On January 1^st, 2018, I made predictions about self driving cars, Artificial Intelligence and machine learning, and about progress in the space industry. Those predictions had dates attached to them for 32 years up through January 1^st, 2050.

I made my predictions because at the time I saw an immense amount of hype about these three topics, and the general press and public drawing conclusions about all sorts of things they feared (e.g., truck driving jobs about to disappear, all manual labor of humans about to disappear) or desired (e.g., safe roads about to come into existence, a safe haven for humans on Mars about to start developing) being imminent. My predictions, with dates attached to them, were meant to slow down those expectations, and inject some reality into what I saw as irrational exuberance.

I was accused of being a pessimist, but I viewed what I was saying as being a realist. Out of emotion I wanted to give some other predictions that went against the status quo expectations at the time. Fortunately I was able to be rational and so I only made one prediction that I view in retrospect as being hot headed (and I ‘fess up to it in the AI&ML table in this post). Any one can, and they do, make predictions. But usually they are not held to those predictions. So I am holding myself accountable to my own predictions.

As part of self certifying the seriousness of my predictions I promised to review them, as made on January 1^st, 2018, every following January 1^st for 32 years, the span of the predictions, to see how accurate they were. This is my third annual review and self appraisal, following those of 2019 and 2020. Only 29 to go!

I think in the three years since my predictions, there has been a general acceptance that certain things are not as imminent or as inevitable as the majority believed just then. So some of my predictions now look more like “of course”, rather than “really, that long in the future?” as they did in 2018.

This is a boring update. Despite lots of hoopla in the press about self driving cars, Artificial Intelligence and machine learning, and the space industry, this last year, 2020, was not actually a year of any surprises.

The biggest news of all is that SpaceX launched humans into orbit twice in 2020. It was later than they had predicted, but within my predicted timeframe. Neither they nor I are surprised or shocked at their accomplishment, but it is a big deal. It changes the future space game for the US, Canada, Japan, and Europe (including the UK…). Australia and New Zealand too. It gives countries in Asia, South America and Africa more options on how they plot their space ambitions. My heartfelt congratulations to Elon and Gwynne and the whole gang at SpaceX.

This year’s summary indicates that so far none of my predictions (except for the one hot headed one I mentioned above) have turned out to be too pessimistic. As I said last year, overall I am getting worried that I was perhaps too optimistic, and had bought into the hype too much.

Repeat of 2019’s Explanation of Annotations

As I said in 2018, I am not going to edit my original post, linked above, at all, even though I see there are a few typos still lurking in it. Instead I have copied the three tables of predictions below from 2020’s update post, and have simply added a total of eight comments to the fourth columns of the three tables. As with last year I have highlighted dates in column two where the time they refer to has arrived.

I tag each comment in the fourth column with a cyan colored date tag in the form yyyymmdd such as 20190603 for June 3^rd, 2019.

The entries that I put in the second column of each table, titled “Date” in each case, back on January 1^st of 2018, have the following forms:

NIML meaning “Not In My Lifetime, i.e., not until beyond December 31^st, 2049, the last day of the first half of the 21^st century.

NET some date, meaning “No Earlier Than” that date.

BY some date, meaning “By” that date.

Sometimes I gave both a NET and a BY for a single prediction, establishing a window in which I believe it will happen.

For now I am coloring those statements when it can be determined already whether I was correct or not.

I have started using LawnGreen (#7cfc00) for those predictions which were entirely accurate. For instance a BY 2018 can be colored green if the predicted thing did happen in 2018, as can a NET 2019 if it did not happen in 2018 or earlier. There are five predictions now colored green, the same ones as last year, with no new ones in January 2020.

I will color dates Tomato (#ff6347) if I was too pessimistic about them. No Tomato dates yet. But if something happens that I said NIML, for instance, then it would go Tomato, or if in 2020 something already had happened that I said NET 2021, then that too would have gone Tomato.

If I was too optimistic about something, e.g., if I had said BY 2018, and it hadn’t yet happened, then I would color it DeepSkyBlue (#00bfff). None of these yet either. And eventually if there are NETs that went green, but years later have still not come to pass I may start coloring them LightSkyBlue (#87cefa).

