For any doubters of the Boston Dynamics dancing robots video that dropped over the holidays, let’s start this newsletter by saying that everything they showed us is well within the known capabilities of their motion control tech.
Yes, it’s another Octoboxy newsletter about the evolving world of artificial intelligence. Drawing from our background in computer engineering, media, business, and trend-watching, we’re here to point out the highlights of 2020’s AI news.
Past newsletters are here:
The letter is a little longer this time – reflecting an industry that’s heating up. If you want to jump ahead to your favorite parts, here’s the links:
For now, let’s hop back to Boston Dynamics.
Their dancing bots were gorgeous, but the real news is that Boston Dynamics has been sold and bought, again. Previously they were owned by SoftBank, but Hyundai now controls 80%.
B.D. has been doing a lot of brand image spin lately, probably forced into it by the Netflix show Black Mirror, which we all loved exactly because that one episode had robotic villains eerily similar to B.D.’s dog Spot. For what it’s worth, Spot is commercially for sale, and at only $75,000 you too can join the wait list.
Demand has outstripped production though, and apparently somewhere south of 1000 units have shipped so far. This is why the Hyundai purchase is so important, because it gives B.D. the big resource they critically need: manufacturing capacity.
One thing totally missing from Boston Dynamic’s public image is any hint of military bots. As Black Mirror reminded us, this company started out as a project to build combat mules. That early effort was abandoned because the bots were too noisy, which is probably why the company sometimes is known to describe Spot as “the quietest robot [they] have built.”
In-house, we’re fans of Spot, but we’re even more interested in the company’s new bot Handle. We foresee incredible opportunities to streamline many warehouse and fleet operations with Handle, if they can just build them fast enough.
Let’s touch quantum computing while we’re still in miscellaneous news. There’s been a lot of iterative advancement in quantum hardware, but nothing that’s directly applicable to AI, yet.
As this nascent field gels, some of the buzz words we’re seeing used are “quantum advantage” and “quantum supremacy,” which are not quite the same thing. The first term means that the quantum computer will greatly accelerate the pace of whatever computation we’re doing on binary hardware today. The second term means the quantum computer will be able to solve problems that we can not solve at all right now.
We’re not going to expand the contents of what actual advancements have been happening here – it’s all stuff like material sciences, or even which elemental particles are being used for qubits. Instead we’re going to make a forecast and say:
We’re at most 2 years from quantum advantage in machine learning.
If you need some details, here’s a few links:
A buzz word we want to kill early is “Edge AI”. What does Edge AI mean for AI research? Answer: not a thing.
Let’s pull a quote from that article we linked, so that we can throw shade at it:
“First, we saw the shift from mainframe to computers to cloud, now, the cloud is moving to Edge, and so is AI.”
No, summer child, in our life we’ve seen several cycles of computation shifting from mainframes, to smart terminals, to LAN servers, to desktops, to WAN servers, into the cloud, to client-side mobile devices, as well as the internet of things, and…
The point here is that just as computation has gone everywhere, so will AI. The actual term “Edge AI” is simply this year’s sales jargon for “LAN server”, because a number of companies have managed to accumulate more Big Data than they have Cloud Bandwidth. At the end of the day, “Edge AI” is a phrase meant to sell rack-mount hardware.
But Edge AI does hint at the true big trend of the year: MLaaS
The real trend in AI is Machine Learning as a Service, or MLaaS. All the big tech companies have new offerings in this field. Let’s glance at some highlights:
Amazon hosts some ridiculous percent of the entire internet, and AI projects are no exception. In one fun keynote at AWS re:Invent they dropped a statistic that 92% of all cloud-based TensorFlow projects run on AWS. Their PyTorch number was similar.
The scale of these projects can be mammoth as well. The presenter cited a car-company customer who was processing 7 petabytes (7 million gigabytes) of data inside the AWS SageMaker service. (Now there is a likely candidate for Edge AI.)
Besides generic platforms, the company has a whole range of specialized AI solutions you can subscribe to. To showcase just one, our favorite is their business intelligence service QuickSight Q.
On the surface, QuickSight Q is an AI data-scientist who constructs charts and graphs on-the-spot to answer your natural-language questions. But as we imagine companies actually using this service, we realize Amazon not only gets to collect the intimate financial metrics of each of their customers, but they also get to record the decision-making journey that those companies’ executives take as they guide their firms.
This year Amazon’s product is QuickSight Q, which is just an AI data-scientist. But we’ll be shocked if this doesn’t grow into a full business-analyst bot over time.
Google is the underdog in web hosting, and they know it. Their biggest MLaaS announcement is that BigQuery Omni now lets you analyze data hosted on Amazon.
But if you want to see a glimpse into what Google thinks about the future of business, trade them an email and read one of their ebooks, like maybe, ‘Unlock the transformative power of AI/ML in Retail.’ Here’s an excerpt:
“Over 60% of retailers we surveyed have more than 20 AI/ML use cases in pilot. On average, retailers are testing out 24 use cases at a time, while fully implementing an average of 11.”
