Wednesday, June 22, 2016

Early Indications June 2016 What do we make of Artificial Intelligence?

For context, an old joke:

Q: What’s harder than solving the problem of artificial intelligence?

A: Fixing real stupidity.

In many current publications, the technical possibilities, business opportunities, and human implications of artificial intelligence are major news. Here’s just a sample:

-a computer beat a grand master at Go about 10 years before many predicted such an outcome would be possible.

-Google is repositioning itself as an AI company, with serious credibility. IBM is advertising “cognitive computing,” somewhat less convincingly, Watson notwithstanding.

-Venture capital is chasing AI-powered startups in every domain from ad serving to games to medical devices.

-Established players are hiring top talent from each other and from academia: Toyota, Amazon, Uber, and Facebook have made noise, but Google remains the leader in AI brainpower.

-Corporate acquisitions are proceeding apace: just this week Twitter bought a London-based company called Magic Pony, which does image-enhancement, for a reported $150 million. Those kinds of numbers (shared around a team of in this case only 11 PhDs) will continue to attract talent to AI startups all over the world

Despite so much activity, basic answers are hard to come by. What is, and is not, AI? By which definition? What is, and is not possible, for both good and ill? The more one reads, the less clarity emerges: there are many viable typologies, based on problems to be solved, computational approaches, or philosophical grounding. I can recommend the following resources, with the caveat that no consensus emerges on anything important: the whole concept is still being sorted out in both theory and practice.

The Wikipedia entry is worth a look.

Here's a pretty good explainer from The Verge.

The Economist reports on the sizable shift of top research talent away from universities into corporate AI research.

Here’s a New Yorker profile of the Oxford philosopher Nick Bostrom.

Oren Etzioni has a piece on “deep learning” (neural networks at very large scale, best I can make out) in a recent issue of Wired:

Elon Musk called AI humanity’s “biggest existential threat” in 2014.

Frank Chen at Andreessen Horowitz has a very good introductory podcast explaining the recent boom in both activity and progress.

Apple is trying to use AI without intruding into people’s identifiable information using something called differential privacy.

Google’s AI efforts, by contrast to Apple’s, build on the vast amount of information the company’s tools know about people’s habits, web browsing, searches, social network, and more.

Amazon has multiple horses in the AI race, and recently made a high-profile hire.


Despite the substantial ambiguity related to the macro-level abstraction that is AI, several generalizations can be made:

1) Defining AI with any precision is problematic. Vendors including Google (“deep learning”) and IBM (“cognitive computing”) are well served by a certain degree of mystery, while the actual mechanics of algorithm tuning are deeply technical and often tedious. There are live questions over how completely the use of a given algorithm (a Kalman filter, used in both econometrics and missile guidance, or a simulated annealing optimization, used in supply chains and elsewhere, to take two examples) is “doing” AI.

2) AI can work spectacularly well in highly defined domains: ad placement, cerebral games, maps and directions, search term anticipation, and more and more, natural-language processing as in Siri/Cortana/Alexa. Leave the domain, however, and the machine and its learning are lost: don’t ask Google Maps to pick a stock portfolio, or Siri to diagnose prostate cancer. The challenge of “general AI” remains a far-off goal: people are more than the sum of their map-reading, pun-making, and logical generalization abilities.

3) Hardware is a key piece of the recent advances. Computer graphics processors feature a parallel architecture that lends itself to certain kinds of AI problems, and the growth of gaming and other image-intensive applications is fueling better performance on the computing frontiers of machine learning. Google also recently announced a dedicated hardware component, the Tensor Processing Unit, built specifically to handle machine learning problems.

4) “Big data” and AI are not synonymous, but they’re cousins. Part of the the success of new machine learning solutions is the vast increase in the scale of the training data. This is how Google Translate can “learn” a language: with billions of examples, not a grammar, dictionary, or by ear.

5) It’s early days, but one of the most exciting prospects is that humans can learn from AI. Lee Sedol, the Go player, says he is now playing better than before his loss to the Google computer. Whether with recipes (for tire rubber or salad dressing), delivery routes, investment strategies, or even painting, getting inspiration from an algorithm can potentially spur people to do great new things. Shiv Integer is a bot in the 3D printing site Thingiverse, and the random shapes it generates are fanciful, part of an art project. It’s not hard to envision a more targeted effort along the same lines, whether for aircraft parts or toys. I would also bet drug discovery could benefit from a similar AI approach.

6) The AI abstraction is far more culturally potent than the concrete instances. The New Yorker can ask “Will artificial intelligence bring us utopia or destruction?” (in the Bostrom article) but if you insert actual products, the question sounds silly: “Will Google typeahead bring us utopia or destruction?” “Will Anki Overdrive (an AI-enhanced race-car toy) bring us utopia or destruction?” Even when the actual applications are spooky, invasive, and cause for concern, the headline still doesn’t work: ”Will the FBI’s broad expansion of facial recognition technology bring us utopia or destruction?” The term "AI" is vague, sometimes ominous, but the actual instantiations, while sometimes genuinely amazing (“How did a computer figure that out?”), help demystify the potential menace while raising finite questions.