In summary then: Green splashes mean I got things exactly right. Red means provably wrong and that I was too pessimistic. And blueness will mean that I was overly optimistic.

So now, here are the updated tables.

Self Driving Cars

Oh gosh.  The definition of a self driving car keeps changing.  When I was a boy, three of four years ago, it meant that the car drove itself. Now it means that a company gives a press release that they have deployed self driving cars, reporters give breathless headlines about deployment of said self driving cars, and then bury deep within the story that actually there is a person in the loop, sometimes in the car itself, and sometimes remotely.  Oh, and the deployments are in the most benign public suburbs imaginable, and certainly nowhere that a complete and sudden emergency stop is an unacceptable action.  And no mention  of restrictions on time of day, or weather, or particular roads within the stated geographical area.

Dear technology reporters, after this table, where almost nothing has changed, save for two comments dated Jan 1^st 2021, I’ll give you some questions to ask before you anoint corporate press releases as breaking technology stories.  Just saying.

I’m still not counting the Waymo taxi service in Chandler, a suburb of Phoenix, as a driverless taxi service. It has finally been opened to members of the public, and has no human driver actually in the car, but it is not a driverless service yet. What about the October 8^th headline: Waymo finally launches an actual public, driverless taxi service with lede Fully driverless technology is real, and now you can try it in the Phoenix area. No. It is not fully driverless. When you read down in the story you eventually find that, just as for the last year with a more restricted customer base, there is still a human in the loop. The story proudly proclaims that “[m]embers of the public who live in the Chandler area can hail a fully driverless taxi today”, but a few sentences later say the economics may not work out “because Waymo says the cars still have remote overseers”. That would be a person in the loop. Although Waymo doesn’t give a straight answer of how many cars a single overseer oversees, they refuse to say directly that is more than one. I.e., they are not willing to say that there is less than one full person devoted to overseeing the control of each “driverless” taxi. This is not a driverless taxi service. Tech press headlines be damned. Oh, and by the way, Waymo’s aim was to get to 100 rides per week by the end of 2020. Not exactly a large scale deployment, and not open to the public at any sort of scale, and certainly not driverless.

But what about this December 9^th headline: Cruise is now testing fully driverless cars in San Francisco? Again it pays to read the story. Carefully. There is a Cruise employee in the passenger seat of every one of the “fully driverless” cars and they can stop the car whenever they choose. And just to be sure, every vehicle is also monitored by a remote employee. Two safety humans for every one of these so called fully driverless cars. A story from the same day in the Washington Post reports that the testing will only be in the Sunset district of the city, and only on certain streets. SF residents know that that is a very residential area with very little traffic, and four way stop signs at intersections every block. The traffic moves very slowly even when the streets are deserted, and it is always safe to come to a complete stop.

Reporters should ask all companies saying they are testing fully driverless cars, whether as a taxi service or not, the following questions, at least.

• Is there an employee in the “driverless car”? Are they able to stop the car or take over in any way?
• Is there a remote person able to monitor the car? Are they able to take over in any way, even just commanding that the car stop, or being able to set a way point which will let the car get out of a stuck situation? How many cars does such a remote person monitor today (as distinct from aspirations for how many in the future)?
• Is there a chase vehicle following the driverless car to prevent rear end collisions from other random drivers? (I understand that Cruise does this in San Francisco; and I think I have seen it myself with a Waymo vehicle in the last week, but I may have misinterpreted.) If there is a chase vehicle does someone in that vehicle have the ability to command the driverless vehicle to stop? Note that a chase vehicle is a whole second car driven by a person just to enable the driverless car to be safe…
• Are there any restrictions on the time of day that the system operates?
• Will testing change if there is really bad weather or will it go on regardless?
• What is the area that the test extends over, and are there any streets that the cars will not go on?
• If it is a taxi service are there designated pickup and drop off sites, or can those be anywhere? And that includes 3 meters further (regular taxis and rideshare services do this) at the passenger’s request.
• Once a passenger gets in can they change the desired destination? Can they abort the ride and get out whenever they want to?
• Can the passenger communicate with the car (not a remote person) by voice?

