the key architects of Deep Learning (which has had undeniable impact on the world) said:

“People should stop training radiologists now. It is just completely obvious that within five years deep learning is going to be better than radiologists.”

More people asking me whether this was true. It wasn’t in five years and it isn’t now. We need more radiologists than ever. And yes they do use deep learning tools to help them see some things they wouldn’t otherwise see. But they also understand anomalies using causal reasoning and we would be in a sorry state if all radiology was done by programs today.

Now look at those plum colored paragraphs above again as you take yourself way back in time to a year or so ago when ChatGPT was just a baby AGI, You can find stories just like this one if you substitute “ChatGPT” for “Watson” and “Microsoft” for “I.B.M.”

The things confidently predicted in 2011 (and in 1979, and in 2016) about the end of doctors didn’t happen then and it is not happening now. Nor are all the other jobs ending.

Today I get asked about humanoid robots taking away people’s jobs. In March 2023 I was at a cocktail party and there was a humanoid robot behind the bar making jokes with people and shakily (in a bad way) mixing drinks. A waiter was standing about 20 feet away silently staring at the robot with mouth hanging open. I went over and told her it was tele-operated. “Thank God” she said. (And I didn’t need to explain what “tele-operated” meant). Humanoids are not going to be taking away jobs anytime soon (and by that I mean not for decades).

You, you people!, are all making fundamental errors in understanding the technologies and where their boundaries lie. Many of them will be useful technologies but their imagined capabilities are just not going to come about in the time frames the majority of the technology and prognosticator class, deeply driven by FOBAWTPALSL, think.

But this time it is different you say. This time it is really going to happen. You just don’t understand how powerful AI is now, you say. All the early predictions were clearly wrong and premature as the AI programs were clearly not as good as now and we had much less computation back then. This time it is all different and it is for sure now.

Yeah, well, I’ve got a Second Coming to sell you…

</rant>

Self Driving Cars

As with flying cars the definition, or common understanding, of what self driving cars really means has changed since my post on predictions seven years ago.  At that time self driving cars meant that the cars would drive themselves to wherever they were told to go with no further human control inputs.

Now self driving cars means that there is no one in the drivers seat, but there may well be, and in all cases so far deployed, humans monitoring those cars from a remote location, and occasionally sending control inputs to the cars. The companies do not advertise this feature out loud too much, but they do acknowledge it, and the reports are that it happens somewhere between every one to two miles traveled. These inputs are not direct control of the normal human mechanism of control the steering wheel, the brakes, and the accelerator.  Rather they are advice that overrides some of the algorithms.  For instance, “steer out into the next lane and go around this truck” as the human realizes that the truck is just not going to move (see an anecdote below on the first night I took the new Waymo taxis in San Francisco (I had previously last ridden a Waymo in 2012 in Mountain View)).

Why is this difference important?  One of the motivations for self driving cars was that the economics of taxis, cars that people hire at any time for a short ride of a few miles from where they are to somewhere else of their choosing, would be radically different as there would be no driver. Systems which do require remote operations assistance to get full reliability cut into that economic advantage and have a higher burden on their ROI calculations to make a business case for their adoption and therefore their time horizon to scaling across geographies.

But wait, you might say, isn’t that electric car company that used to be based in California and is now based in Texas going to roll this out imminently and have a fully digital taxi service. They demoed it on a Hollywood movie studio lot just this year, and the cars were painted gold. Hmm. The location of the demo and the fact that the cars, even down to the tires, were painted gold tells you everything you need to know. Both the cars and the humanoid robots at that event were presented as autonomous but in reality they were all tele-operated directly by people (see below in the humanoid section for more details). And that same electric car company is actively hiring people into paying jobs as remote operators.

There was a reasonably balanced appraisal from Reuters just after the event, though it does not go into details of the demos. Here is a direct quote from the story:

“We do expect to start fully autonomous unsupervised FSD in Texas and California next year.” Musk said.

The astute reader will note that this is the 11^th year in a row that the CEO of Tesla has made this prediction of the same milestone happening the next year. We can admire the consistency.

Actual self-driving is now generally accepted to be much harder than every one believed.

The reason that this bait and switch is important to understand is that the promise of inevitable fully self driving technology upended a historical way that new transportation systems have been adopted.

In the past whenever we have introduced new transportation mechanisms there have been large investments in infrastructure and that infrastructure is shared and used by everyone. The Romans built roads so soldiers and traded goods could travel long distances–in Europe those road networks are still the basis of today’s road networks. When steam engine driven trains were the new transportation technology vast networks of rails were built allowing goods to move long distances in mere hours or days. When Ford started mass production of automobiles he built roads and the local governments followed and the the Federal government followed, and those roads are what we use today.

Actual fully self driving cars promised that no infrastructure changes would be needed to revolutionize how vehicles would be controlled. Each individual vehicle would do what was needed all by itself. As sensors and networks got better there was no need for expensive new infrastructure because of this promise.

The promise was false. If government and private partnerships in building smart roads, which was a hot topic in the 1990s. had continued, every one of us would now have smarter safer cars, but still with onboard human drivers taking over in many situations. But we would have had smart freeways where once you were on it your car would be self driving. The road would have had lots of sensors effectively shared across all cars, as that data would have been transmitted to all passing cars. It would have been a fraction of the cost per car compared to the sensing on today’s almost but not really self driving cars like those of Waymo. And we would have had much more accurate congestion data where the root causes of local congestion would have been sensed with semantic understanding rather than just inferring it from the aggregate collection of location data from phones, individual cars, and historical data from roadside sensors.

Instead we now have individual corporate actors using a mixture of partial self driving and remote human supervision. The big question is whether the economics of this works at scale, and whether the fake promises will drive out the human drivers in cheaper services and we’ll all end up paying more. Will the level of hype we saw push our decentralized transportation system into the hands of a few wealthy companies, and in effect make it a centralized system where everybody has to pay private companies to be part of it?

