Shulou(Shulou.com)11/24 Report--

The original title: "OpenAI has just raised 10 billion yuan, DeepMind is in a hurry?" CEO accepted an exclusive interview with time, calling on AI circles to reduce scientific research competitions! "

Recently, time magazine invited Deepmind CEO and founder Demis Hassabis to express his views and concerns about the current development of AI, but netizens don't seem to buy it.

DeepMind has always been the pride of Google.

As a subsidiary of Google's parent company Alphabet, DeepMind is one of the world's leading artificial intelligence laboratories. In the 13 years since its establishment, it has handed over a very bright report card.

I have to say, DeepMind has something to do with the algorithm.

In July 2022, AlphaFold exploded. In October 2022, AlphaTensor, developed by DeepMind, appeared on the cover of Nature, the first AI system to discover novel, efficient, and correct algorithms for basic computing tasks such as matrix multiplication.

Now, everyone is aware that AI is on the cusp of the times.

Although ChatGPT's writing ability is comparable to that of human beings, it is difficult to avoid misinformation and toxic biases in training data. To make matters worse, computing power has become cheaper and cheaper, and the threshold of AI for the general public is getting lower and lower.

DeepMind CEO Demis Hassabis is worried about the development trend of AI. He is worried that AI will be used to make terrible tools and destroy human civilization.

However, some netizens complained, "I appreciate his position, but how can these AI development companies not act quickly when competitors are trying to move towards the same goal as quickly as possible?" "

Some netizens said that he was trying to buy time hhhhhh to catch up with OpenAI.

DeepMind's mission-- to build a general artificial intelligence AGI, let's take a look at what important information Hassabis reveals in this conversation.

Time magazine columnist BILLY PERRIGO interviewed Deepmind CEO and founder Demis Hassabis to discuss DeepMind's mission and how to build AI to solve some of the world's thorniest problems.

Standing in the middle of the spiral staircase, Demis Hassabis examines his cathedral, the three-story DNA sculpture that is at the heart of DeepMind's recently opened London headquarters

The 46-year-old Hassabis has been working on the eternal question about AI:

What is AI? What are the possibilities buried in it? How can human beings gain more power from it?

Hassabis says his team has been trying to achieve a more ambitious goal:

The goal is to build a machine that can think, learn and solve the most difficult problems of human beings to achieve general artificial intelligence (AGI).

Today's artificial intelligence is narrow, fragile, and not intelligent at all.

Hassabis believes that AGI will become an epoch-making technology that will change the structure of human life just like the use of electricity.

Even he himself changed the signature of the Twitter home page to "committed to AGI, trying to understand the basic nature of reality."

But the prospect of artificial intelligence is also accompanied by risks.

The AI program built by researchers to design new drugs can easily produce deadly biochemical products, and large-scale training AI models may be used by hate talkers.

ChatGPT, a popular model developed by OpenAI, has been criticized for its vulnerability to racism and misinformation, as has small company Prisma Labs, where many users complain that Lensa sexualizes their images and exposes biases in its training data.

It was in this atmosphere of uncertainty that Lensa, the AI enhanced selfie software developed by Prisma Labs, Hassabis agreed to give a stern warning about his growing concerns. "I suggest slowing down the pace of AI, at least not too fast."

Just like when the number of Facebook users soared to 3 billion, everyone was cheering, but the company was caught off guard when false information, hate speech and even incitement to genocide began to appear on the platform.

Hassabis sees a worrying trend in the development of artificial intelligence, he said:

Artificial intelligence is now at the forefront of creating tools that could wreak havoc on human civilization and urging its competitors to be more cautious than before, but most people are not aware of the dangers.

From the chess player to the upsetting CEO, unlike his peers, Hassabis12 was already the world's second-largest chess player at the age of age and graduated from high school a year early.

Hassabis (left) was captain of England's under-11 chess team at the age of 9.

When he was still in junior high school, he began to think about the big issues that changed human history. He said:

I want to understand the big questions, the big topics that are usually related to the field of philosophy or physics, and I think building AI will be the quickest way to answer some of these questions.

Hassabis was only 15 years old when he entered Guildford's Bullfrog video game studio.

He also competed in a video game magazine to win an internship in a famous studio, and a space invader-style game he designed was the second-best seller of the quarter.

