Generative AI is only a transition, AI will be free in the future, and interactive AI will change mankind.

Editor: Run

Mustafa Suleyman, co-founder of DeepMind and founder of Influence AI, accepted an online interview with him by MIT Technology Review. In the interview, he threw out his views:

At present, the generative AI is only a technical stage, and then it will enter the era of interactive AI: AI will become a robot that can call other software and people to complete the work according to the different task requirements of each user.

The goal of the company he founded, Influence AI, is not only a chat assistant, but also to make the vast majority of people in the world get more equal resources and information through AI.

Finally, he also called on the government to actively supervise AI, and believed that it was not difficult to supervise AI under the existing system.

There may not be many people who have similar views with Suleyman, but unlike most people, he is the leader of Influence AI, a company with a valuation of billions of dollars.

This company has top talents from DeepMind, Meta and OpenAI, and has the largest professional AI hardware computing power reserve in the world through cooperation with NVIDIA.

We are in the second wave of three waves of AI technology.

In Sulyman’s view, the first wave of AI was about Classification. Deep learning shows that humans can train computers to classify various types of input data: images, videos, audio and languages.

Now in the second wave, generative AI, input data and generate new data.

The third wave will be interactive AI. He always thinks that dialogue is the future interactive interface. Users not only click buttons and type words, but also have a direct dialogue with AI.

These AIs will be able to take action on their own. Just give it a general and final goal, and AI will use all tools to achieve this goal.

They will talk to others and talk to other AI. This is what he will achieve with Pi.

This is a huge upgrade of technical ability. This is a very, very far-reaching moment in the history of technology. He thinks that many people underestimate the potential of technology.

AI technology is still static. It basically does things the way you tell it.

But technology will become alive. If the user gives AI full freedom, AI can use this freedom to take action.

This is indeed a qualitative leap in human history, and we are creating tools with this agent ability.

Humans will always be the dominant players. Humans need to make sure that AI can’t cross the boundaries, and ensure that these boundaries can provide provable security from the actual code to the interaction with other AI or humans, to the motivation and incentives of the companies that create technology.

Influence of Early Experiences on Their Values and Entrepreneurship

When asked about his unique professional experience, which is completely different from the traditional "Silicon Valley Style", and what impact these experiences have on his business in Silicon Valley, Suleyma replied:

He has always been interested in power and politics. He thinks that the basic principles of human beings are the product of compromise. It depends on a dynamic balance that can be achieved through continuous consultations between different and conflicting forces.

Although human beings are constantly solving problems-but human beings will still be limited by their own prejudices and cognitive blind spots.

Non-governmental organizations, local governments, central governments, international organizations, etc.-will all be limited by themselves and cannot operate efficiently and without mistakes.

Imagine what would happen if human beings didn’t make mistakes.

He feels that it is possible for human beings to build an AI that truly reflects their own interests, and ultimately make better trade-offs on behalf of all people more consistently and fairly.

(In 1984, Sulyman was born in North London to a Syrian father and an English mother. He grew up in poverty. At the age of 16, his parents separated and both of them moved abroad, leaving him and his brother to fend for themselves.

Sulyman went to Queen Elizabeth School (Barnet’s grammar school for boys). Around that time, he met Demis Hassabis, the co-founder of DeepMind, through his best friend. Sulyman said that he and Demis Hassabis often discuss how they will influence the world.

Later, he was admitted to Oxford University to study philosophy and theology, but dropped out a year later.

Later, together with his college friend Mohammed Mamdani, he founded the Muslim Youth Helpline, which is a telephone consultation service. The organization later became the largest Muslim mental health support service in Britain.

Subsequently, Sulyman worked as a human rights policy officer under Ken Livingstone, the mayor of London, and later founded Reos Partners, a consulting firm of "systematic change", which used conflict resolution to solve social problems.

As a negotiator and mediator, Suleyman has worked for a series of clients such as the United Nations, the Dutch government and the World Wide Fund for Nature. )

Suleyma said that after leaving DeepMind, he was financially free. You don’t need to start a business or write a book to make money to support yourself.

Money has never been the motivation for work. Money is a by-product of work.

For him, the goal of his work has never changed, that is, how to make the world better and how to push the world forward in a healthy and satisfactory way.

Even in 2009, when he began to consider entering the technology industry, he saw that AI technology represented a fair and accurate way to provide services in the world.