Most of these questions will have answers that the PR people, at this point, will not want to be forthcoming about. Sometimes they will be “unable to answer” even when it is a simple yes/no question (e.g., see the Cruise story linked in the questions above)–that is likely not a sign of them not actually knowing the answer…

The only company “that says it has” tested fully driverless cars, with no chase vehicles on California roads is Nuro, but that is a long way from a taxi service since they use special purpose tiny vehicles that can not accommodate even one person.

Towards the end of the year AutoX, a company based in China, has said they are operating a driverless taxi service in a number of cities. First note, however, that the service is not open to the public, and is only available for demonstrations by company employees. They say they have removed both drivers in the cars and remote drivers, but none of the articles I have seen have answered the comprehensive set of questions I have above, including whether there is a person doing remote monitoring rather than remote driving. The one video I have seen is from inside an empty vehicle as it drives on empty(!!) roads. I have spent a lot of time in China in a lot of different cities, and I have never seen an empty road.

I am not saying there is anything wrong with these small steps towards driverless cars. They are sensible, and necessary steps along the way. But if you look at the headlines, and even the first few paragraphs of most stories about driverless cars, they breathlessly proclaim that the future is already here. No it is not. Read the fine print.  We are very far from an actual driverless taxi service that can compete with a traditional taxi company (which has to make money), or even one of the ride share companies which themselves are still losing money. The deployed demonstration systems have many more people per car working full time than any of these other systems, are only operating a tiny number of rides per day, and would not be profitable even if each ride was charged at a thousand times that charged by the incumbents. We are still a long, long way from a viable profitable autonomously driven taxi service.

The last two years have seen a real shakeout in expectations for deployment of self driving cars.  The most recent example of this is was in December 2020 when Uber sold their self driving car efforts for minus $ 400 million. That’s right.  They gave away the division working on self driving cars, once thought to be the way to finally make Uber profitable, and sent along with it $ 400 million in cash.  [[… If only they had asked me. I might have bought it for minus $ 350 million! A relative bargain for Uber!]]

To illustrate how predictions have been slipping, here is a slide that I made for talks, based on a snapshot of predictions about driverless cars from March 27, 2017. The web address still seems to give the same predictions with a couple more at the end that I couldn’t fit on my slide. In parentheses are the years the predictions were made, and in blue are the dates for when the innovation was predicted to happen. The orange arrows point at predictions that I have since seen companies retract. I may have missed some retractions. The dates highlighted in red have passed, without coming to fruition. None of the other predictions have so far borne fruit.

In summary, twelve of the twenty one predictions have passed and vanished into the ether. Of the remaining nine I am confident that the three for 2021 will also not be attained. And then there are none that come due until 2024. I think it is fair to say that predictions for autonomous vehicles in 2017 were wildly overoptimistic. My blog posts of January 2017 and July 2018 tried to tamp them down by talking about deployment issues that were not being addressed, and that, in my opinion, needed large amounts of research to solve. It turns out that before you get to my concerns just the long tail of out of the ordinary situations, and not even the really long tail, just some pretty common aspects of normal driving in most cities and on freeways, have delayed the revolution.

Two more car topics

A year ago, in my update at the start of 2020, I had a little rant about a prediction that Elon Musk had made about how there would be a million Tesla robo-taxis on the road by the end of 2020. His prediction certainly helped Tesla’s stock price. Here is my rant.

Well, surprise!, everyone.  2020 has come and gone and there are not a million Tesla robo-taxis on the road. There are zero. I, and Kai-Fu Lee, both had complete confidence in our predictions of zero. And we were right. Don’t worry. Undaunted by complete and abject failure to predict accurately, twice again, Elon has new predictions. In July he predicted level 5 autonomy in 2020, and in December he made predictions of full autonomy in 2021, with his normal straight face. Those predictions have again have pumped up Tesla’s stock price. Never give up on a winning strategy!

