AI Update, Late 2020 - dumpster fire

2020 is a very strange year and a dumpster fire in many respects. Everything is still holding together but it feels like the news we get are just progressively more absurd. Similarly is the case with AI where a slow motion train wreck is progressing eliminating more and more hyped up companies and researchers. There are still areas where money is pouring into the field, but it feels like it's a far cry of what it was during the peak hype a few years ago. I fully expect a shift from private to more public money, as the government are always late in the hype cycle. So there could still be more waste to come, though there is a feeling of decline in the air. Anyway let's jump into a few highlights of the recent months. 

DeepMind Alpha Fold

DeepMind has been rather quiet and even popular press noticed a significant decrease in the level of hype, but recently they managed to show some progress on protein folding problem. This problem is of high practical importance in biology so at least it's good to see that the company uses their incredible resources on something that we all may eventually experience as a benefit instead of clearly non practical problems such as playing the game of go. I'm not really an expert in protein folding problem so I can't really make a definite judgement on this result, all that I know is that the general protein folding in 3d space appears to be NP-hard so I highly doubt they actually "solved" the problem. I'm pretty certain that if Demis Hassabis actually managed to solve P=NP problem, DeepMind PR machine would go on such an overdrive that I'd probably find news about it in my freezer. That seems not to be the case and there is no shortage of blog post from people who seem know more about protein folding who diffuse much of the hype. Nevertheless DeepMind appears as a wonder-company making all these seemingly amazing breakthroughs in areas where lots of people were already trying to make a dent for years. These mystical properties dissipate once one looks into their financials. As reported by CNBC (based on official filings), the company burned $649 million in 2019. They also were forgiven a loan from Google valued at $1.1 billion. Such sums are equivalent to a full budget of a large university (though arguably Deep Mind does not actually need to train anybody aside from the few interns they get). For example the yearly total budget of MIT in 2020 is ~$3.7billion. Anyway, given these numbers, they are one of the worlds richest research labs, able to deploy enormous computational resources and relying for free on a highly reliable and scalable infrastructure of one of the world's biggest internet companies. Whether the results they present periodically are actually still impressive given all that burn rate, I'll leave it for the reader to decide.  

Element AI fiasco

In my AI update last year I mentioned a Canada based company Element AI, which at that time was apparently in the process of raising a flat round of financing. I expressed a warning, that flat rounds are typically not a sign of a healthy company and suggested that other than hiring some deep learning big shots for hype and some students to do some "busy work", they really had no idea for any business. Lo and behold one year later the company got liquidated [1](though euphemistically it is described as a "sale"). The company was apparently sold for $230 million to US based ServiceNow while they raised $257 million over four years of existence. Since they were "sold" for less than they raised, obviously common stock holders (including most of the employees) were wiped out, only some of the investors (typically holding preferred equities) were payed part of the money back. In essence this is bankruptcy and a liquidation of assets. From all I know they had some really fancy office space, but that apparently is not sufficient to create any value.  

Anthony Levandowsky in Jail

Once a star of the Google self driving car project, once a millionaire selling his Otto company to Uber with great fanfares (I wrote about this before here in 2016 and here in 2019), is now a bankrupt criminal serving his jail time. Apparently justice still exists in some corners of the world. 

Uber ATG - Aurora  SNAFU

Graveyard of Uber self driving cars.

While all the way until October 2020 Uber was still assuring they were in the autonomous car game, only two months later in December 2020 news broke that Uber is dumping their ATG (Advanced Technology Group) unit to Aurora. Now again this deal is euphemistically described as sales, but in reality it's more complex. Uber is actually investing $400 million into Aurora alongside of dumping ATG and agreeing among other investors to value the resulting company at $10B. All of this is just marking up price on paper, in reality Uber payed Aurora to take their ATG and play with it. Interestingly Aurora took most of the assets and employees from ATG but decided to pass on research lab Uber ATG had in Toronto lead by Raquel Urtasun.  Urtasun was hired by Uber to great fanfares only three years ago, but apparently Aurora doesn't need her skills. After Element AI, this is another big blow into Canadian AI party.   

