Musk and Zuckerberg Got Into a Fight Over AI. What They Said — and What They Left Unsaid

All of these men are in fact AI's supporters, and none of them care how many jobs it takes away.

Two Silicon Valley tech leaders, Elon Musk and Mark Zuckerberg, picked a fight with each other from opposite ends of the country.

It started with a barbecue in Zuckerberg's backyard. In a Facebook Live broadcast last weekend, Facebook CEO Mark Zuckerberg stood in that backyard and answered questions from users.

Asked about the theory that AI is a threat, Zuckerberg gave a very optimistic answer: "AI is going to make the world better." He said the word "optimistic" twice, then added, "Some people hype up doomsday scenarios about AI. I think that's irresponsible."

No one in Silicon Valley preaches the AI apocalypse louder than Tesla founder Elon Musk.

So someone tagged Musk on Twitter, and the CEO, who picks fights on social media all the time, answered fast: "I've talked to Mark about this. His understanding of the subject is limited."

It was not a persuasive argument. But the two men do stand for the two mainstream attitudes toward AI in Silicon Valley.

Zuckerberg rarely speaks up about the AI-threat theory. Neither do the CEOs of the other biggest tech companies — Google, Facebook, Amazon, Apple. All of them are pouring money into AI.

Musk stands for a minority. He, Stephen Hawking and others signed an open letter urging the world to take the threat of AI seriously. With another preacher, Sam Altman, president of Silicon Valley's biggest incubator, Y Combinator, and others, he set up the nonprofit OpenAI, meaning to build an industry association for AI development.

In the middle of this month, Musk called on U.S. states to act at the National Governors Association's summer meeting, telling them to start writing AI laws. He called AI the biggest problem our civilization faces. As an example, he said AI could start a war by manufacturing fake news: manipulating information is sometimes more powerful than weapons.

No One Is Worried About AI Taking Your Job

Over the past two years, as Lee Sedol and then Ke Jie lost at Go to artificial intelligence, worry about the fate of the human race has flared up again.

Then came the practical uses — self-driving cars, voice assistants, automated stock trading, automated insurance sales, automated news writing — and more and more people began to worry that AI will take their jobs.

Using AI to recommend information, predict crime and watch over the public, meanwhile, has raised further questions about how the technology gets used.

None of this is what Silicon Valley's tech leaders worry about. On this they agree almost word for word.

Musk thinks falling demand for labor will bring a golden age in which everyone has an income: "With the automation AI brings, we'll very likely see a new era in which everyone has a basic income. I'm not sure what people will do once they don't have to work, but perhaps they'll have time for other things that are more complex and more interesting. And of course there will be a lot more leisure."

Bill Gates, who also worries about an AI apocalypse, thinks taxing robots would fix the social problem: "Robotic automation should be taxed too. If a human worker is taxed on income — income tax, social security and so on — a robot can be taxed for the same amount of work. The world needs the same goods and services, but freeing up labor can give the elderly a better life, narrow the gap between classes and provide care for children with special needs."

Andrew Ng, the former Stanford professor who ran major AI projects at Google and Baidu, says many of the AI projects he has worked on could wipe out tens of thousands of jobs. AI can create more wealth, and though it will be hard, it can create a fairer society. People who lose their jobs can get a better education and find other work.

Google founder Larry Page, who never comments on the AI-threat theory, thinks a new system of working hours will fix the problem of shrinking work. He said: "Most people want more time with their families. In the AI era we might have two weeks off a month, or a four-day week. Cutting working hours would solve the jobs problem. Add a few more positions on top of that and you can make sure more people have work."

Roughly: AI will reduce the amount of work, but people will have more time, new trades will appear later on, and you may even get paid without working at all.

That claim is doubtful. In reality, tech companies have used technology to raise efficiency and cut the need for physical labor, and they have not offered to compensate the people they replaced. The basic laborers they do keep on the payroll earn no more than their peers elsewhere, and some earn less: Amazon's warehouse workers and Apple Store staff both hold low-paid jobs. The thousands of content moderators who work for Facebook are all outsourced.

Research over the past decade or more into what robots do to employment and wages has found the exact opposite of what Silicon Valley predicts.

But this optimism is the Silicon Valley leaders' consensus on AI.

What They Are Arguing About Is Skynet

The only real disagreement between the Musk camp and the Zuckerberg camp is probably whether an AI apocalypse can happen at all.

They have never sat down in one room to argue it out, but each of them has addressed the question somewhere. We have arranged their remarks into an imaginary debate. Every line below is a direct quote.

Stephen Hawking:

The development of full artificial intelligence could end the human race! All of us should stop and shift our research from raising AI's capability to maximizing AI's social benefit.

Amazon founder Jeff Bezos:

To keep a company vibrant, you have to make decisions that are both fast and good... Speed matters in business. Besides, an environment where decisions get made fast is more fun.