As a reminder of how strong the hype was and the certainty of promises that it was just around the corner here is a snapshot of a whole bunch of predictions by major executives from 2017.

I have shown this many times before but there is one new annotation here for 2024. The years in parentheses are when the predictions were made. The years in blue are the years are the predicted years of achievement. When a blue year is shaded pink it means that it did not come to pass by then. The predictions with orange arrows are those that I had noticed had later been retracted.

The prediction that Jaguar and Land-Rover made that they would have fully autonomous cars by 2024 did not come to pass, so I have shaded it pink,

Note that every single blue year up until now is shaded pink, and that every one that is shaded pink has still not come to pass. None of the predictions that were out there in 2017 for the next few years have happened.  None. There are three more for 2025, and I am sure that a year from now they will all be shaded pink also.

One of the big selling points of self driving cars was that they would be safer than cars driven by humans. So far that is not holding up with real data. One electric car maker with self driving software had it disengage when it sensed there would be an accident, supposedly so that the human could take over in a split second. And then the company did not report the incident as the fault of the software as it was no longer controlling the car when the impact occurred. It was reported, and I had this experience myself in my last ride in a Cruise in 2023, that Cruise vehicles would freeze when an accident looked likely, and then not report it as their software’s fault as the car was stationary and was hit by another car. In many reported cases, and in my case, simply continuing to move forward would avert any likely accident (fortunately for me the human driver of the other car slammed on the brakes and did not hit my robot vehicle).

In this story from the Washington Post about Federal investigations into the safety incidents with self driving cars, they report that the companies involved claim they have vast amounts of driving on our roads under their belt. Not so.

An industry association says autonomous vehicles have logged a total of 70 million miles, a figure that it compares to 293 trips to the moon and back. But it’s a tiny fraction of the almost 9 billion miles that Americans drive every day. The relatively small number of miles the vehicles have driven makes it difficult to draw broad conclusions about their safety.

To put that into perspective, the total number of miles driven by all autonomous (sort of) vehicles over the last decade is less than 1% of the miles driven by humans every day in the United States. It is a tiny, tiny portion.

Take a look at this embedded video from the Wall Street Journal about investigations of crashes (many of which have been fatal) involving autonomous driving systems.

From the audio: “The kinds of things that tend to go wrong with these systems are things like it was not trained on, pictures of an overturned double trailer. It just didn’t know what it was. There were some lights there, but the lights were in unusual positions. A person would have clearly said something big is in the middle of the road. But the way machine learning works is it trains it on a bunch of examples and if it encounters something it doesn’t have a bunch of examples for it may have no idea what’s going on.”

[[My own take is that the fetish of end to end learning leads people to leave out well known algorithms that might solve many of  these problems (e.g,, the incredibly simple time to collision algorithms based on looming). Yes, end to end learning made speech understanding systems better, but that does not mean it is the appropriate fetish to apply everywhere.]]

Pro tip: Think about this history of industry prognostications about fully autonomous driving being just around the corner when you read today’s prognostications about LLMs taking jobs, en masse, in the next couple of years, or humanoid robots being dirt cheap and being able to learn how to do any human manual task real real soon now. You know you have seen this movie before…

My own experiences with Waymo in 2024

I have two sorts of experiences with Waymo vehicles. First, as a driver of my own vehicle and sharing road space with them every single time that I drive. And second, as a user of their ride service.

The streets of San Francisco had been thick with Waymo vehicles with no driver in them especially in the second half of 2024. As I drive across the city every morning to head down to my robotics/AI startup half way down the peninsula I see them everywhere until I get on to 101.  I see them in front of me and behind me and in adjacent lanes as I drive on multilane one way streets. Sometimes I see four of them in a single block. Twice I’ve seen four of them in a line, in my block and could see four of them in a line in the block ahead of me.  When I am at four way intersections with no traffic lights I see them participating in the social ritual of taking your turn to drive through the intersection in the order you stopped, except when a pedestrian is crossing in front of you. They do that pretty well. They do less well when they accidentally get into a line of parents’ cars snaking around a corner for school drop off or pickup.

Over the last few months I have noticed that in general they are getting more aggressive about stretching the rules, just like people do. Otherwise human drivers (including me) take advantage of their politeness. That aggression is not always welcomed. One morning I saw a workman with a group doing some digging on a road, and holding a sign with SLOW on one side and STOP on the other side have to jump in front of a Waymo to get it to do what he was trying to tell it to do with the sign. STOP. It wasn’t stopping for no stinking sign!

The only time I have seen a Waymo go into reverse, ever, was when I was illegally driving the wrong way down a single lane street and we were heading straight at each other.

As a rider I feel they are not quite aggressive enough with human drivers some time, so a ride in a Waymo takes longer than with an Uber or Lyft.

It is hit and miss where they drop me off. Sometimes they take a place to pull over half a block from my house, even when it is raining. There is no way to adjust what they happen to decide that day, even though I know that they will always be able to pull in right in front of my house.

The first time I took a Waymo this year, on the way home it picked me up at a restaurant and then was about to make a right turn. But at that corner there was an 18 wheeler with its lights flashing and surrounded by green cones. It pulled right in behind that truck and waited a long time before it drove forward. I am guessing a remote operator intervened told it to go around because eventually it pulled around it in the lane just to the left. Based on seeing Waymos interact with orange cones I suspect it would have done better if the cones had been orange rather than green.  This easily illustrates that the learning that this robot does, and indeed any robot does, is nothing like the learning that people do (see my rant about the seven deadly sins and mistaking performance for competence in the section below on advances in AI and ML).

I mostly feel safe when I am a passenger in a Waymo.  Sometimes I don’t feel that my driver of an Uber that I am taking rides with Uber that are not as safe as I would prefer.

