硅谷大佬为什么都在豪赌人工智能

刘颖

Last week, Tesla CEO Elon Musk and fellow tech kingpins committed $1 billion to researching artificial intelligence. The group’s findings would be made available for the world. The possibilities where AI might help include the ability to detect anomalies in images of cells to detect cancer, programming robots that can interact with humans, and building programs that could help teach kids at their pace of learning in a more individual style.

Behind this feel-good effort is a hint at the priorities of some of the biggest names in Silicon Valley. It also provides an understanding of how AI or machine learning, as the technology is often called, has the potential to remake the tech world in the same way the web did in the mid-’90s.

Worries about artificial intelligence have sparked headlines exclaiming that AI could bring about the death of humanity as smart machines become so much smarter than us they wipe us out, not out of malice, but because we’re simply in the way of their own goals. The most optimistic ones focused on the possibility of sex robots that can carry on conversations.

But in reality, AI has existed for over a decade. It already plays a big role in technologies that we take for granted like Apple’s Siri personal assistant, IBM’s Watson Jeopardy-winning computer, and even the autopilot feature that Teslarolled out in its cars earlier this year.

And before AI can destroy humanity, or provide sexual satisfaction, it has to get better. Much better. And the launch of OpenAI, the billion-dollar nonprofit research center announced this week, opens a window into what some of the big thinkers in computer science and business consider as opportunities and challenges.

First, as analyst Ben Thompson, who writes over at the site Stratechery, pointed out in an essay about the topic, OpenAI’s creation can be read as a manifesto, or as a recruiting ad for top research talent.

Thompson looked past the do-gooder language of the OpenAI blog post, which talks about ensuring that commercial interests don’t hijack the promise of artificial intelligence research. Instead, he focused on the final line of the third paragraph of the introduction, which reads “We hope this is what matters most to the best in the field.”

The fear is that Google, Facebook, and Chinese search engine Baiduare luring all of the machine learning talent to their companies using a sales pitch that hires can work on some of the most complex social problems of our era. Each company uses huge pools of data to help train sophisticated machine learning algorithms.

Data is the lifeblood of AI. To train computers to learn more like humans, you have to feed them tens of thousands of examples of something. Depending on what type of outcome you are hoping for, the examples can be photos, maps, or words. The computers try to understand what elements of those examples define what makes a cat a cat in an image or what gives meaning to a certain word. The algorithm then gives a statistical weight to each guess that helps the computer “learn” what the right answer is. The computer scientist helps train the algorithm by giving feedback and more examples along the way.

That’s why none of these companies ever wants to throw away data. It may come in handy for AI training someday. And that’s why the promise of using something like Tesla’s car data for building algorithms might be enough to get a researcher excited to work with OpenAI instead of Google.

Sam Altman, a co-chair at OpenAI, tells Fortune that data from Tesla would be made available to researchers working at OpenAI. He said he would also work to make data from startups that go through Y Combinator, the accelerator program he leads, available for researchers at OpenAI as well.

“There are also plenty of publicly available data sets on the Internet,” Altman said. Researchers could use those to come up with new tools and algorithms that will advance AI as well.

The second element designed to attract talent to OpenAI is its nonprofit status and its pledge of openness. It’s not that Facebook and others aren’t open with their research. They publish their research fairly quickly. Google, however, tends to wait until it has gained a significant strategic advantage from a new findings before publishing. But it is still made public.

SerkanPiantino, director of Facebook’s AI research program, emphasized the importance of openness in a conference call ahead of premiering his company’s new servers designed especially for training computers to learn earlier this month. Facebook’s engineers expect the work they do to be contributed back to the open source community. Thus, Facebook contributes code to the community in part because that keeps its civic-minded engineers happy.

But the race for talent isn’t the only reason OpenAI exists. The development of true artificial intelligence is going to remake software. And every business wants to be part of that shift.

“The way software is eating the world today, well, AI will do that to software,” says Amir Husain, CEO of Spark Cognition, an AI security startup in Austin, Texas.

He explained that many kinds of business software that replaced paper documents and in filing cabinets will eventually be transformed into a new format. And that format will likely be more user-friendly because of hard work done by artificial intelligence behind the scenes.

“All of these categories will be destroyed and remade, so there’s a lot of economic potential locked up in this,” says Husain. “It’s sort of like being the only guy in 1995 who knows HTML.”

And that, more than anything, is why the big brains in Silicon Valley and at other companies left out of the OpenAI effort are hustling to stake a claim in this space. Rob High, an IBM Fellow, and VP and CTO of IBM’s Watson Group, explained that the computing giant is interested in learning more about the organization and getting involved.

IBM, which learned about the OpenAI group on Friday like nearly everyone else, has a decades-long program in artificial intelligence through Watson. The company hopes that it will help it weather the shift from web-based software to new A.I.-related services.

But IBMis also building an entirely new type of chip designed for artificial intelligence modeled on the human brain, called a synaptic chip. As far as hardware for AI goes, IBM is the most serious player in the space.

Following is Nvidia, which makes graphics processors that are actually the preferred chip used today for training computers to learn.

That gets us back to Altman, from OpenAI, and the plans for the nonprofit. The short-term goals, he said are to build tools and algorithms that will be shared publicly. But in the long term, better hardware is needed to build AI that can perform more like a human.

“If you think about building better AI and modeling it after the human brain, more hardware research and better hardware will be important,” Altman says. “But today that is not our primary focus.”

That might be why Altman says OpenAI only spoke very casually with a person who was involved with Watson at IBM, instead of going through formal channels to try to get Big Blue involved with the project. (And why IBM found no record of someone from OpenAI contacting it at all). Or perhaps there’s simply a divide between the Silicon Valley practice of calling anything with machine learning involved AI and promoting its involvement in new product launches. Meanwhile, IBM, which brands all of its AI efforts under Watson and cognitive computing, may have confused the public.

Plenty of other companies have their own efforts in artificial intelligence. For example, Apple has hired researchers, but has reportedly found it to be tough to recruit experts. In part, it’s because the company doesn’t want to share the research results. Microsoft MSFT 1.29% also has AI research in natural language for its Skype translation efforts and computer recognition that are worth mentioning as well.

This is a cheaper way to solve the problem and more Amazon-like. Outside of the giant tech firms, startups, industrial giants, researchers and more are all experimenting with using AI. If OpenAI really does build broadly useful tools, that could help advance science for everyone.

Altman says it’s too soon to list OpenAI’s research priorities. It will work on tools and algorithms, but the specific areas where it will focus are unsure. But he said he would consider it a success if the organization, within one year, publishes “some seminal paper that drives the state of the art forward.”

However, it’s clear that technologists supporting the project and those working on AI in general, have much larger goals.