Finally I want to say something about flying cars.  As with “self driving cars” the definition has been changing as PR departments for various companies try to claim progress in this domain. My predictions were about cars that could drive on roads, and could take to the air and fly when their driver wanted them to. Over the last couple of years “flying car” has been perverted to mean any small one or two person flying machine, usually an oversize quadcopter or octodecacopter, or some such, that takes off and lands and then stays exactly where it is until it takes off again. These are not cars. They are helicopters. With lots of rotors. They are not flying cars.

Artificial Intelligence and Machine Learning

I had not predicted any big milestones for AI and machine learning for the current period, and indeed there were none achieved. In the world of AI hype that is a bold statement to make. But I really do believe it. There have been overhyped stupidities, and I classify GPT-3 as one, see the analysis at the end of this section.

No research results from this last year are easily understood by someone not in the thick of the actual research. Compare this to the original introduction of the perceptron, or back propagation, or even deep learning. Geoff Hinton was able to convey the convergence of three of four ideas that came to be known as deep learning, in a single one hour lecture, that a non-specialist could understand and appreciate. There have been no such innovations in the last year. They are all down in the weeds.

I am not criticizing research being down in the weeds. That is how science and engineering gets done. Lots and lots of grunt work making incremental improvements. But thirty years from now we probably won’t be saying, in a positive way, “remember AI/ML advances in 2020!”  This last year will blend and become indistinguishable from many adjacent years.  Compare that to 2012, the first time Deep Learning was applied to ImageNet. The results shocked everyone and completely changed two fields, machine learning and computer vision. Shock and awe that has stood the test of time.

In last year’s review I had some thoughts and links to things such as the growing carbon foot print of deep learning, on comparison to how quickly people learn, on non generalizability of games that are learned with reinforcement learning. I also argued why game playing has been such a success for DeepMind, and why it will probably not generalize as well as people expect. These comments occurred both just before and just after last year’s version of the table below. I stand by all those comments and invite you to go look at them.

In the table below I have added only three items, two of which relate to how people look at AI and ML, and one to a missing capability, rather than technical advances.

There have been many criticisms in the popular press over the last year about how AI is not living up to its hype. Perhaps the most influential such one is from the The Economist. The headline was: An understanding of AI’s limitations is starting to sink in, with a lede of After years of hype many people feel that AI has failed to deliver. Such rationality has not stopped breathless other stories in outlets that should know better, such as the AAAS journal Science, and sometimes even in Nature. The ongoing amount of hype is depressing. And it is mostly inaccurate.

Attention

In the table above I mentioned that both common sense and attention have surfaced as something that current ML methods are lacking and at which they do not do well.

Although learning everything from scratch has become a fetish in ML (see Rich Sutton and my response, both from 2019) there are advantages to using extra built-in machinery. And of course, when ever any ML-er shows off their system learning something new they have been the external built-in machinery. They, the human, chose the task, the reward, and the data set or simulator, with what was included and what was not. Then they say “see, it did it all by itself!”. Ahem.

We have seen reinforcement learning (RL) being used to learn manipulation tasks, but they are both contrived, by the humans writing the papers, and can be extraordinarily expensive in terms of cloud computation time. How do children learn to manipulate?  They do it in evolutionarily pre-guided ways. They learn simple grasps, then they learn the pinch between forefinger and thumb, and along the way they learn to direct force to those grasps from their arm muscles. And in some seemingly magical way, they are able to use these earlier skills as building blocks.

By the time they are 18 months (and I just observed this in my own grandson), presented with a door locked with an unknown mechanism (four different instances in this particular anecdote), children use affordances to know where to grasp and how to grasp. They ignore applying arbitrary forces to the wood of the door, except to try to pull it open, testing whether they have gotten the lock in the unlocked state, and instead apply seemingly random forces and motions (just like a machine RL learner!) to just the lock mechanism, experimenting with different grasps and forces. They direct their manipulation attention to the visible lock mechanism, rather than treating the whole thing as one big point cloud as a machine RL learner might. They know where to pay attention to.

Of course this strategy might fail, and that is precisely the allure of secret lock wooden boxes. But usually it works. And the search space for the learner is a couple of orders of magnitude smaller than the “learn everything” approach. When the bill for cloud computer time is tens of millions of dollars, as it is for some “advances” promoted by large companies, an order of magnitude or two can make a real cash difference.