Self Driving Taxi Services

There have been three self driving taxi services in the US in various stages of play over the last handful of years, though it turns out, as pointed out above that all of them have remote operators. They are Waymo, Cruise, and Zoox.

Waymo and Cruise are similar in that they use conventional cars adorned with lots of sensors. Zoox has purpose built vehicles that have no steering wheel or pedals for brake or accelerator.

Waymo and Cruise went for deployments in large parts of two or more cities and have had ride services callable by apps, just as one can do with Uber or Lyft. Zoox is smaller scale, much more restricted in geography, and really not comparable.

At this time last year Cruise was in trouble has it had suspended all of its San Francisco operations under pressure from regulators after some bad accidents that happened in a way that never would happen for human driven cars.  Briefly, their cars were getting hit at night by emergency vehicles with lights flashing as the Cruise cars crossed intersections. Human drivers see the reflections of lights from such vehicles flashing even if they don’t see the vehicles themselves. The Cruise vehicles were only reacting to flashing lights that they could perceive directly. But the accident that tipped the scales was when a pedestrian crossing in front of a human driven vehicle was hit and went flying in the air landing right in front of a Cruise. The Cruise hit the person (who now disappeared from sight) as a human driver would most likely have done. But then it proceeded to drive 20 feet with the human underneath the vehicle being dragged along as it went into a mode where it was supposed to get off the road. A human driver would not have reacted that way to having been in a collision, even if it was not their fault.

The hammer finally fell in December of 2024. General Motors shut down Cruise. The leading paragraphs from this linked story from the Wall Street Journal are:

While there are 40 companies that have permits to test autonomous driving in California, alone, the demise of Cruise leaves just one company, Waymo, trying to make an actual go of a digital taxi service in the United States. They have an enormous significant lead over anyone else who wants get into this business and have spent billions of dollars (probably very much north of $10 billion) on this endeavor over the last 15 years. In an email they sent me a couple of weeks ago as a user of their services they reported that they provided 4 million customer rides in 2024. That is approximately 4 million more than any other company in the United States.

Waymo

Despite being so far out in front it has not been all smooth sailing for Waymo.

Early in the year the operations center for Waymo somehow neglected to realize it was Chinese New Year in Chinatown in San Francisco. So Waymo vehicles were routed through that area on the biggest night of celebration. Any human driver would have realized that the streets, i.e., the street surfaces where cars usually drive, were completely packed with humans, no doubt some of whom were intoxicated as well as just being out having a good time. Not so the Waymo vehicles. They tried pushing through the very very dense crowds, no doubt annoying many people. And what do people have at Chinese New Year?  Fireworks. So some revelers decided to push back on this robot car invading their space. Here are a couple of pictures of the results.

Not pretty.  And an example of how taking away people’s agency is never a good idea for robots (see my second law of robotics).

Throughout 2024 Waymo has been investigates for various accidents such as those described in this Wall Street Journal article. “Reports included collisions with stationary or semistationary objects, such as gates, chains or parked vehicles, according to the regulator.”

In the middle of the summer Waymo added a feature where they would honk their horns at cars in their way. But this backfired when hundreds of Waymos were coming back to their parking lot in the very early hours of the morning, and they started honking at each other and waking up human neighbors. Eventually that got fixed.

In late September a motorcade for Kamala Harris in San Francisco was brought to a halt by a Waymo that stopped in the middle of California Street doing a U-turn in front of it. I’m sure this incident was of great concern to the Secret Service. Eventually a San Francisco police officer got into the car and drove it out of the way–this is shown in a video included with the story above. I do not know how the officer got access to the vehicle and whether Waymo remote operations were cooperating.

More disturbingly humans outside the Waymos started harrassing humans inside them. The most concerning cases come from the realization that if a woman is in a Waymo at night she will be dropped off, outside, on a public road at the end of her journey with no option but to get out of the car where it has stopped. So groups of men have followed Waymos with women in them and then harassing the woman when she gets out. If she was driving her own car she might be heading to an off road parking space or she might choose not to stop if she knows she is being followed. There are no such options in a Waymo so taking a Waymo at night is less safe than other means of transportation–just follow it and eventually the preyed upon woman will have to get out. Here is a very recent disturbing story about this practice.

Meanwhile Waymo managed to raise $5.6B to expand to new cities in 2025. It already operates in parts of San Francisco, Los Angeles, and Phoenix. The new money will let it expand to Austin and Atlanta in the United States and to start operating in parts of Tokyo in Japan. That is expensive expansion.

Here is the question for the future of watered down remote monitored “autonomous” driving systems (let’s call it “watered down autonomy”), and it is up to Waymo now. Can Waymo expand fast enough in these new markets in 2025 and take enough business from what is left of traditional taxi operators, along with those operating under the Uber and Lyft models, and do it in a way which is in sight of profitability, so that it has a case to raise the stupendous amounts of money needed to operate in all large cities in the US in the next 10 t0 20 years?

If Waymo can not succeed at this in the next two years I think the idea of large scale use of watered down autonomy will be dead for at least a decade or two. Right now full autonomy everywhere is already dead.

Electric Cars

Last year US manufacturers pulled back on their planned production of EVs. In data from this report we can see that sales dropped at the start of 2024 but have now picked up again.

There is steady growth in sales but my prediction of 30% of US car sales being electric by 2027 now seems wildly optimistic. We need two doublings to get there in three years and the doubling rate seems more like one doubling in four to five years.

Note that some sources include hybrids and hydrogen powered cars in electric vehicles but I am using the battery electric vehicle (BEV) numbers.

To see how the trends are across brands you can see a breakout for Q2 of 2024 here.