7) Who will train the future generations of researchers and teachers in this vital area? The rapid migration of top robotics/AI professors to Uber, Google, and the like is completely understandable, and not only because of money. Alex Smola just left Carnegie Mellon for Amazon. In his blog post (originally intended for his students and university colleagues), he summarized the appeal: less bureaucracy, more data, more computing power.

          “Our goal as machine learning researchers is to solve deep problems (not just in deep learning) and to ensure that this leads to algorithms that are actually used. At scale. At sophistication. In applications. The number of people I could possibly influence personally through papers and teaching might be 10,000. In Amazon we have 1 million developers using AWS. Likewise, the NSF thinks that a project of 3 engineers is a big grant (and it is very choosy in awarding these grants). At Amazon we will be investing an order of magnitude more resources towards this problem. With data and computers to match this. This is significant leverage. Hence the change.”

It’s hard to see universities offering anything remotely competitive (across all 3 dimensions) except in rare cases. Stanford, MIT, University of Washington, NYU, and Carnegie Mellon (which lost most of an entire lab to Uber) are the schools I know about from afar with major defections; those 4 (absent NYU) are among the top 5 AI programs in the country according to US News, and I wouldn’t feel too comfortable as the department chair at UC-Berkeley (#4).


As in so many other domains (the implications of cheap DNA sequencing; materials science including 3D printing; solar energy efficiency), we are seeing unprecedentedly rapid change, and any linear extrapolations to predict 2025 or even 2020 would be foolish. Perhaps the only sound generalization regarding AI is that it is giving us strong reinforcement to become accustomed to a world of extreme, and often troubling, volatility. Far from the domain of machine learning, for example, a combination of regulations, cheap fracking gas, and better renewable options led the top US coal companies to lose 99% — 99%! — of their market capitalization in only 5 years. Yet other incumbents (including traditional universities) can still look at our world and say, “I’m immune. That can’t happen here.” Helping expand perspectives and teach us flexibility may be one of AI’s greatest contributions, unless human stupidity is too stubborn and wins the day.

Friday, June 10, 2016

Early Indications May 2016: Technology and inevitability

Human nature drives us to look backwards and see a series of developments neatly explaining the current situation: we all exhibit hindsight bias is some form. It’s much harder to look back and recapture the indeterminacy in which life is lived in the present tense. Technological history is particularly prone to this kind of thought and rhetoric: the iPhone was famously (but not universally) mocked upon its introduction, to take but one example; looking for “the next Microsoft” or the “next Google” is another manifestation. The project “singularity” of digital cognition surpassing the human kind builds on this kind of logic. Coming in June, longtime tech observer Kevin Kelly’s new book is called simply The Inevitable.

It’s important to remember, however, that merely inventing (or imagining) a technology is a far cry from getting it into garages, factories, living rooms, or otherwise achieving successful commercialization. The low success rate for university technology transfer offices bears this out: a great molecule, material, or method does not a successful product make, absent entrepreneurship, markets, and other non-technical factors. This month I’ll run through a few technologies, some well-known and visible, others largely forgotten, that failed to achieve market success. I do this less out of nostalgia and more in the interest of tempering some current projections with a reminder that luck, competition, timing, and individual drive and vision still matter.

1) Very Light Jets
Led by Eclipse but also joined by Honda, Embraer, and others, the late 1990s stand as the height of the promise of a small, cheap (under $1 million new) aircraft that could both lower the barrier to personal jet ownership and fuel the rise of short-hop air taxi services. Eclipse shipped far later than promised at more than twice the projected cost, and performance problems were numerous: tires needed frequent replacement, the windscreens cracked, fire extinguishers leaked corrosive chemicals into sensitive components, the computerized “glass cockpit” failed to perform, and so on. A few air taxi firms went live (such as DayJet), but failed in the 2008 financial crisis, as did Eclipse. Honda, meanwhile, is prone to showing HondaJets in company advertising, but as of last December, had delivered a total of one plane to a paying customer. Sale prices are in the $4 million and up range — more than a used Hawker or similar mainstream business jet.

2) Flying cars
However intuitive the appeal, flying cars remain a niche market occupied primarily by mad-scientist visionaries rather than established production teams and facilities. The latest attempt, the Aeromobil, is claimed to be ready for market in 2017. The video is pretty impressive. Much like VLJs, flying cars have failed as much for economic reasons as technical ones. Building such a complex vehicle is not cheap, and safety considerations raise the product’s cost in multiple ways: FAA certification, spare parts management, expensive short-run production, and insurance factor into the actual operational expenses. Some of these expenses are out of the control of the aforementioned visionary (and in the Eclipse case, Bert Rutan has throughly impressive credentials), while other business challenges, including marketing, are common in tech-driven startups: who will buy this and what problem does it solve for a critical mass of real people?