We need to figure out the right mechanisms for attention and common sense and build those into our learning systems if we are to build general purpose systems.

GPT-3 (or the more things change the more they are the same…)

Will Bridewell has likened GPT-3 to a ouija board, and I think that is very appropriate. People see in it what they wish, but there is really nothing there.

<rant>

GPT-3 (Generative Pre-Trained Transformer 3) is a BFNN (that’s a technical term) that has been fed about 100 billion words of text from many sources. It clusters them, and then when you give it a few words it rambles on completing what you have said. Reporters have seen brilliance in these rambles (one should look at the reports about Eliza from the 1960’s), and some have gone as far as taking various output sentences, putting them together in an order chosen by the journalist, and claiming that the resultant text is what GPT-3 said.  Disappointingly, The Guardian did precisely this getting GPT-3 to generate 8 blocks of text of about 500 words each, then the journalists edited them together to about 1,100 words and said that GPT-3 had written the essay. Which was true in only a very sort-of way. And some of these journalists (led on by the PR people at Open AI, a misnamed company that produced GPT-3) go all tingly about how it is so intelligent that it is momentous and dangerous all at once; oh, please, the New York Times, not you too!! This is BS (another technical term) plain and simple. Yes, I know, the Earth is flat, vaccines will kill you, and the world is run by pedophiles. And if I wasn’t such a bitter pessimist I would hop on the GPT-3 train of wonder, along with all those glorious beliefs promoted by a PR department.

Not surprisingly, given the wide variety of sources of its 100 billion training words, GPT-3’s language is sometimes sexist, racist, and otherwise biased. But people who believe we are on the verge of fully general AI see the words it generates, which are rehashes of what it has seen, and believe that it is intelligence. A ouija board. Some of its text fragments have great poetry to them, but they are often unrelated to reality, and if anyone is even a little skeptical of it, they can input a series of words where its output clearly shows it has zero of the common human level understanding of those words (see Gary Marcus and Ernie Davis in Technology Review for some examples). It just has words and their relationships with no model of the world. The willfully gullible (yes, my language is harsh here, but the true believers or journalists who rave about GPT-3 are acting like cult followers–it is a mirage, people!) see true intelligence, but its random rants that it produces (worse than my random rants!) will not let it be part of any serious product.

And finally, GPT-n was actually invented, using neural networks, back in 1956. It just needed more and bigger computers to fool more of the people more of the time. In that year, Claude Shannon and John McCarthy published a set of papers (by thirteen mostly very famous authors including themselves, John von Neumann, and Marvin Minsky) titled Automata Studies, under the imprint of Princeton University Press (Annals of Mathematical Studies, Number 34).

I love that the first paragraph of their preface begins and ends:

Among the most challenging scientific questions of our time are the corresponding analytic and synthetic problems. How does the brain function? Can we design a machine which will simulate a brain? … Currently it is fashionable to compare the brain with large scale electronic computing machines.

Apart from the particular sparse style of their language, an equivalent passage could easily be thought to have been written by any number of modern researchers, sixty four years later.

On page vi of their preface they criticize a GPT-n like model of intelligence, one that is based on neural networks (the words used in their volume) and able to satisfy the Turing Test (not yet quite called that) without really being a thinking machine:

A disadvantage of the Turing definition of thinking is that it is possible, in principle, to design a machine with a complete set of arbitrarily chosen responses to all possible input stimuli (see, in this volume, the Culbertson and the Kleene papers). Such a machine, in a sense, for any given input situation (including past history) merely looks up in a “dictionary” the appropriate response. With a suitable dictionary such a machine would surely satisfy Turing’s definition but does not reflect our usual intuitive concept of thinking.

The machine they refer to here, a neural network, is just like the trained GPT-3 network, though Culbertson and Kleene, in 1956, had not figured out how to train it.

Steven Kleene was a very famous logician and academic Ph.D. sibling of Alan Turing, under Alonzo Church. James Culbertson is a forgotten engineer who speculated on how to build intelligent machines using McCulloch-Pitts neurons, here and elsewhere–I rehabilitate him a little in my upcoming book.