There appear to be two main headwinds for BEV adoption. Firstly, if one doesn’t have on property residential parking it is hard work in the US to find a place to recharge, and it takes hours for the charging to finish. This will stop many city dwellers from adopting. Secondly the increased tire wear adds up to real money. The maintenance requirements for BEVs are much less than for cars with an internal combustion engine. On the other hand tires do not last as long (I have had to buy four new tires in less than two years owning my first BEV), apparently due to the increased weight of the car.

Flying Cars

Flying cars are another category where the definitions have changed. Back when I made my predictions it meant a vehicle that could both drive on roads and fly through the air.  Now it has come to mean an electric multi-rotor helicopter than can operate like a taxi between various fixed landing locations. Often touted are versions that have no human pilot. These are known as eVTOLs, for “electric vertical take off & landing”.

Large valuations have been given to start ups who make nice videos of their electric air taxis flying about. But on inspection one sees that they don’t have people in them. Often, you might notice, even those flights are completely over water rather than land. I wrote about the lack of videos of viable prototypes back in November 2022.

Nevertheless there have been wild predictions.  I ended a longer version of this component in last year’s annual review with:

Also note the size of this vehicle. There are many fossil fuel powered helicopters that are much smaller. This is not going to be a personally owned vehicle for the masses.

Don’t hold your breath. They are not here. They are not coming soon.

Nothing has changed. Billions of dollars have been spent on this fantasy of personal flying cars.  It is just that, a fantasy, largely fueled by spending by billionaires.

Robotics, AI, and Machine Learning

So what happened in Robotics, AI, and Machine Learning this year?

Many, many, many people got just a little bit over excited. That’s what happened.

There have been a lot of party tricks and it is the researchers who often play the tricks on themselves without realizing it. This is not new, none of it is new. But there are orders of magnitude more people watching it now, and more people are out to make a buck by being hypesters, promising riches to those who will invest in their irrationally overpriced companies.

How could this be?

We are seeing mass sinning, lots and lots of people committing some of the seven deadly sins of predicting the future of AI  which I wrote about back in 2017 here (or here you can see a professionally edited version of that blog post of mine).

Four of those seven sins seem most relevant to today’s hyped up atmosphere around robotics, AI, and machine learning.

Here now are short descriptions of these particular four sins, edited down from my earlier much more detailed descriptions. Then I will weave them together to explain how it is still pretty much business as usual, and I mean that in a good way, with steady progress on both the science and engineering of AI.

Performance versus Competence

One of the social skills that we all develop is an ability to estimate the capabilities of individual people with whom we interact. We use cues from how a person performs any particular task to estimate how well they might perform some different task. We are able to generalize from observing performance at one task to a guess at competence over a much bigger set of tasks.

These estimators that we have all inherited or learned do not generalize well to other creatures or machines. We are not good at guessing which smart things other species might be able to do, and we are not good at guessing what an AI system can do when we have seen it do a few tasks in a limited domain. We get it wrong all the time.

Indistinguishable from Magic

When people cannot explain how something works they cannot know its limits as they do not have any sort of model (nor have they seen enough examples of it before). Arthur C. Clarke said that any sufficiently advanced technology is indistinguishable from magic.

In our minds UFOs can do all sorts of amazing things as we have no way of knowing their limits–they may as well be magic, And that is what they become in speculation about them.

Isaac Newton spent half his working life on alchemy as he did not know that the nucleus of atoms were not subject to mere chemistry. He would have been just as ignorant of the limitations of an iPhone screen (different sort of apple…), despite his own ground breaking work in optics. Remember, he was a really really smart dude. But even he was not able to develop all the theories needed to understand the world around him, despite his successes with calculus and gravity and the makeup of white light. He attributed properties to chemistry that were way beyond its limits.

Exponentialism

We have just lived through sixty years of the most phenomenal growth of a technology in the history of humankind. It is the story of silicon-based computation. Everyone has some idea about Moore’s Law, at least as much to sort of know that computers get better and better on a clockwork like schedule.

This reality has trained people to think that probably a lot of other things in tech will change exponentially, especially when that thing has a strong computational component. The sin of exponentialism is to argue that some other process is going to follow a Moore’s-like law when it is unwarranted to so argue.

Moore’s law worked for so long because in the starting technology of the 1960s the currents used to represent digital information were many many orders of magnitude beyond the minimal physical limit needed to determine whether they  were present or not, and hence distinguish a 1 from a 0. Those currents could be halved many times without breaking physics limits.

Speed of Deployment

New technologies get deployed much more slowly than people imagine. Even software technologies.

The old internet protocol, IPv4, can only address two billion, or 2×10⁹, devices, which is way less than the number of people on our planet. A new protocol, IPv6, which can address more than 3×10³⁸ devices was meant to replace it over a two year period of dual use by about 2003. But in 2024 IPv4 was still there and carrying over half the world’s internet traffic despite its inadequacies.

Must functioning businesses that operate in the physical world are very averse to taking up new technology as it dramatically increases existential risk to their business. They must foresee immediate and incredibly high return on investment (ROI) to be tempted to move to new technologies.

Even the military is slow to adopt new technologies. The US Air Force still flies the B-52H variant of the B-52 bomber. This version was introduced in 1961, making it 63 years old. The last one was built in 1963, a mere 61 years ago. Currently these planes are expected to keep flying until at least 2040, and perhaps longer–there is talk of extending their life out to 100 years.

What does this all mean?

Right now there is incredible hype for both Large Language Models (LLMs), and all their variations, and for humanoid robots, especially humanoid robots that are going to learn how to do things.

The hype is driven by the four sins above.

LLMs

LLMs have proved amazing facile with language. They have been trained on pretty much all the text that is available on the Web and all the digitized historical books that exist. Miraculously LLMs seem to be able to infer a representation of some sort, that is somewhat independent of the particular human language that they read. So they are able to translate between human languages, and when you ask them just about anything they produce text in the language that you asked in, and that text often seems entirely reasonable and informative.