TuSimple $350M 

Tu Simple - a self driving truck company claims to have raised $350M, bringing the total the company is about to burn to $650M. I would not be too surprised if these guys ended much like Element AI. I didn't really follow them closely, I know they've been rolling a truck on I-10 between Phoenix and LA (actually between hubs located next to the freeway, their trucks don't enter the cities). I-10 in that area is a freeway in a middle of nowhere - a desert. I don't think it's particularly challenging to run an autonomous truck along that particular route. After all, Otto was already delivering beer in Colorado in 2016. Nevertheless, there is something slightly fishy about this company, and I may investigate them in a bit more detail later.    

Boston Dynamics third owner

My favorite robotic youtube clip making company Boston Dynamics is changing hands again. According to sources SoftBank actually made quite a bit of money on the deal (they reportedly acquired BD from Google for $165 million). Though it should be noted that it is unclear how much SoftBank actually pumped into the company over those three years - from sources in the article above they've actually tippled their staff and bought a new headquarters, so the deal might not be as sweet as it looks like on the surface. BD apparently sold on the order of 400 of their dog robots for hefty $75 000. The robot is pretty much useless beyond an expensive amusement, the claimed use case for  "inspections" (as a remote controlled quadruped), can be already accomplished using drones for 1/10 of the price. In the meanwhile, tweets below https://twitter.com/jetpack/status/1 and [tweet 2] match my feelings about Boston Dynamics perfectly. 

 

 

 

 

 

 

Now don't get me wrong, I think they make some awesome robots. And if only we had brains we could put inside them, they could potentially be incredibly useful. But we don't have "robotic brains". And until we do, the creations of Boston Dynamics will be better described as "works of art" rather than anything practical. 

Tesla Fool Self Driving

My yearly AI update would be incomplete without Tesla. In April 2019 Elon Musk promised (and repeated that promise many times since then) that by the end of 2020 there would be a million Tesla robotaxis on the road (via a magical software update that will turn already sold cars to self driving). I [1] and a few others claimed there would actually be zero. And here we are, 2020 is over and there are zero Tesla robotaxis on the road. Meanwhile the company has been sending software updates which added new features to the autopilot but although impressive looking at the surface, the software is still hopeless in any sort of "unusual" conditions and often fails unpredictably even in seemingly benign looking conditions (which is why I think it is unsafe and should be taken off the road). Some folks tested the new autopilot on Lombard street in San Francisco where it failed spectacularly, while Waymo brought back some recording from the same place in 2009 while it was still "Google car" and the car managed to do much better. I will upkeep my bet that there won't be any Tesla robotaxis in 2021 and 2022, though in his typical manner Elon Musk might ship some slightly improved version of current tech, call it "Full Self Driving", cover his ass by some unknown regulations that won't allow him to unleash the full potential of the software and claim victory. Nonetheless, don't expect to be driven by your Tesla without being fully liable for any deadly accident it causes within the next few years and possibly ever.  

Bald or ball?

While people who have no idea what they are talking about are spreading panic about Boston Dynamics robots taking over the world, in reality AI systems can't tell a difference between a soccer ball and a bald head of the referee. There are plenty examples of such mistakes and they all clearly point to the obvious: contemporary AI system have very rudimentary understanding of reality and cannot be trusted to make similar judgements as we do.  

Other 

There has been some subtle claims going under the radar that AI alone may not be able to cut it for self driving cars and there might be a need for teleoperation. Companies have been popping out lately in support of that use case. I personally don't think teleoperation is really a viable solution for driving and certainly not a solution that will scale. This is mainly due to the hard real time requirement for this problem and inability of a teleoperator to acquire situational awareness within the time required to react to a dangerous situation. Meanwhile the MIT task force predicts autonomous vehicles won't arrive for at least 10 years, giving some extra support to my skepticism expressed in this blog. 

Summary

This year was crazy in many ways. There has been a lot of things going on so it's no wonder that AI lost some of the headlines. But the implosion of several flagship projects certainly gives a strong sense of deflation. I think the real AI winter will come once everyone will finally realize that self driving cars are not coming anytime soon. Once that fact of life finally hits the portfolios of those heavily invested in the current AI bubble, money stream will freeze for many years. And while the winter is needed to cool down some of the craziness, it will generally be harmful to legit researchers distancing themselves from the hype and probably slow down the progress towards real AI, but at this point, with all the bald promises already made, it seems inevitable. 

If you found an error, highlight it and press Shift + Enter or click here to inform us.

Comments

comments