S. C. Kleene, Representation of Events in Nerve Nets and Finite Automata, pp 3-41.

James T. Culbertson, Some Uneconomical Robots, pp 99-116.

</rant>

AlphaFold

Finally, I want to mention AlphaFold, which right at the end of the year got incredible press, breathless incredible press I would say, about “solving” an outstanding question in science.

Given a sequence of amino acids making up a protein, how will that protein fold due to physical forces between parts of nearby acids? This is a really important question. The three dimensional structure, what is inside, and what is on the outside, and what adjacencies are there on the surface of the folded molecule, determines how that protein will interact with other molecules. Thus it gives a clue as to how drugs will interact with a protein.  People in Artificial Intelligence have worked on this since at least the early 1990’s (e.g., see the work of Tomás Lozano-Pérez and Tom Dietterich).

AlphaFold (from DeepMind, the subsidiary of Google that is losing the better part of a billion dollars per year) is a real push forward. But it does not solve the problem of predicting protein folds.  It is very good at some cases, and very poor at other cases, and you don’t know which it is unless you know the real answer already. You can see a careful (and evolving) analysis here. It is long so you can skip to the four paragraphs of conclusion to hear about both the goods and the bads. It likely will not revolutionize drug discovery without a lot more work on it, and perhaps never.

While AlphaFold is another interesting “success” for machine learning, it does not advance the fields of either AI or ML at all. And its long term impact is not yet clear.

Space

In 2020 SpaceX came back from the explosion of a Crewed Dragon capsule that happened during a ground test on April 20^th, 2019.  They were able to launch their first two people to space aboard Demo-2 on May 30^th, 2020.  Bob Behnken and Doug Hurley, both NASA astronauts, docked with the ISS on May 31st, and stayed there until their return and splashdown on August 2^nd.  On November 16^th, Crew-1 launched with three NASA and one JAXA astronauts on board, and is currently docked to ISS where the astronauts are working.

Boeing’s entry into manned space flight, the CST-100, is behind SpaceX. A software problem in December 2019 meant that an unmanned flight did not make it to the ISS, though it did return safely to Earth. In April of 2020 Boeing announced it would refly the unmanned test in the fourth quarter of 2020. That has now slipped to the first quarter of 2021 due to further software issues. I think Boeing is one glitch away from not to getting a manned flight in 2021. So my optimism for them is going to be tested.

There was only one attempt at a manned suborbital flight in 2020 but it did not get to space. On December 12^th the rocket motor of Virgin Galactic’s SpaceShipTwo Unity did not ignite after the craft was dropped as planned from its mothership at over 40,000 feet in altitude.

On October 13^th, 2020, Blue Origin’s New Shepard 3 had its 7^th successful unmanned test flight. A new craft, New Shepard 4, was slated to undergo an unmanned test flight in the last two months of the year, as the 14^th test flight of a New Shepard system. That did not happen. New Shepard 4 is currently expected to fly the first manned mission for Blue Origin.

As noted in the table SpaceX originally predicted that they would send two paying customers around the Moon in 2018, using Falcon 9 Heavy as their launch vehicle. During 2018 they revised that to a new vehicle, now known as Starship, with a date of 2023. Progress is being made on Starship, and the 55m long second stage flew within the atmosphere for more than a simple hop in 2020. It did controlled aerodynamic flight as it came into land, and impressively and successfully flipped to vertical for a landing. There were some fuel control issues and it did not decelerate enough for a soft landing, and blew up. This is real and impressive progress.

So far we haven’t seen the 64m first stage, though there are rumors of a first flight in 2021. This will be a massive rocket, and given the move fast and blow up strategy that SpaceX has very successfully used for the second stage, we should expect its development to have some fiery moments.

A 55m long vehicle to loop around the Moon seems like a bit of overkill, but that could be the point. However, I will not be surprised at all if the mission slips beyond 2023.  If fantastic progress happens in 2021 I will get more confident about 2023, but 2021 will have to be really spectacular.