(Suleyman and Hasabiss co-founded DeepMind Technologies, an artificial intelligence company, as the chief product officer. DeepMind was acquired by Google in 2014, and he became the head of DeepMind’s application of artificial intelligence.

In January 2022, Suleyman left Google and joined Greylock Partners as a venture partner.

In March 2022, Suleyman and Reid Hoffman of Greylock co-founded the AI laboratory Influence AI.

The company was founded with the goal of "using artificial intelligence to help people talk to computers", recruited former employees from companies such as Google and Meta, and raised $225 million in the first round of financing.

Then in July this year, it received $1.5 billion investment from Microsoft, NVIDIA, Bill Gates and LinkedIn founder Reid Hoffman.

Immediately after Influence AI got the financing, it purchased NVIDIA H100 GPU worth $1.3 billion and built one of the largest GPU clusters in the world.

Earlier this year, he released a ChatGPT competition called Pi. He also co-authored a book about the future of AI with writer and researcher Michael Bhaskar, entitled The Coming Wave: Technology, Power and the Biggest Dilemma in the 21st Century. )

How to treat the negative effects of technology

He thinks it doesn’t make much sense to dwell too much on who is optimistic or pessimistic. This is a biased way of looking at things. He just wants to calmly analyze the interests and challenges.

From his standpoint, it is very clear that with the further increase of the scale of large language models, they become more controllable.

So two years ago, the view at that time-he thought it was wrong at that time-was "Oh, they will only produce toxic, ruminant, biased and racist content. 」

But these situations are only temporary. People ignore the continuous progress of technology and don’t really understand the track of its development.

Now a model like Pi is very controllable. Influence AI can prevent Pi from producing racism, homosexual discrimination and sexism-any form of toxic content.

It can make it refuse to teach users to make biological or chemical weapons, or encourage users to smash neighbors’ windows.

As for how to realize it, Suleyma didn’t want to disclose too many technical details, but he said, please feel free to experiment boldly! Pi is online, and users can try all possible attacks.

Any jailbreak, prompt hacker attack or anything against Pi will not work. This is an objective fact.

Influence AI has the most powerful team in the world and has created all the largest language models in the past three or four years.

This is an excellent team, working in a very efficient environment and having a lot of computing resources.

They made safety their first priority from the beginning, so Pi was not as "hot" as other companies’ models.

Look at Character.ai (Character is a chat robot, users can design different "characters" for it and share them with others online. Its function is mainly used to create unconstrained characters, but Influence AI thought it was not a good direction from the beginning-they wouldn’t do such a function.

Pi will respect you very much. If you start complaining that immigrants in your community have robbed you of your job, Pi will not accuse you or tell you what to do.

Pi will ask with concern and provide you with language and psychological support. It will try to understand where users’ emotions come from and gently encourage you to sympathize with others. These are the values that I have been thinking about for 20 years.

But I was asked why not open source if these technologies can make the world a better place.

Sulyman stressed that after all, he was running a company and had to earn money to support everyone. Before he bought a 1.3 billion GPU from NVIDIA, he had to earn money to meet investors’ demands for returns.

Now the open source ecosystem is booming and doing quite well, and the open source community has also found similar methods. Suleyman only assumes that they are six months ahead at most.

How to effectively supervise AI?

For him, the technology of developing AI recursive self-improvement needs to be supervised. It is dangerous for a small AI to update its code without your supervision. Perhaps this should even be an act that requires permission.

It may be necessary to have a special mechanism to supervise this behavior, just like anthrax or nuclear materials.

He also believes that the fear of AI is completely unnecessary. AI supervision can completely settle within the current regulatory framework. AI is just another component of the governance system.

Humans have done a very good job in regulating ultra-complex things. Take the Federal Aviation Administration as an example: We are all used to flying at an altitude of 40,000 feet in these aluminum pipes, which is one of the safest modes of transportation in history. This is an example of successful supervision.

Or a car: every component has undergone extreme stress tests, and users must have a driver’s license to drive it.

Some industries, such as airlines, regulate themselves well from the beginning. Because they know that if they don’t do a good job in safety, everyone will be afraid and they will lose their business.

References:

https://www.technologyreview.com/2023/09/15/1079624/deepmind-inflection-generative-ai-whats-next-mustafa-suleyman/