Google’s analysts suggest that most companies are not wisely investing their AI efforts. AI is good at optimization-type problems, is their argument, such as inventory management and supply chain. But at the same time, they also suggest many other problems to optimize, like real estate contracts.
Maybe in a year or two our newsletter will get to say that Google’s AIs have started negotiating business deals. That would be one heck of a service offering.
We pulled these examples – Amazon QuickSight Q and Google predicting AI-optimized real estate deals – because we believe they’re related. It’s our take that both companies are racing to build an AI with an MBA. If you made us throw a dart at the calendar and make another forecast:
Within 3 years Amazon and/or Google will release an AI with an MBA.
We have some other trends to look at, so let’s skim over the rest of the miscellaneous news as fast as we can. All these folks are doing something, it’s just maybe a more obvious thing.
Full self-driving has dropped for elite beta testers. Given Tesla’s aggressive engineering schedule, that means other Tesla drivers will almost certainly see it later this year. They’re shy on details though, so we don’t actually know how “full” the “full self-driving” really is.
The OpenAI project has built the next version of their text-synthesis bot. This one they’re not releasing to the public directly, but instead they’re planning on spinning up a hosting service and granting developers access via an API.
This is probably a good move. OpenAI is very concerned with their tech being used nefariously, and this is going to grant them the most control over who’s doing what with it.
The research side of Google has unlocked the next level of medicine, basically. This is probably the best actual AI news of the year. Sadly, it’s relegated to this small spot in our report, because for a while now we’ve been hearing that protein-folding simulations were the next great computational barrier due to be overcome. Yay, we solved them!
It happened pretty much when we expected it would, and was done by the R+D arm of the world’s largest AI company. It’s lovely, but it feels more like Google is trying to buy themselves some social forgiveness than they actually want to help the world.
Western countries are accidentally going to put AIs in charge of our economies when Amazon and Google succeed in building bots with MBAs. China, by contrast, seems to be aiming that way deliberately.
In a late 2019 paper called On the Measure of Intelligence, a well-known AI researcher named F. Chollet poses that we’ve hit a plateau in the advancement of AI. The author argues that we’re iteratively making our dumb, narrow AIs deeper and bigger, but there’s still a fundamental level of synthesis that our biological brains can do and silicone ones can not. Chollet proposes a challenge in this paper, a set of questions that are all solvable by humans, but are extremely difficult to answer with current machine learning science.
It’s a fascinating paper, and we’ve been mulling over the implications all year.
Other authors have explored the idea in journalistic forms. This article, for instance, is trying to make the same point, but then starts talking about computational resources. They also give us the fun statistic that the computational power we’re throwing into deep learning is doubling every 3.4 months, a pace that’s 7x Moore’s Law!
While we believe Chollet’s paper is seminal, the Davis article’s twist toward computational power gives us a hint at forming a rebuttal. The idea we pose is that engineering problems rarely have breakthroughs where new levels of understanding are suddenly available. Instead, diverse teams invent small iterative improvements across a wide range of related technologies, until one day we realize it’s a whole different world.
To use a totally unrelated field as an example, check out this chart from the National Renewable Energy Laboratory on advancements in solar energy efficiency over time.
Perhaps current deep learning techniques are insufficient for certain classes of problems, but research into the kinds of problems we can solve is still pushing forward at a staggering rate. Even more, we’re not actually sure what the boundaries are on what problems are solvable.
For example, in our newsletter last year we made the challenge that, ”…any company not researching large-scale simulations is currently missing the whole point of AI.” Our context was half-a-billion games of hide-and-seek. But just a few months later, a team at SalesForce released results from a research project where they had done exactly what we wanted to see: they had simulated an economy and solved for optimal tax rates.*
* – punchline: trickle-down policies do not work, sorry
Another direction where AI science makes progress, even though deep learning as a technique stays the same, is the idea of You-Only-Look-Once, or YOLO.
YOLO was introduced in 2016 as a technique for computer vision. Imagine you’re building an object classifier AI, and you teach it to recognize dogs, cars, and bicycles. You give it an image to analyze, and you might imagine the computer evaluates three different networks, one for dogs, one for cars, and one for bicycles, then gives you back the strongest match of the three. Indeed, that’s exactly what used to happen.
But YOLO changed this and was a technique in which the same deep learning network could recognize dogs, cars, and bicycles all at the same time. But the magic went even deeper yet. In YOLO, an input image is broken up into a bunch of little tiny sub-images. Each sub-image is compared against the deep-learning network, and nothing ever considers the whole image at once.
Say the photo you’re analyzing has a dog. YOLO looks at your image as a grid. Some of the sub-rectangles have parts of a dog: just a paw, a muzzle, a shoulder, or an ear, for instance. But from just considering each part separately, YOLO confidently predicts, in one computational step, the entire span of the dog in the image.
YOLO revolutionized computer vision when this technique was invented. Since, the team has made several iterative improvements and currently are on version 4 of their algorithm. Derivative work is also starting to show up. In late 2019 a similar approach was applied to whole-body pose estimation from live video of real people.
