This is part of a four post series spanning two blogs.
- One post gives a general historical overview of the artificial intelligence business.
- One post specifically covers the history of expert systems.
- One post (this one) gives a general present-day overview of the artificial intelligence business.
- One post explores the close connection between machine learning and (the rest of) AI.
1. “Artificial intelligence” is a term that usually means one or more of:
- “Smart things that computers can’t do yet.”
- “Smart things that computers couldn’t do until recently.”
- “Technology that has emerged from the work of computer scientists who said they were doing AI.”
- “Underpinnings for other things that might be called AI.”
But that covers a lot of ground, especially since reasonable people might disagree as to what constitutes “smart”.
2. Examples of what has been called “AI” include:
- Rule-based processing, especially if it is referred to as “expert systems”.
- Machine learning.
- Many aspects of “natural language processing” — a term almost as overloaded as “artificial intelligence” — including but not limited to:
- Text search.
- Speech recognition, especially but not only if it seems somewhat lifelike.
- Automated language translation.
- Natural language database query.
- Machine vision.
- Autonomous vehicles.
- Robots, especially but not only ones that seem somewhat lifelike.
- Automated theorem proving.
- Playing chess at an ELO rating of 1600 or better.
- Beating the world champion at chess.
- Beating the world champion at Jeopardy.
- Anything that IBM brands or rebrands as “Watson”.
That last bit is awkward, as IBM is doing the industry a major disservice via its recklessly confusing Watson marketing, which is instantiating Monash’s First Law of Commercial Semantics — Bad jargon drowns out good. I suspect there’s an interesting debate under it all, in which IBM stands almost alone against the whole rest of the industry by sticking to the old academic belief that sophisticated knowledge representation is the key to AI. But it’s hard to be sure, because IBM’s Watson marketing is so full of smoke that reality, if any, doesn’t show through.
3. When I think of present-day AI commercialization, what comes to mind is mainly:
- Multiple efforts in speech recognition, from Google, Microsoft, Apple, and Nuance Communications. (I’m not sure whether Apple’s is mainly in-house or mainly outsourced.)
- Other natural language efforts, such as Google’s in machine translation.
- Technology related to robots and autonomous vehicles, specifically in machine vision, other senses (e.g. touch), and reactions (e.g. driving decisions).
- Google is the most visible player here. It’s gotten a lot of press for driverless automobiles, and it bought up a lot of robotics companies when they were hurting due to a hiatus in DARPA funding.
- Large auto companies will surely compete.
- Gesture interpretation and similar kinds of recognition.
- Microsoft has the most visibility here, due to Kinect, and is trying to bring similar technology to general computing.
- Facebook, Google et al. are making major investments into the closely related area of virtual reality. Facebook is also building an AI team.
- Machine learning.
- Machine learning in general can be regarded as part of AI, at least historically.
- Machine learning is a key component of many AI efforts. Google in particular has made a big fuss about it, suggesting that data is generally more important than algorithms.
- Whatever parts of the IBM story, if any, are actually real.
So with one big exception, commercial AI seems to be concentrated at a small number of behemoth companies. The exception is machine learning itself, which is being adopted and developed on a much broader basis.
5. Some of the reasons for AI’s concentrated industry structure lie in general business and economics.
- A large company can risk research with unclear payoffs a lot more easily than a small one can.
- AI is prestigious and/or cool. Some large companies like to indulge in stuff like that.
Yes, those reasons are somewhat counteracted by the facts that:
- VCs know they’re investing in companies whose eventual exit will likely be an acquisition.
- Some of those acquisitions are for a LOT of money.
But I think they apply even so. And by the way — to date, most AI companies have not been acquired for very high prices.
6. Some of the reasons for AI industry concentration are more specifically technological.
- Some AI — e.g. speech recognition or autonomous vehicle navigation — could be the “sizzle” that differentiates offerings in huge business sectors. Thus, a “win” in AI could have more value to an already-large electronics, search or automobile company than to a startup.
- The largest companies in those huge sectors can afford huge amounts of training data, or may even get it as a byproduct of their other activities. Hence they can more easily afford massive exercises in the relevant machine learning.
My paradigmatic example for the latter point is Google with anything connected to search, such as translation (which it does of search results) or natural language recognition (which it does of search queries).
If you want to do an AI startup, those are some of the competitive factors that you need to beat.
- An earlier version of some of this material was in my January, 2014 post on The games of Watson.
- Earlier this year, I posted about robotics.
- There is quite a bit of AI humor.