Elon Musk, who founded the nonprofit AI organization OpenAI:

I'll keep ringing the alarm bell for you! Until people see robots killing people in the street, they won't know how to deal with AI, because it sounds too unreal.

Former head of AI at Google and Baidu, Andrew Ng:

There may be a killer-robot race in the future, but right now I am certainly not working on making AI "not evil" — for the same reason I'm not worrying about overpopulation on Mars.

Google founder Sergey Brin:

If all the world's information were wired into your head, or into the head of an AI smarter than you, you would certainly be better off than you are now.

Facebook founder Mark Zuckerberg:

When I hear people say AI is going to hurt people in the future, I think: technology can always be used for good or for ill, you have to handle it carefully, build it carefully and apply it carefully.

Microsoft founder Bill Gates:

Humans have a general learning algorithm that came out of evolution. It runs on an extremely slow computer with very little storage, and we have to use a ridiculous tool called a mouth to send data to other "computers." Every new "computer" that gets made has to start all over again — it doesn't even know how to walk. We have never built an algorithm like that in software. But once that algorithm is written into silicon, it will pass humans at once and become superhuman. As soon as it is plugged into the internet it can read every book, every magazine and everything else. We have already prepared the raw material for a superintelligence. I try not to push this view too hard. But when someone tells me it isn't a problem, I really can't agree. How can they not see what an enormous challenge this is?

Chairman of the Silicon Valley incubator YC, who founded OpenAI with Musk, Sam Altman:

We still don't know how human intelligence works, so it's hard to prove how far today's machine "simulation" of intelligence has come. The direction could be completely wrong, or we could be one algorithm away... Plenty of people believe superintelligence is dangerous but think it's far off, maybe something that will never happen. That is a careless and dangerous way to think.

Author of Zero to One and a well-known Silicon Valley investor, Peter Thiel:

If we had an AI stronger than humans, it would be like aliens landing on Earth. If aliens showed up, we certainly wouldn't ask them whether they'd like to help fix the economy. We would ask them whether they'd come to kill us.

Founder of DeepMind, the AI company Google bought, Demis Hassabis:

The problems we face now are complicated ones. Quantum physics, macroeconomics, disease — they all involve huge amounts of data. Let's face it: without help, humans can barely make sense of them. AI is only a tool for us to solve society's problems...

Andrew Ng:

As a veteran AI researcher... I think AI replacing jobs is a big problem, and that's what we should be paying attention to. All the science-fiction and dystopian stuff is a distraction.

Demis Hassabis (not finished yet):

My personal dream is that once we've got the AI technology working, I'll use AI to create a new system of scientists, and use AI to speed up scientific and medical research. The faster the better.

Peter Thiel, who is also an investor in DeepMind:

I remember the last DeepMind investor meeting. One investor there told me this might be his last chance to take out Demis Hassabis — to put a bullet in him. (Not long after that board meeting, Google bought DeepMind.)

What's the Difference Between the Superintelligent AI Musk and His Camp Fear and the Xiaoice on Your Phone?

Unlike Skynet in The Terminator, which fires nuclear missiles and enslaves humanity, the AI that actually touches most people's lives today is still very weak.

Microsoft's Xiaoice, the smoothest talker among Chinese-language chatbots, can keep a conversation going for a few dozen turns, guess which celebrity you have in mind, and finish by singing you a song. But she is a long way from being a real friend.

Google Assistant, the strongest of the intelligent assistants, can put together your schedule from your chat records, pick your route out in advance, or tell you the status of a flight you once looked up. But it still can't write a diary entry for you from an outline you hand it.

Google Translate and Microsoft Translator know more languages than anything else. They can translate nearly every language on earth, and even understand High Valyrian and Klingon. Faced with a new language they haven't studied, they can do nothing.

In other words, today's AI is at most an automated tool for one particular function. Read "artificial intelligence" strictly and these systems are artificial but not yet intelligent. In the AI industry's own terms, they are "weak AI."

On top of this basic AI you could build general AI, and finally reach the level of superintelligent AI.

But even with the tiers laid out, the most we know is that "general AI" would think and learn roughly the way humans do, draw inferences from one case to another, perhaps even develop an independent consciousness. What a superintelligent AI beyond human capability would actually be is, for now, pure imagination.

And so-called artificial intelligence had nothing to do with human "intelligence" to begin with. The biggest problem is that nobody yet knows how the human brain works, and there is no algorithm for simulating how it thinks.

The story of Jeff Hawkins — Palm founder, well-known Silicon Valley tech entrepreneur and brain scientist — shows the limits of AI research well.

After Hawkins joined Intel in 1979, a strong interest in brain science took hold of him. But mainstream research at the time knew nothing about the brain. In September 1979, Scientific American ran a special issue on the brain in which Dr. Francis Crick, one of the discoverers of DNA, wrote that humanity knew almost nothing about it. Brain science, Crick said, lacked not only a theory but even a framework.