Peter Molyneux, co-founder of Bullfrog, still remembers meeting Hassabis for the first time. Molineucks said:

He looked like an elf in the Lord of the Rings, a child who walked past him and would not even notice, but his eyes sparkled with wisdom.

In a casual conversation on the bus to Bullfrog's Christmas party, the teenager was fascinated by Molyneux.

They talked about the philosophy of the game, what makes victory so attractive, and whether you can instill these same characteristics into the machine, and then he knew that the young man was destined to do great things.

After graduating from Cambridge University while studying for a PhD, Hassabis returned to Bullfrog to help Molyneux develop his most popular game to date: Theme Park, a simulation game in which players run the amusement park business from God's point of view.

Before deciding to pursue a doctorate in neuroscience, Hassabis continued to build his own gaming company.

He wants to understand the algorithmic aspects of the brain: not the interactions between microscopic neurons, but larger structures that seem to produce powerful human intelligence.

Thought is the most elusive thing in the universe, from chess to games to neuroscience, my life experience and career trajectory are to discover the mystery.

In 2013, the third year of DeepMind, Google knocked on the door.

A group of Google executives flew to London on a private jet, and Hassabis amazed them by showing them the team's work-the AI prototype that trained to play the computer game Breakout.

Reinforcement learning (Reinforced Learning), the iconic technology behind the DeepMind algorithm, was something Google didn't do at the time.

It is inspired by the way the human brain learns, an understanding formed by Hassabis during his time as a neuroscientist.

AI is rewarded for every score he has trained millions of times in the game and learns the best strategy through a point-based reinforcement process.

Hassabis and his colleagues firmly believe in training artificial intelligence in a game environment, and Google executives are impressed by the benefits of this approach.

"I fell in love with them right away," said Alan Eustace, a former senior vice president of Google who led the expedition.

Currently, one of Hassabis's favorite games is called Polytopia. In this strategy game, the player's goal is to develop a small village into an empire that rules the world through technological progress.

Musk is also a fan of the game, Hassabis said. The last time the two talked was a few months ago, mainly about Polytopia.

While Hassabis is cautious, it's easy to understand why games resonate with him. He still believes that technological progress is essentially beneficial to mankind. Under the capitalist system, it is possible to predict and mitigate the risks of AI. "the progress of science and technology is the driving force of civilization," he said.

Abandoned by Musk, but as DeepMind continued to climb new heights, things began to get complicated.

In 2015, its first two investors, billionaires Peter Thiel and Elon Musk, symbolically turned their back on DeepMind by funding rival startup OpenAI.

The lab, which then received a $1 billion grant from Microsoft, also believes in the possibility of AGI, but has a very different idea of how to achieve that goal.

It's not that interested in games. Most of its research is not focused on reinforcement learning, but on unsupervised learning, which is a different technology that involves grabbing large amounts of data from the Internet and extracting it through neural networks.

While DeepMind, Google and other AI labs have been conducting similar research behind closed doors, OpenAI prefers to let the public use its tools.

In 2022, DeepMind launched DALL E 2, which can generate images of almost any imaginable search term, as well as the chat robot ChatGPT. Since both tools are trained based on data collected from the Internet, they are plagued by structural biases and inaccuracies.

DALL E 2 is likely to portray lawyers as old white men and flight attendants as young beauties, while ChatGPT is apt to assert false information confidently.

But despite Hassabis's call to slow down the pace of the AI competition, DeepMind does not seem to be immune from competitive pressure.

In early 2022, the company released a blueprint for a faster engine. The study, called Chinchilla, shows that many of the most cutting-edge models in the industry are inefficient in training and explains how they can provide more functionality with the same level of computing power.

Hassabis said DeepMind's internal ethics committee discussed whether it was unethical to release research because it could allow less cautious companies to release more powerful technologies without strict protection.

The moral red line of the artificial intelligence giant is in the middle, and Hassabis is well aware of the dangers that artificial intelligence can pose.

"he understands that AI technology has a profound impact on society, and he wants to know something about it before it is invented," Eustace said. "it's like chess. How to face the end of the situation? How to predict the development after 20 steps? "

Eustace promised Hassabis that Google shares the same concerns and that the two share the same interests. Eustace says Google's mission is to index all human knowledge, make it easy to access, and ultimately improve the world's IQ.

Alan Eustace, a former senior vice president of Google, "I think it resonates with him. In 2014, Google bought DeepMind for about $500 million. Hassabis rejected a higher offer from Facebook.