I used the word “miraculously” as we do not really understand why they are able to do what they do. We, of course, know that the architecture for them is built around noticing correlations in vast amounts of text  that connect some tens of thousands of tokens which are the components of words in each language that is digested. It is a surprise that they work as well as they they do, and produce coherent sounding language on just about any topic.

Here is the original architectural diagram from the 2017 Attention Is All You Need paper:

Each column from bottom to top is a pure feed forward network, with no search, no iteration, no conventional algorithm at all. There are inputs at the bottom and then layer upon layer of linear neurons that have numbers or weights stored in them that multiply and add their inputs and threshold that sum to provide an output. The detail in the architectural diagram is how the connections between layers are organized.

On the left is an input or question, in a linear string of words, from a user. That gets injected half way up the network on the right and remains constant while a single iteration process runs. The stack on the right outputs a word (or token) and that gets fed back to the bottom of that stack, and a new token pops out the top. All the output tokens that have so far been produced remain in the right bottom input buffer as ordered input.

What the network has been trained to do, is given the user input on the left, and what the network has output so far, choose a very likely next word, given the billions of examples it has seen in training. Some randomness is used to choose among a small number of very likely next words at each stage.

There are hundreds of billions of weights that get learned and stored in the layers of network to act as multipliers for each individual input to each layer.

So now us humans are faced with looking at this system running and our human nature just makes us commit the first two sins from above.  It is in our nature and we cannot help ourselves.

First, we see really impressive examples of responses to input questions, and if a human was giving those answers we would estimate that person to be quite clever and able to reason. Often though, because they have so many billions of examples on which they were trained LLMs are essentially looking up the question in the weights. The weight if gained from all of human knowledge that is out there on the network in language form. Invisibly the network is perhaps (but not in any intentional way) merging some similar questions, and then merging the answers which were already in the vast data that it has seen.

But us dumb humans just think the damn thing is really really smart.

Then, since we don’t have a real explanation in our heads for what it is doing we start thinking it is magic, and that there is no real limit to what it is extracting from all that data (that it used a significant portion of the energy budget for many different countries to compute) and how general its capabilities will be. It becomes magic. And then researchers try to show that it can reason, that it has inferred a spatial understanding of the world, that language can be used to do all sorts of things that Moravec’s paradox tells us it can’t. There is a lot of magical thinking that humans do about LLMs.

Of course it can diagnose diseases like a doctor talking about them. Of course it can teach a student as well as a human teacher. Of course it can program as well as a human computer programmer. It is magic after all.

But in reality the fact that it is just picking likely next words means that in fact we can’t trust its output. Some outputs are great. Some are pure confabulations (most people use the word “hallucinations” for this, but I prefer “confabulations”). And we do not know which we will get ahead of time, or more perniciously how much of each we will get, trustworthy pieces of output and confabulated pieces of output all jumbled together.

Not to worry say the proponents, More learning will fix it. Fire up a nuclear power plant (I am not making this up–the tech companies are getting more nuclear power built or activated so that their LLMs can learn what a human learns using just 20 watts powering their brain; I am not confabulating this!!), and we’ll feed it more data and it will become more trustworthy.  It is magic after all. But the magic is not going as well as the proponents imagined and promised as this Wall Street Journal story explains. Their imaginations were definitely encourage by exponentialism, but in fact all they knew was that when the went from smallish to largish networks following the architectural diagram above, the performance got much better. So the inherent reasoning was that if more made things better then more more would make things more better. Alas for them it appears that this is probably not the case. But rabid exponentialists have not yet given up. Expect a bunch of VCs to adversely affect the growth of pension funds around the world as pension funds are a prime source of capital that VCs spend.

More serious academics are working on boxing in the LLMs with more external mechanism beyond just feeding the output tokens back in as a linear string of input. Many of these mechanisms look a lot like more conventional AI mechanisms, and we will see where these additions prove to be useful, how much of the wheel will be reinvented, and how long (months?, years?, decades?) to get there.

And the answers to those last questions will tell us how much sinning has been done by companies in predicting fast deployments. Back in rant at the beginning of this post I gave the example of I.B.M. and Watson and their completely optimistic predictions of how any problems of applying Watson (which seemed extremely competent based on its performance on live TV) to the real world would be solvable. The areas that it was predicted to be applicable came from magical thinking.

Surely no one today could be as dumb as that big company was back in 2011. Surely not. No, not us smart inhabitants of 2025. Its us. We are nowhere near as dumb as them!!

Humanoid Robots

The other thing that has gotten over hyped in 2024 is humanoids robots.  The rationale for humanoid robots being a thing is a product of the four sins above and I think way less rooted in reality than the hype about LLMs. In fact I think it is pretty dumb. [[I suspect many people will reason that I cannot have a valid opinion about this precisely because I happen to have built more humanoid robots than anyone else on the planet. So read ahead with caution.]]

My first law of robotics states:

The visual appearance of a robot makes a promise about what it can do and how smart it is. It needs to deliver or slightly over deliver on that promise or it will not be accepted.

The first sentence describes, I think, what is sucking people into believing that humanoid robots have a big future. It looks like a human, so its performance will be like a human, so it will be competent like a human.  It’s the performance/competence sin without even waiting for the performance part!

The second sentence describes how the humanoid fever will break, and how the hundreds of millions of dollars put into many of these companies (billions of dollars overall) will disappear. The puppets will not perform at acceptable levels. It is easy to see this as you hear all the things investors and CEOs of humanoid robots say they will be able to do. They have hardly even got to the lab demonstration phase.  My third law of robotics is:

Technologies for robots need 10+ years of steady improvement beyond lab demos of the target tasks to mature to low cost and to have their limitations characterized well enough that they can deliver 99.9% of the time. Every 10 more years gets another 9 in reliability.

For real work, robots need to operate with four, five, or six nines. We are a long way from that. The zeitgeist is that we will simply teach the robots to do stuff and then they will be able to do it.

