2020-01-07

A Survey of the State of AI

2020-01

Most of a year ago we learned that AI has already reshaped our society, present tense, not the future tense we hear so much speculation about. No singular news item since has been as revolutionary, but yet industry is never idle. A steady stream of smaller advancements demonstrate that AI is a growing force and more society-bending changes are sure to come. Now, as 2020 kicks off, it seems a good time to take a new look at the industry. In this paper we begin with a high level analysis of current trends in AI, both where tech is winning as well as where it’s off course, then we dive deep into the specific details of each major firm’s recent announcements.

The Biggest Trend

Science fiction often depicts the AI takeover as an army of robots – walking, talking, android employees, maybe something like like Star Trek’s Lieutenant Commander Data. Truly, the modern race seems to be instead which company can first build the Starship Enterprise: a device that rarely talks back unless we ask it something, we live inside it, and it takes care of all our needs, automatically.

Every tech behemoth right now is working towards better and more ubiquitous voice interfaces. Better means a deeper understanding of the nuanced questions that we ask our devices. More ubiquitous means we invite cloud-connected smart speakers to sit in every room in our homes, offices, autonomous cars, or even on our bodies in the shape of wearable glasses.

This trend brings to the forefront a constant discussion about personal privacy. This is an important issue, true, but by-and-large privacy is beyond the scope of our analysis today.

Another related sub-trend, also beyond our scope, is there seems to be growing tension in the industry between Big Tech™ companies and their own employees. Internal activism within many large corporations is growing as senior leadership make decisions about privacy, human rights, environment, globalism, and other social outcomes. In response, it seems some large tech firms have begun forcing out the voices of dissent from within their ranks.

But again, while privacy and activism are trends worth watching, they are actually beyond the questions we’re focusing on here. Here we are interested in what the tech is doing, not whether it is morally okay to do it.

Winning the Race

Returning to the main trend of voice interface, there has been emerging evidence that only Google and Microsoft have the resources needed to win the race to a smart assistant. The reason for this is that smart assistants can only grow as intelligent as their underlying knowledge base.

No matter how sophisticated a voice recognition system becomes at understanding the questions we ask, answering those questions is a different problem entirely. In this way, Google and Microsoft have the advantage because only those two firms have large indexes of the whole internet from which they can build a good knowledge base.

Missing the Mark

Pulling back to an even broader scale, we would pose that most of the rest of the tech industry is still totally missing the mark on what AI is truly capable of.

There exists an ocean of advertising for companies using machine learning to reduce a pile of your best data to a single confidence number. To possibly over-generalize, it feels like many of the companies researching AI for their products are treating deep learning like nothing more than a new way to calculate probability.

For example, there are multiple CRMs that try to rank your sales leads by a statistical likelihood of conversion. All your digital conversations are inputs to the algorithm, and it pops out a single measure of whether a particular lead is worth your time to pursue.

But machine learning has been around for decades, and applications like this are absolutely nothing new. What’s actually different about the tech these days is not the idea of machine learning itself, but that exponential growth in hardware has let us brute-force our way through ever larger piles of data and ever more massive numbers of subtly varied experiments on how to interpret it.

No longer is a linear regression, principle component analysis, or any other favorite number crunching algorithm the goal of data research. Instead, savvy researchers are organizing armies of cloud computers to run vast simulations across arbitrarily complex systems until emergent behaviors begin to appear. As processing power continues to scale, the breadth and depth of these large scale simulations will also continue to grow. This is the true power of modern machine learning – not crunching your data to a confidence score, but simulating whole realities, over and over and over again until we get it right.

For example, this is how autonomous vehicles learn to drive: we don’t care about a confidence number that another car is going to make an unexpected lane change, instead we simulate the whole world with as much variation as we can manage and let the AIs develop their own strategies for navigating through it.

To see a nice sandbox example of what we mean here, take a look at the OpenAI project where they simulated half a billion games of hide-and-seek. The digital agents in this experiment began with only very simple set of rules for a world, then they evolved their game tactics without any human guidance at all. At some point the aggressor agents figured out how to exploit flaws in the world’s physics engine, and the defending agents learned how to block such attacks.

The true power of deep learning is that our modern computers can run more generations of natural evolution – by multiple orders of magnitude – than biological life is capable of. We strongly pose that any company not researching large-scale simulations is currently missing the whole point of AI.

Specifics

Who to Watch

Leaving the trends now, let’s dive into details and see what the movers and shakers of AI are up to. We’ll begin with the big names and work towards the smaller players.

Elon Musk oversees several companies that are advancing the science of AI.

First, Tesla is always dedicated to improving their self-driving car technology. The most noteworthy news here is that they have componetized the AI “brain” of the car into a single circuit board that can be swapped in and out of an underlying vehicle. It is trivial to imagine this pluggable control module being productized and applied to other complex control systems in fields that are not automotive.