One reason, he says, is that, unlike Meta, Google is "very happy" to accept the moral red line drawn by DeepMind.

Hassabis hopes that DeepMind will become a representative of safety and ethics in Al research. Lead by example and be yourself in this field full of pursuit of faster and stronger.

DeepMind has issued a "red line" against the unethical use of its technology, including surveillance and weaponization. But it failed to stop its parent company, Google, from using its own AI system to cooperate with the Pentagon in technology.

In 2021, Alphabet ended years of negotiations with DeeMind. According to the outcome of the negotiations, DeepMind will maintain its current position at Alphabet, and its future work will be overseen by Google's Advanced Technical Review Board, including two DeepMind executives.

Hassabis does not deny DeepMind's plan to "run away from Google", but does not express its concerns about the company's current structure.

But he added that DeepMind's moral structure had "evolved into what it is now" since the acquisition.

Hassabis said that both DeepMind and Alphabet are committed to the public ethics framework and have incorporated security into their tools from the start.

DeepMind has its own internal ethics committee, the Institutional Review Committee (IRC), which is made up of representatives from all departments of the company and chaired by Lila Ibrahim, chief operating officer.

Lila IbrahimIbrahim, chief operating officer of DeepMind, said IRC meets regularly and any differences will escalate to a final decision by DeepMind's executive leaders. "We operate very freely and have a separate review process," she said. "

Asked what would happen if DeepMind's leadership team disagreed with the decision of Alphabet executives, or if the company's moral "red line" was crossed, Ibrahim said: "We haven't encountered this problem yet. "

AlphaGo: a man-machine war that shocked the world

In the glass cabinet on the far wall of the lobby of DeepMind headquarters, there is a large square piece of wood painted with black graffiti.

This is a souvenir of DeepMind's first major success. Shortly after Google's acquisition, the company set a challenge for itself to design an algorithm to beat the world's best players in go.

At that time, chess had long been conquered by brute force computer programming, but go was much more complex; the best artificial intelligence algorithms were still not comparable to top human players.

DeepMind solved this problem by cracking Breakout. It has established a program that can play millions of simulated battles with itself after learning the rules of the game by watching human games.

Through reinforcement learning, the algorithm updates itself, reducing the "weights" of decisions that make it more likely to lose the game, and increasing the "weights" that make it more likely to win.

In March 2016, an algorithm called AlphaGo faced Lee se-dol, one of the world's top go players. AlphaGo beat him four to one. The defeated Li scribbled his signature on the back of the chessboard with a black marker.

Hassabis celebrates DeepMind's victory over go player Lee se-dol (right) in South Korea in 2016. Will the AGI-driven future be utopia? Hassabis believes that if the era of wealth creation by AGI has come, then that wealth should be redistributed. "I think we need to make sure that as many people as possible benefit-preferably all mankind. "

He likes the idea of a national basic income (universal basic income). Under the concept, every citizen will receive subsidies and basic services from the government every month. At the same time, the state will ensure citizens' basic living standards such as transportation or housing.

Hassabis said the AGI-driven future should be more focused on wealth equality than the world today, but he did not explain how the system would work.

"if we live in a world of extreme affluence, the less room for inequality and the lower the probability of inequality. So if AGI's vision is realized, this is one of its positive effects. "

However, given that the rapid growth of the technology industry has coincided with a sharp rise in wealth inequality over the past few decades, others are less optimistic about this utopian future.

Paris Marx, host of the podcast Tech Won't Save Us, said: "the top priority for large companies, including the parent company of DeepMind, is to ensure maximum value for shareholders."

If they don't make a profit, these companies won't focus on solving the climate crisis at all. When the whole goal of the company is to accumulate more wealth for shareholders, they are certainly not interested in redistributing wealth. "

"what we lack is not the technology to solve the climate crisis or redistribute wealth," Marx said. "what we lack is political will. At present, it is difficult to see how the birth of a new technology will generate political will to achieve real social transformation. "

Back to the spiral staircase of DeepMind, in the middle of the staircase, there is a huge DNA sculpture. There is a message on the wooden stool in front of the sculpture: "Please don't touch it," it says. "it is very fragile and easily damaged. "

Reference:

Https://time.com/6246119/demis-hassabis-deepmind-interview/

Https://www.youtube.com/watch?v=Gfr50f6ZBvo

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.