Inspired to throw himself into brain science, Jeff Hawkins applied to study at MIT's Artificial Intelligence Laboratory as a way into research on the brain. As a computer science researcher, he believed then that artificial intelligence and human intelligence — that is, the science of the brain — could not be separated.

MIT's AI Lab turned his application down. MIT held at the time that artificial intelligence was computer science and had little to do with the brain.

Having made money running companies, Hawkins founded the Redwood Neuroscience Institute in California in 2002, devoted to research in brain science. In a talk he said that brain research matters enormously: only by understanding how the brain works can we build machines that have intelligence.

But AI research still works on the premise of knowing nothing about the human brain. Today's AI — whether a chatbot like Xiaoice or a self-driving machine like the Tesla Model S — is built on the model of "input data, output behavior."

The most talked-about AI project of the past year, the Go-playing AlphaGo, shows fairly typically how so-called "weak AI" learns.

AlphaGo rests on an algorithm that teaches the machine the basic rules of Go, but raising its playing strength takes a lot of samples — game records. At first AlphaGo learned by imitating human game records. Its biggest breakthrough was a mode called "reinforcement learning": let the machine play itself, generate more game records, and keep enough variation among the samples that they don't converge. In the end AlphaGo passed human players.

Clearly, AlphaGo's way of learning is nothing like a human's.

In August 2016, QDaily interviewed Qi Lu, then Microsoft's global executive vice president, at the Xiaoice launch. Asked how computer science can build artificial intelligence at a time when biologists still don't know how humans think, Lu gave an answer that explains a lot.

"There's an idea in engineering that even though humans know how birds fly, inventing the airplane was not an imitation of birds," he said. Computer engineering, he said, can build a system that imitates the process by which humans learn things and produce behavior — and that is artificial intelligence.

In other words, today's AI is still only an imitation of intelligent behavior. Put it in human terms: weak AI is about behavior — you say something to it, it answers, and intelligent behavior has occurred. But a general AI would be a "person" who keeps thinking even when you aren't talking to it, the way you or I lie in bed and let our minds work. Even in silence, your intelligence is still there.

General AI, all in all, remains fairly far off. As for the more advanced superintelligent AI, we can only imagine it from what humans do: you give an order — "predict how this war will go" — and the AI searches the internet, analyzes the generals' characters, the trend of the news, shifts in the economy, the outcomes of similar battles in history, and produces a report within a second. That would roughly count as superintelligent AI.

The slow pace of research is also why the other side thinks Musk and his camp are overthinking it. Silicon Valley, after all, is so far only building more, faster, more usable "tools," and doesn't look any closer to "intelligence."

What Musk and his camp propose now is a system for assessing degrees of intelligence, real technical contingency plans for an AI disaster, and rules like the "Three Laws of Robotics" to govern how AI behaves. Regulation has to step in here too.

On Keeping AI From Doing Harm, Musk and His Camp Have Left Some Things Unsaid

OpenAI attracts investment and pulls small companies in. Besides calling for regulation and guarding against the AI threat, the other thing it does is gather the strength of small companies to close the gap with the big ones in AI research and development.

You can see it in the original participants alone.

OpenAI has a noble mission: to keep AI research from going astray. But of the world's five biggest tech companies — Apple, Google, Facebook — not one takes part in it.

Microsoft and Amazon do appear on the list of supporters, but only because OpenAI's participants run their computing on those companies' cloud platforms.

Each of these five companies has more than a billion users, holds a huge amount of data, and has assembled a team of at least several hundred people to develop AI.

OpenAI's participants, meanwhile, are all companies that want to do AI research but, next to those giants, are short on data and capital.

Elon Musk's own company, Tesla, puts its weight on autonomous driving — a future that leans hard on AI.

As boss of the incubator YC, Altman has incubated dozens of startups valued anywhere from a few billion to tens of billions, among them many AI projects: the chatbot msg.ai, the open-source deep learning framework Skymind.

Two other participants, LinkedIn CEO Reid Hoffman and Peter Thiel, are both well-known investors. Thiel especially. His own startup, Palantir, is a data company built around big data and AI, whose main business is helping government law-enforcement agencies analyze potential threats, among other things. And as an early investor in DeepMind — the company that made AlphaGo and was later bought by Google — he holds more than 25% of the shares, far more even than founder Hassabis.

The venture capital firm Andreessen Horowitz once said that in frontier fields like AI, the resources big companies command let them attract more researchers — Google, Facebook, Amazon, Microsoft, and later Apple too. For AI researchers today, the data used to train AI is precious and scarce, and small companies joining forces can close the gap with the big ones.

Nothing suits sharing data better than an open, nonprofit alliance.

They may genuinely fear that AI will destroy the world one day. But as usual, the side these business leaders take does not run against their companies' interests either.

Cover image: Westworld, stocksnap