3) ATT Picturephone

Here is ATT’s website, verbatim:

"The first Picturephone test system, built in 1956, was crude - it transmitted an image only once every two seconds. But by 1964 a complete experimental system, the "Mod 1," had been developed. To test it, the public was invited to place calls between special exhibits at Disneyland and the New York World's Fair. In both locations, visitors were carefully interviewed afterward by a market research agency.

People, it turned out, didn't like Picturephone. The equipment was too bulky, the controls too unfriendly, and the picture too small. But the Bell System* was convinced that Picturephone was viable. Trials went on for six more years. In 1970, commercial Picturephone service debuted in downtown Pittsburgh and AT&T executives confidently predicted that a million Picturephone sets would be in use by 1980.

What happened? Despite its improvements, Picturephone was still big, expensive, and uncomfortably intrusive. It was only two decades later, with improvements in speed, resolution, miniaturization, and the incorporation of Picturephone into another piece of desktop equipment, the computer, that the promise of a personal video communication system was realized."

*I’m sure the story of exactly _who_ in the Bell System drove this $500 million boondoggle is fascinating, if heavily revised.

4) Voice recognition software
Bill Gates is very smart, and obviously has connected some pretty important dots (as in the Internet pivot Microsoft executed in the late 1990s). On voice recognition, however, “just around the corner” has yet to come to pass. His predictions began in earnest with his 1995 book The Road Ahead, and in numerous speeches since then (well into the 2000s), he doubled down. Even now, in the age of Siri/Alexa/Cortana, however, natural-language processing is a very different beast compared to replacing a keyboard and mouse with talking. Compare two statements to see the difference: “What is the temperature?” vs “highlight ‘voice recognition software’ and make it bold face.”

5) Nuclear civilian ships
President Dwight Eisenhower wanted both Americans and citizens of other nations to temper their fears of military atomic and nuclear weapons by encouraging peacetime uses (his “Atoms for Peace speech” was delivered in 1953). The NS Savannah, a nuclear cargo ship, was intended to be a proof of concept, and it remains a handsome vessel a half-century on. The ship toured many ports of call for publicity and drew good crowds. Reaction was mixed, and the fear of both nuclear accidents and waste leaking into the oceans proved prescient as the US vessel and, later, a Japanese civilian ship both experienced losses of radioactive water. Although operational costs are low, the high up-front investment and, more critically, unpredictable decommissioning and disposal costs presented unacceptable risks to funding agencies or banks. Despite 700 military nuclear vessels becoming standard pieces of national arsenals, nuclear civilian craft have never caught on (with the exception of a few Russian icebreakers). A great BBC story on the Savannah (now moored in Baltimore) can be found here.


6) 3DO gaming console
After the 1980s, in which Sony’s Betamax format lost out to Matsushita’s VHS, consumers remained wary of adopting a technology in the midst of a format war. The lesson has been learned and relearned in the succeeding decades. In the early 1990s, Trip Hawkins (who founded Electronic Arts) helped found a new kind of console company, one based on licensing rather than manufacturing. The effort attracted considerable attention, but numerous problems doomed the effort. Sony and Nintendo can subsidize the cost of their hardware with software royalty streams; this is a basic element of platform economics as seen in printer cartridges and other examples. The 3DO manufacturers lacked this financial capability, so a high selling price was one problem. In another basic of platform economics, software and hardware must be available in tandem, and there was only one game — Crash ’n Burn, ironically enough — available at US product launch. In Japan, a later launch helped enable better reception as six game titles were available, but within a year the platform had become known for its support for pornographic titles, so general adoption lagged. 3DO clearly had some technically attractive elements (some of which were never included in the Nintendo 64 and Sony Playstation that followed) but the superior technology failed to compensate for market headwinds.

7) Elcaset
Unless you’re an audio enthusiast of a certain age, this is deep trivia. Sony introduced this magnetic tape format in 1977, and it was clearly technically superior to the audio cassette that had become entrenched by this time. The tape was twice as wide, and moved twice as fast, improving the signal/noise ratio and allowing for more information to be recorded, thus increasing fidelity. Like a VHS deck, tape handling was done outside the plastic shell, improving performance further. Unfortunately, the added performance came at a cost, and few consumers saw any reason to embrace the odd new format, which was supported by TEAC, Technics (Matsushita), and JVC as well. Also, no pre-recorded titles were available: this was the time when “home taping is killing music” — the 1980s UK anti-cassette campaign was later dusted off for the Internet age by the Norwegian recording industry association — and label execs were of two minds with regard to cassettes. In a curious twist I only recently learned about, Sony sent the remaining inventory of players and tapes to Finland after a distributor there won the wholesale auction, where many of the machines, well-made as they were, continued to work well for decades afterward.

This somewhat random collection of technologies holds very few generalizations. Having high-ranking executive sponsorship — up to and including the President of the United States — failed to compensate for deep fears and uncertain cost projections. Some failures came from corporate labs, others from entrepreneurs. Platform economics prove to be critical, whether for hardware and software, spare parts and airfields, or communications technologies. In the end, the only true generalization might be that markets are fickle, and there’s very little technology that is truly inevitable in its adoption.