Another Musk company, Neuralink, has been operating in stealth mode for a very long time, but then this last year they finally tipped their hand and showed us what they’re working on. It’s crazy stuff: digital fibers embedded directly into brain tissue to bridge the gap between our minds and digital neural networks. The word on the street is the company broke their silence for recruiting reasons – it was getting too hard to hire top talent into a totally secret company, no matter what tech rock star runs it.

Musk was also one of the founders of the OpenAI project that we mentioned above, though he has since stepped down from management. OpenAI is a pure research lab. This group’s biggest news is they’ve released the largest and most complex version of their GPT-2 natural language model. This is the back side of a smart language interface: not recognizing our words, but instead generating words for us. The GPT-2 language model is sophisticated enough that using it to rapidly create massive amounts of fake news is actually a major concern, and why the group staged out release over most of a year.

Google is clearly a behemoth in AI research.

Google’s most interesting news is they have managed to make their language recognition nets small enough to fit directly onto mobile devices, thus eliminating the latency inherent when sending audio data to the cloud for processing. It only works on newer Pixel phones, but it vastly improves a user’s experience.

Unfortunately they haven’t done much else of interest lately. We learned last year that Google’s YouTube AI is the singular most socially radicalizing force on the planet, yet there has been pretty much zero movement from the company on addressing this problem.

Similarly, Google probably has the current smartest voice assistant of all the big tech companies (see link under Winning the Race, above), but the gains are incremental and no news story about any specifics is particularly worthy of calling your attention to.

Amazon is doing very little AI themselves, but they do host the computers that run most of the internet.

Amazon’s big news this last year was partnering with others to bring quantum computing into their portfolio of Amazon Web Services. Quantum computing is expected to be a step function in our computational ability, once we’ve learned enough to take advantage of it. Deep learning algorithms are very computationally expensive, so whosoever first figures out how to make a quantum deep learning system will likely change the face of our world forever.

China is the other large player to note.

If you measure by patents, last year China claimed two and a half times as many patents for AI technology as US companies did. A lot of this tech hasn’t made it to our side of the ocean yet, but it’s indisputable that eastern tech leads the field when using neural networks for facial recognition and social surveilence. Mostly they seem to be using this to subdue minority cultures, which is itself a trend we probably should be aware of.

So Who’s Losing?

The field of AI is broad, and there’s a lot of underdogs who are still big enough they shouldn’t be forgotten.

Microsoft has shown little output of their own, but they are still trying to buy their way into relevancy.

For example, they contributed the next billion dollars of operating capital to the OpenAI project, and they’ve joined Amazon’s Voice Interoperability Initiative. But direct gains from either of these ventures into Microsoft’s own service offerings have yet to materialize.

Apple is supposedly building a car to rival Tesla, but we know pretty much nothing more. Truly, we’ve seen nothing at all from Apple to suggest that AI is a priority in their daily business.

Facebook is similarly behind in the race for AI. They sponsor some research teams around the world, but if this has any external reach we have yet to see the evidence. We know they use AI internally to personalize your news feed, but it seems unlikely they’re ever going to productize any of their research.

Salesforce is trying to put a smart administrative assistant in a box, just like everyone else, but so far it seems to be just a CRM with a confidence score.

IBM is probably last place among big tech companies. If they have any offering that’s winning in any race, we couldn’t find it.

Still Important

To close out our survey we want to list a couple last companies that, while not big tech, are still doing interesting things.

Boston Dynamics is making androids. Some of their less mobile systems are designed for smart factories, but any media attention is usually given to their walking models. This company’s motion control systems are probably top of the line, though we feel it’s unlikely that anyone except militaries are looking to buy them. As we stated above, the big race in the industry is for Starship Enterprise, not Commander Data.

Nuance owns the Dragon brand, which has lead the world in speech recognition for a couple decades. The company claims – probably rightly – that taking dictation is a different problem from making a voice assistant. That said, the company these days is focusing on smart assistants like everyone else. They still seem to have an edge by some metrics, but it’s unclear if they can keep it given the pace that Google and others are advancing.

Closing

The trend of most industry research is smart assistants, as we’ve seen. Old challenges, like picking the next item on your news feed, or how likely a sales lead is to convert, are beginning to look like simple, well understood problems. These are no longer science, but engineering – we’re just trying to optimize a process that’s already been invented.

Yet the true power of AI is large scale simulations of complex systems, like self-driving cars, or interacting smart agents in a virtual world. Hardware advances are enabling ever larger experiments over literally billions of generations of evolution of the underlying models that we’re simulating.

As these trends continue, even by the end of this year, we expect to see our digital agents get better at predicting us, taking care of us, and shielding us in a generally made up media reality.

Social implications are still beyond the scope of our discussion, but the obvious conclusion is our world is poised for some very big changes.