BUT, we do not know yet whether that is going to work. In order for it to work you have to both collect the right sort of data and then learn the right things from that data. It is not at all clear to me that we know the answers to make either of those things true. I think it will be an active place for lots of good research for many years to come.

There is an excellent survey paper of current research state of the art called Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes. Unfortunately I think the title of the paper is going to confuse many people. “Real-World Successes” to someone like me, who these days deploys robots that people pay for and that provide real ROI, sounds like it is about systems that have been deployed. But on reading the paper it turns out that they mean that it is learning and demonstrations done in a lab setting on physical hardware rather than just in simulations and simulators.  And, to me the lab demonstrations are shakier (literally) than I imagined in my third law above.

I think we are a long way off from being able to for-real deploy humanoid robots which have even minimal performance to be useable and even further off from ones that have enough ROI for people want to use them for anything beyond marketing the forward thinking outlook of the buyer.

Despite this, many people have predicted that the cost of humanoid robots will drop exponentially as their numbers grow, and so they will get dirt cheap. I have seen people refer to the cost of integrated circuits having dropped so much over the last few decades as proof. Not so.

They are committing the sin of exponentialism in an obviously dumb way. As I explained above the first integrated circuits were far from working at the limits of physics of representing information. But today’s robots use mechanical components and motors that are not too far at all from physics based limits, about mass, force, and energy. You can’t just halve the size of a motor and have a robot lift the same sized payload. Perhaps you can halve it once to get rid of inefficiencies in current designs. Perhaps. But you certainly can’t do it twice. Physical robots are not ripe for exponential cost reduction by burning wastes in current designs. And it won’t happen just because we start (perhaps) mass producing humanoid robots (oh, but the way, I already did this a decade ago–see my parting shot below). We know that from a century of mass producing automobiles. They did not get exponentially cheaper, except in the computing systems. Engines still have mass and still need the same amount of energy to accelerate good old fashioned mass.

This Year’s Prediction Update

A Parting Shot

I recently read a research paper on humanoid robots working in built for human environments. It was based on the argument that the best form for a robot that is to operate in human environments is something tallish and skinny-ish, and probably dynamically balancing, with arms that can reach down to table tops etc., and with a sensor system that can look down from above, as that is what our human environments are optimized for. Here is the first paragraph of the paper:

The past decade has seen an explosion of research in humanoid robotics. The stated motivations for this work have varied widely. Many teams have concentrated on bipedal locomotion, some have been interested in human level social interactions, understanding human intelligence, modeling human learning capabilities and others have been more interested in entertainment. Some humanoid robots have had manipulation capabilities on static humanoid platforms and some of that work is aimed at dexterity, plus there has been simple two armed grasping on mobile humanoid platforms. Overall there has been very little work combining dexterous manipulation with humanoid robots, static or mobile–much of that which has appeared, has been concerned with dynamic tasks like pole balancing and juggling rather than manipulation, or has used teleoperated manipulation.

Apart from the weird references to pole balancing and juggling this all sounds pretty reasonable and consistent with what is happening today, and with recent history.  In fact this is the very first paragraph of the very first paper in the very first issue of the very first volume of the International Journal of Humanoid Robotics.

And it was published in 2004, with me as first author.  Let me spell that out in case you thought there was a typo in the year. This is from a paper that I and my students and post-docs wrote in the year two thousand and four. Here is the beginning of the contents page for that first issue.

Another pro tip: Don’t stand anywhere near a walking or balancing wheeled humanoid when they are moving or doing any task. I have had some near misses for myself with my own humanoids twenty years ago and more recently with some of the humanoids from new start ups. And more generally never be below any sort of walking robot, no matter how many legs it has, when it is walking up stairs.

HUMAN SpaceFLIGHT

The numbers of flights in 2024 was not much different from those in 2023 (I neglected to include the flights by China last year).  It does not feel like a golden age of human spaceflight, though there were other highlights from SpaceX.

Orbital Crewed Flights

Three countries put 28 people into orbit in 2024, the United States launched 16 people on five flights and Russia and China launched 6 people each with two launches. So there were nine crewed orbital flights total. Two were private and seven were government flights.

The United States: There were four US flights to the International Space Station, starting with the private Axion-3 mission with a crew of four on January 18^th. The launch vehicle for this was a SpaceX Falcon 9, and the crew vehicle was a SpaceX Dragon. The remaining US flights to the ISS were paid for by NASA. Two of them were SpaceX flights, with four people on March 4^th, the Crew-8 mission, and two people on board Crew-9 on October 25^th. The remaining US flight to the ISS was the inaugural crewed flight of Boeing’s Starliner, launched on June 5^th atop an Atlas V rocket with two people aboard. They are still stuck in space and will be for a few more months–see the section on Boeing below.

The other US mission was also a SpaceX launch and vehicle flight, this time known as Polaris Dawn. It was the second mission paid for by billionaire Jared Isaacman, with him as commander. There was a former US Air Force fighter pilot as mission pilot and two SpaceX employees as mission specialists, giving a total crew size of four. They stayed aloft for five days, launching on September 10^th, This mission flew higher above Earth than any mission since Apollo 17, the last lunar landing mission, in 1972. Two of the crew “spacewalked” with their feet inside the Dragon capsule but with their bodies outside. This was the first private spacewalk ever. Now Isaacman has been tapped by the incoming US President to be the administrator of NASA.

Russia: There were two Soyuz launches, each with three people, up and down, but different people coming back. The launch dates were March 23^rd and September 11^the. The six people that launched on Soyuz in 2024 were 3 Russian Cosmonauts 2 NASA Astronauts and one Belarusian commercial airline flight attendant who won a national competition with 3,000 applications. She was the only one not set for a long duration mission and was off the ground for slightly less than 14 days. So there were no space tourists per so, but the Belarusian flyer was most likely included as part of Russia’s efforts to keep in good favor with Belarus which has aided it in its war in Ukraine, and was certainly not part of the regular scientific program of the ISS.

