DeepMind Co-founder: Generative AI is Just a Transition, Future AI Will Gain Freedom, Interactive AI Will Transform Humanity
-
In a recent interview with MIT Technology Review, DeepMind co-founder and Inflection AI founder Mustafa Suleyman shared his insights. He claimed they have successfully removed toxicity from AI and asserted that generative AI is only a transitional phase, with future AI expected to achieve human-like freedom.
Suleyman, who participated in an online interview with MIT Technology Review, proposed that the current era of generative AI is just a technological stage. The next phase will usher in the age of interactive AI, where AI will evolve into robots capable of completing tasks by coordinating other software and human resources based on individual user needs.
The mission of his company, Inflection AI, extends beyond creating a chatbot. It aims to leverage AI to provide more equitable access to resources and information for the vast majority of people worldwide.
Finally, he called for proactive government regulation of AI, expressing confidence that regulating AI under existing frameworks is entirely feasible.
While many may share Suleyman's views, what sets him apart is his role as the leader of Inflection AI, a company valued at billions of dollars.
This company has top talent from DeepMind, Meta, and OpenAI, and through its collaboration with NVIDIA, it possesses the world's largest reserve of professional AI hardware computing power.
We are currently in the second wave of AI technology.
According to Sulyman, the first wave of AI was about classification. Deep learning demonstrated that computers could be trained to classify various types of input data: images, videos, audio, and language.
Now, we are in the second wave—generative AI—where input data is used to generate new data.
The third wave will be interactive AI. He has long believed that conversation is the interface of the future. Users will not just click buttons or type text; they will engage in direct dialogue with AI.
These AIs will be capable of taking autonomous actions. Simply give them a general, ultimate goal, and they will use all available tools to achieve it.
They will converse with people and other AIs. This is what he aims to achieve with Pi.
This represents a massive upgrade in technological capability. It is a profoundly significant moment in technological history, and he believes many people underestimate its potential.
AI technology is still static at present. It essentially operates based on the instructions given to it.
However, technology will become more dynamic. If users grant AI sufficient freedom, it can utilize that freedom to take actions.
This truly represents a qualitative leap in human history, as we are creating tools with such agentic capabilities.
Humans will always remain the dominant force. It is crucial for humans to define the boundaries that AI must not cross and ensure these boundaries—from the actual code to interactions with other AIs or humans, and even the motivations and incentives of the companies developing the technology—provide provable safety.
Early Experiences and Their Impact on Values and Entrepreneurship
When asked how his unconventional career path, which differs from the traditional 'Silicon Valley style,' influenced his entrepreneurial journey in Silicon Valley, Suleyma responded:
He has always been interested in power, politics, and related fields. He believes that fundamental human principles are the result of compromise, relying on a dynamic equilibrium achieved through continuous negotiation among different, often conflicting forces.
Although humans constantly solve problems, they remain constrained by their own biases and cognitive blind spots.
Civil organizations, local governments, central governments, international organizations—all are limited in their ability to operate efficiently and flawlessly due to their inherent constraints.
Imagine what would happen if humans never made mistakes.
He believes it's possible for humans to create AI that truly reflects their own interests, ultimately making better trade-offs more consistently and fairly on behalf of everyone.
(Born in 1984 in North London to a Syrian father and British mother, Suleyman grew up in poverty. At 16, his parents separated and both moved abroad, leaving him and his younger brother to fend for themselves.
Suleyman attended Queen Elizabeth's School, a boys' grammar school in Barnet. Around that time, through his best friend, he met Demis Hassabis, co-founder of DeepMind. Suleyman says they often discussed how they would change the world together.
He was later admitted to Oxford University to study Philosophy and Theology but dropped out after a year.
He then co-founded the Muslim Youth Helpline, a telephone counseling service, with his university friend Mohammed Mamdani. The organization later became the UK's largest Muslim mental health support service.
Subsequently, Suleyman served as a human rights policy officer under London Mayor Ken Livingstone before founding Reos Partners, a 'systemic change' consulting firm that uses conflict resolution methods to address social issues.
As a negotiator and mediator, Suleyman has worked for clients including the United Nations, the Dutch government, and the World Wildlife Fund.)
Suleyman said that after leaving DeepMind, he has achieved financial freedom. He doesn't need to start a business or write books to make a living.
Money has never been the motivation for work. Money is a byproduct of work.