We really like the idea of this approach, and we might use the word “intuitive” to try to name why. The YOLO algorithm is being trained to make lots of wild intuitions way outside of available data, rather than be limited to only what’s provably correct. This strikes us as an approach that might have far-reaching applications we haven’t seen yet.
This is our challenge to the industry this year: let’s look for other cases where intuitive good answers are better than computationally right ones.
The final trend we want to cover today is the big one, the idea of “responsible computing.” That is, how are we doing on the challenge of building AI for the betterment of humanity?
Everything is going, in a word, abysmally.
Let’s start at the top:
Two years ago when penning our first AI essay, our takeaway point was this:
“When any single AI picks the news content that hundreds of millions of people believe as true, it doesn’t much matter what intentions the AI is built with, it’s going to affect society.”
The best news on this front is that the Netflix documentary The Social Dilemma came out this fall. The movie is brilliantly made. If you haven’t seen it yet, it’s definitely worth a watch.
Followers of our newsletter will find no revelations in the film, because it’s basically made by many of the same sources whom we cite for these essays. But it’s good that the film exists, because now we can point our friends and family towards it and say, “watch this movie, it does a good job of explaining what’s so terrifying about social media.”
Facebook still pretends very hard that they’re not building AI. Yes, they publish a few papers, and they’re superficially part of the MLaaS movement, trying to lure innovative young developers to their platform. Yet they heavily downplay how their entire business model relies on some of the richest AI in the world.
Then again, downplaying their own business model is the very hallmark of Facebook’s global strategy. For what it’s worth, searching for the phrase “Responsible AI” at Facebook yields a page with 2 followers.
Microsoft’s public image for AI is much better than Facebook’s, and they clearly spent a lot of advertising budget in crafting it. For example, they made a video that explains their approach to Responsible AI.
Bear with us for a moment, as we’re going to rip this video apart. Let’s watch it together. Ready?
It opens, as it should, with a visual of renewable energy. Then it follows with a non-white lady, and they even gave her ink, a nose-ring, and a coffee cup for character. As we keep watching, the video is a veritable parade of ethnicities and disabilities. Everyone is seen at Microsoft! Each pleasant-looking person sports perfectly manicured nails and expert makeup. Truly, the first face we see with a pimple is an extreme close-up of someone with deep, asian-seeming eyes, behind thick black-rimmed glasses, that reflect a computer monitor.
So yes, geeky engineer who was allowed to keep their pimple, Microsoft sees you, at least enough to let you be the public face of any woke stereotype that’s trendy right now.
The point they’re trying to make is very real: AI bias is a true problem. It’s also the trendy problem, the one they can talk about and sell you consulting time as they do it. But I defy you to find a concrete example where they’ve made any kind of actual impact. They say all the words and pay for expensive spin coverage, but just like the diversity treatment in their marketing video, they manage to stop short of investing any true effort in changing things.
Their video is not all bad though. We deeply love the text flyovers near the beginning:
The potential of AI For the benefit of everyone Machines Move into Rolls of Humans
On the theme of robots taking human rolls, let’s jump back to Amazon for our centerpiece. So far, as the company buys empty shopping malls and converts them into fulfillment centers, they’ve seemed content to build a hybrid human/robot workforce.
However, we had opportunity to tour a new Amazon sort center a couple years back, and it’s that experience we imagine when watching videos about agile dancing robots like Handle. We can not convince ourselves that Amazon isn’t on the brink of automating away half their workforce. The tech is ready, all it really would take at this point is enough high-end production capacity to build a fleet.
Google is still playing whack-a-mole with fake news, handling it case-by-case through means of content moderation, while otherwise ignoring that their core algorithm is designed to promote this stuff.
Update: 2021-03 – Google is systematically muzzling and/or dismantling their entire “ethical AI” team.
As far as responsible AI, if the big companies aren’t announcing one thing while doing another, it’s only because they’re doing nothing at all. In some obvious cases, their entire business model depends on AIs subverting human interests.
The big takeaway here is that, kind of like “Edge AI”, the term “Responsible AI” seems to be jargon used to sell consulting services, and it doesn’t actually have any impact on where we’re going as a people.
After throwing shade at every main enterprise out there, you might wonder if there is any algorithm we here at Octoboxy actually trust?
Well, yes!* Spotify.**
* – but “trust” is a strong word, maybe “don’t totally distrust”
** – with a paid account, we have no idea what their ad-supported accounts are like
The company is not doing anything interesting on a business front. (They keep throwing money into podcasts, which they may earn back someday, but we’re skeptical). But their AI seems to have a half-decent sense of good music, and we think that’s a very hard problem. For one thing, it’s an intuitive problem, where there are good answers but no right ones. As we mentioned above under Philosophy, we want to see more cases where machine learning is applied to problems like this.
Thanks for coming with us on this journey. All the best to you and yours as we cross this next year!
~Wyv, Chief of Octoboxy
We could not log you in, reset your password, sign you up, please try again.