China: There were two flights of  Shenzhou (a larger more modern version of Soyuz) that were crewed in 2024.  Both flights were to the Tiangong Space Station and both took along three Taikonauts, first on April 25^th and then on October 9^th.  Both crews were assigned long duration missions and now the crews are overlapping previous crews at Tiangong so it is now being continuously occupied. The first handover this year took about five days and the second about three and a half weeks.  Both times there were six Taikonauts onboard Tiangong at the same time.

Suborbital Crewed Flights

There have been two companies providing space tourism flights on suborbital flights. Blue Origin launches a capsule on top of a reusable rocket, New Shepard, and the capsule lands using a parachute and a brief rocket blast right before hitting the ground (similar to how Soyuz lands). Virgin Galactic has a winged craft which is carried aloft by a bigger a jet engined airplane, it separates at high altitude within the atmosphere and rockets into space. It flies back and lands on a runway.

Both companies are run by billionaires who made their money in other businesses.  Both billionaires have flown to space on their own craft.

Both companies have aimed to have regular launches with lots of tourists, but neither has gotten to that scale and so far only a very small number of the many people who have paid a substantial deposit have been able to fly.

Blue Origin had a failure with an uncrewed version of the vehicle in 2022 and only flew one flight in 2023 which was also uncrewed. This year they flew three crewed flights on May 19^th, August 29^th, and November 22^nd, each with six passengers (the system is automated and requires no pilots). In 2021 and 2022 they also had three flights, so there has now been nine crewed flights total. The first two took four passengers and the remaining seven have had six passengers, so altogether they have flown 50 people above the Karman line, 100 kilometers above Earth.  This is not yet a regular cadence, nor a large scale tourist business.

In 2024 Virgin Galactic had two flights, each with two crew from the company and four passengers. These flights were on January 26^th and June 8^th. Virgin Galactic flights are now on hiatus, awaiting a new bigger and better vehicle in about two years.  Virgin Galactic has had a total of twelve flights since December 13th in 2018.  Three have had two people on board and nine have had six people on board, for a total of sixty filled seats that have crossed the Karman line. The total number of different people is smaller as the two pilot seats on each flight have been occupied by a small number of people who have flown multiple times.

So, in 2024 thirty people went on suborbital flights, and altogether there have been 110 people on these commercial suborbital flights. Space tourism on suborbital flights has yet to take off in a regular or scaled way.

Boeing’s Starliner

First announced in 2010 Boeing’s Starliner was originally scheduled to fly a human crew in 2018. It carried out its second uncrewed flight in May 2022, and finally did make its first crewed flight on June 5^th. The crew of two docked with the ISS, but there were problems with multiple gas thrusters for fine motion during the docking. The original plan was that the crew would stay on the ISS for about a week and then return to Earth for a touchdown on to hard soil (as all Russian and Chinese crewed missions end along with all Blue Origin sub-orbital flights).

The option of that return was considered, but the thrusters were on a section of the vehicle which is discarded along the way before the landing so there was no possibility of getting a look at the hardware back on Earth.  So a program of tests while docked to the ISS was started delaying the crew return.

Eventually it was decided that it was too risky for the crew to return on the craft and so it returned empty on  September 7^th, landing in New Mexico. As it happened, although there were more anomalies with the thrusters the crew would have landed safely had they been on board.

Now the crew was stranded in space with no designated ride home. It was decided to remove two crew from the Crew-9 launch and have the Starliner astronauts, Barry Wilmore and Sunita Williams, fly back on that SpaceX Dragon with the other two, which after additional delays is now scheduled to happen some time in March 2025. Their one week visit to the ISS will have stretched out to nine months by then.

Boeing has committed to fixing the problems with Starliner. The boosters that it uses are no longer being built, but there are five existing ones reserved for the five additional contracted flights that Boeing has with NASA. They are supposed to happen once per year.

We do not know at this point, but I think it would not be a huge surprise if Starliner never flies again.

SpaceX Falcon 9 

Once again the Falcon 9 launch system has broken all sorts of records for number of launches and reuse.

During 2024 there were 132 single booster launches.  For two of those flights no attempt was made to recover the first stage (there is a performance penalty for the primary payload in order to recover the first stage). One attempted recovery failed when the booster (on its 23^rd flight) caught fire as it landed on the recovery barge. Another booster has since flown a total of 24 times.

In terms of mission success all but one of these flights succeeded; one failed when the second stage failed during re-ignition for adjusting the orbit.

There were also two Falcon Heavy, the three booster version, launches, both of which succeeded. One of the had successful landings for the two side boosters, but there was no attempt to recoer the central booster on that flight and no attempt to recover any of the three boosters on the other Heavy flight.

This brings the total number of launches of the single booster version to 417 along with 11 launches of the three booster Heavy version.  These numbers are way beyond the number of launches for any other orbital booster.  Additionally it is the only flying orbital system that is reusable at the moment, though  Blue Origin and Rocket Lab both plan on joining the club soon.

It is worth, once again, looking at how long it has taken to get to a total (across both single booster and Heavy triple booster versions) of 428 launches, with only three failures to deliver the payload to where it was intended to go.

The first launch occured in June 2010, and there were a total of 4 launches in the first three years.  The first successful booster recover happened on the 20th flight, in December 2015, five and a half years in. The first reuse of a booster occured in 2017, in the 8^th year of the program.

Since 2021 there has been a steady increase in the number of launches per year,

SpaceX had previously gotten satellites to orbit with its first rocket, the Falcon 1.  Falcon 9 has been a spectacular success.  But it was not instantaneous.  It took time to build from the cadence of launches, about 10 years before the hockey stick curve showed up.  Deployment is never sudden but comes after a long build.