For him, the goal of work has never changed: it's about how to make the world a better place and how to move the world forward in a healthy and satisfying way.
Even back in 2009, when he first considered entering the tech industry, he saw AI technology as a fair and accurate way to provide services to the world.
(Suleyman co-founded the AI company DeepMind Technologies with Hasabiss, serving as Chief Product Officer. DeepMind was acquired by Google in 2014, and he became the head of applied AI at DeepMind.
In January 2022, Suleyman left Google to join Greylock Partners as a venture partner.
In March 2022, Suleyman co-founded the AI lab Inflection AI with Greylock's Reid Hoffman.
The company was established with the goal of using 'AI to help humans converse with computers,' recruiting former employees from companies like Google and Meta, and raising $225 million in its first funding round.)
In July of this year, Inflection AI received a $1.5 billion investment from Microsoft, Nvidia, Bill Gates, and LinkedIn founder Reid Hoffman.
After securing the funding, Inflection AI immediately purchased $1.3 billion worth of Nvidia H100 GPUs to build one of the world's largest GPU clusters.
Earlier this year, the company launched Pi, a competitor to ChatGPT. Additionally, Inflection AI's founder co-authored a book with writer and researcher Michael Bhaskar titled The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma.
Perspectives on the Negative Impacts of Technology
He believes that obsessing over whether one is an optimist or pessimist is not particularly meaningful, as it represents a biased way of viewing things. His focus is on calmly analyzing both the benefits and challenges.
From his perspective, it is clear that as large language models continue to scale, they become more controllable.
Two years ago, the prevailing view—which he considered incorrect even then—was that these models would only produce toxic, regurgitated, biased, or racist content. However, these issues were temporary. People overlooked the continuous progress of technology and failed to grasp its developmental trajectory.
Models like Pi are now highly controllable. Inflection AI ensures that Pi does not generate racist, homophobic, sexist, or any other form of toxic content. It can refuse to teach users how to create biological or chemical weapons or encourage them to vandalize their neighbors' windows.
As for how this is achieved, Suleyma prefers not to disclose too many technical details but encourages users to experiment freely. Pi is already live, and users can attempt all possible attack methods. Any jailbreaking, prompt hacking, or other attempts to bypass Pi's safeguards simply won't work—this is an objective fact.
Inflection AI boasts the world's strongest team, having created all the largest language models over the past three to four years. This is an exceptional team working in a highly efficient environment with abundant computational resources. From the outset, they prioritized safety, which is why Pi isn't as 'edgy' as models from other companies.
Take a look at Character.ai. (Character is a chatbot platform where users can design various 'characters' and share them online for others to chat with. It serves as a link.) Its primary function is to create imaginative characters, but Inflection AI has always believed this is not the right direction—they would not develop such features.
Pi will treat you with great respect. If you start complaining about immigrants in your community taking your jobs, Pi won't criticize you or lecture you.
Pi will empathetically inquire and offer you linguistic and psychological support. It will try to understand the source of such emotions and gently encourage you to empathize with others. These are values I've contemplated for 20 years.
When asked why they don't open-source the technology if it could make the world a better place,
Sulyman emphasized that, after all, he is running a company and needs to make money to support everyone. Having previously purchased $1.3 billion worth of NVIDIA GPUs, he must generate returns to meet investors' expectations.
Currently, the open-source ecosystem is thriving and performing quite well, with the open-source community discovering similar methods. Suleyman assumes they are at most six months ahead.
How can AI regulation be effectively implemented?
For him, the technology of AI recursive self-improvement needs to be regulated. It is dangerous for small-scale AI to update its own code without supervision—perhaps this should even be a licensed activity.
Perhaps there should be specialized mechanisms to regulate this behavior, similar to how anthrax or nuclear materials are handled.
He also believes that fear of AI is completely unnecessary. AI regulation can fully be addressed within the current regulatory framework. AI is just another component of the governance system.
Humans have excelled at regulating highly complex systems. Take the Federal Aviation Administration as an example: we have all grown accustomed to flying in aluminum tubes at 40,000 feet, which is one of the safest modes of transportation in history. This is an example of successful regulation.
Or consider automobiles: every component undergoes extreme stress testing, and users must have a driver's license to operate them.
Some industries—like airlines—have done an excellent job of self-regulation from the start. They understand that if they fail to ensure safety, people will become fearful, and they will lose business.