SpaceX Starship

Starship is SpaceX’s superheavy two stage rocket, designed to put 150 tons of payload into orbit, but also be able to go to the Moon or Mars. There is the booster which is designed only to work in Earth atmosphere with 33 Raptor engines both to get the second stage high enough and fast enough and to let the first stage have a controlled return to the launch site. The second stage, called Starship, is both a booster and the payload.  It has three Raptor engines and three Raptor vacuum engines. The Raptor engines are designed to get the Starship into orbit after the first stage drops away, and to guide the Starship as it returns to its Earth launch site. The Raptor vacuum engines are meant for breaking out of Earth orbit and going to the Moon or Mars, and to do soft landings on those two bodies where there is no or almost no atmosphere.

In 2024 SpaceX made steady progress with four launches of the two stages coupled together.  The first two launches lead to both stages blowing up.

The third and fourth launches were a big improvement.  As with earlier flights they launched from the coast of Texas. In both cases the second stage did a reentry burn on it first orbit and then did a soft landing in a target zone in the Indian Ocean.  In the third flight the main booster returned to the launch site and hovered next to the launch tower betweeen two giant arms which then captured it and the engines shot down successfully. It was sifficiently damaged during flight however, that it was not reusable. In the fourth flight there were health anomalies to the first stage was ditched in the Gulf of Mexico.

On the fourth flight there was both less heat shielding and much less damage from heat during reentry. This is definite forward progress. But it is still quite a long way from both being operational and both stages being reusable. And it is even further away from being human rated.

This is the vehicle that the CEO of SpaceX recently said would be launched to Mars and attempt a soft landing there.  He also said that if successful the humans would fly to Mars on it in 2030. These are enormously ambitious goals just from a maturity of technology standpoint. The real show stopper however may be human physiology as evidence accumulates that humans would not survive three years (the minimum duration of a Mars mission, due to orbital mechanics) in space with current shielding practices and current lack of gravity on board designs. Those two challenges may take decades, or even centuries to overcome (recall that Leonardo Da Vinci had designs for flying machines that took centuries to be developed…).

The President of SpaceX may be taking a leaf out of the CEO’s always overly optimistic predictions. In November she said “I would not be surprised if we fly 400 Starship launches in the next four years”. Looking at the success of Falcon 9 it is certainly plausible that I may live to see 400 Starship launches in a four year period, but I am quite confident that it will not happen in the next four years (2025 through 2028).

One more thing. Back when I first made the predictions there had been an announcement by the CEO of SpaceX that in 2018 the company was under contract to send a very rich paying customer in a trip around the moon in 2018, launched on a Falcon Heavy. I was completely skeptical. Over the years the date got pushed back and pushed back, and the proposed flight vehicle was changed to be Starship. As we all know the flight of the Japanese billionaire around the Moon still hasn’t happened. In 2024 Yusaku Maezawa finally gave up waiting and cancelled the contract.

NASA Artemis

NASA’s plan is that the second Artemis mission, using the Orion Capsule, Artemis II, will fly to the Moon with four people aboard, the first crewed Artemis flight. An uncrewed flight of Orion around the Moon flew in 2022.  The crewed flight was scheduled to launch in May 2024, but it was first delayed by six months and then a little more and in the last year it has slipped another full year. It is now scheduled to fly in April 2026.

Artemis III was scheduled to launch in 2025 with a return to the surface of the Moon. However that relied on using a Starship (itself refueled in LEO by 14 (yes, fourteen!!) other Starship launches) to land there.  No one any longer believes that schedule, and willlikely delay a few years, given where Starship is in its development and current capability.  The officieal schedule says mid 2027, but that seems unlikely.

You can find the architecture of the Artemis III mission at this website.

Blue Origin Orbital BE-4 Engines and New Glenn

The suborbital tourist flights that Blue Origin operates are not its main business. It has ambitions to compete head to head with SpaceX. Another billionaire vs billionaire competition.

It has developed the BE-4 engine designed to fly 100 times, and to power the first stage of its massive New Glenn rocket (see below).  But in the meantime it has started selling the BE-4 to ULA (United Launch Alliance) to power their Vulcan Centaur heavy launch vehicle. It’s first stage uses two BE-4 engines, along with a variable number of solid fuel strap ons.

Vulcan Centaur flew two times in 2024 and the BE-4 engines worked perfectly both times, on January 8^th and again on October 4^th. This is a solid validation of the engine’s capabilities.

Blue Origin’s own first orbital class rocket, New Glenn, is massive, and comparable to the Flacon Heavy (three boosters) rather than the Falcon 9 in capability.   It has been in development for a long time, but saw its first visits to a launch pad, fully stacked in 2024. The first stage uses seven BE-4 engines, and is intended to land on a barge and be fully reusable. The second stage uses two BE-3U engines, a variant of the single engine used on their New Shepard sub-orbital space tourism vehicle. There is a project underway to make a fully reusable version of the second stage.

Launch seems imminent.  Here it is at the launch pad in November 2024.

On Friday December 27^th, 2024, it was fully fueled in both stages and went through a countdown and fired its seven BE-4 engines for 24 seconds. Now it will leave the pad to have its payload installed. The launch could be as early January 6^th.  The very first launch will be an all up affair, attempting to get something to orbit and land the booster on its first flight. This is a very different development approach to that used by SpaceX.

Let’s Continue a Noble Tradition!

The billionaire founders of both Virgin Galactic and Blue Origin had faith in the systems they had created. They both personally flew on the first operational flights of their sub-orbital launch systems. They went way beyond simply talking about how great their technology was, they believed in it, and flew in it.

Let’s hope this tradition continues. Let’s hope the billionaire founder/CEO of SpaceX will be onboard the first crewed flight of Starship to Mars, and that it happens sooner than I expect. We can all cheer for that.