Nine orders of magnitude later, he shipped a companion, not an oracle

Mustafa Suleyman tells Sarah Guo and Elad Gil that he was a change-the-world kid: grand visions, doing good, huge impact. Philosophy at Oxford filled him with hypothetical moral quandaries, so at the end of his second year he walked out of theory and into a telephone. A friend was starting a helpline for young British Muslims — non-judgmental, non-directional, secular — about six months after 9/11, as Islamophobia rose and the government talked anti-terrorism. He was 19. He spent almost three years on it full time, with nearly a hundred young volunteers answering the phones. Fundraising was the game, only the numbers were much smaller. He still calls it his first startup.

One-to-one nonprofit work would not scale, so he founded a consultancy that treated conflict as a process problem. Clients included the United Nations and the Dutch government. The work ran through Israel-Palestine, Cyprus between Greeks and Turks, and, via colleagues, South Africa, Colombia, and Guatemala. The skill he came away with is unfashionable and, he says, trainable: learn to speak other people’s social languages. With attention, patience, and care you can make someone feel heard so that they can empathize with people they disagree with. That, he says, is a superpower, and it has been a theme of his whole career.

The limit of large-scale human process showed up in 2009, when he facilitated one part of the climate negotiations in Copenhagen. One hundred ninety-two countries, a thousand NGOs and activists, academics proposing different definitions of the problem: inspiring as a gathering, deeply depressing as a result. Mild consensus did not arrive until 2015. He left thinking the world’s governance systems would not keep up with exponential challenges from carbon or from technology. That was the next thing on the landscape, and it is why a mediator ended up in an AI lab.

Around 2008 he was watching Facebook, then two or three years old, hit 100 million monthly actives. The number was mind-blowing. Coming from conflict resolution, he did not see a neutral pipe for information. He saw a frame: how you organize space, how you prepare people for disagreement, which button you put where. A thumbs-up, the order of controls, the arrangement of a page — those design choices would incentivize behavior at a scale he compared to religion. For thousands of years a male God and a fixed role for women had been the implied undertone of a social structure. Facebook, he thought, was rewriting the default of human connection in a few years.

Gil remembers building Facebook apps in that era and thinking about conversion, not society. Suleyman had already decided Copenhagen was not his path to significant change — the processes should continue, just not with him full time — and his heart was set on technology. He reached out to Demis Hassabis, the brother of his childhood best friend. They loved games, they were competitive, they were good at poker, and one night in a London casino they both got knocked out early. They sat drinking Diet Coke and talking about how to change the world. Suleyman was inspired by platforms and social apps. Hassabis was in robotics and sci-fi: simulate the entire economy to make economic decisions, a frame leftover from simulation games such as *Evil Genius* and *Black & White*.

The third person, Shane Legg, was the one who made the language of artificial general intelligence feel like an engineering project. For his Ph.D. with Marcus Hutter he had gathered something like 80 definitions of intelligence and collapsed them into one: the ability to perform well across a wide range of problems. Suleyman calls that the turning point. It gave the three of them a measurement — a ruler they could hold up to an algorithm to ask whether it was inherently general, whether it could do many things well at the same time. DeepMind was founded in 2010 on that thesis: distill the essence of human intelligence into an algorithmic construct.

Asked whether he still uses Legg’s definition, Suleyman says no. It is a good definition that over-rotated the whole field onto one aspect of intelligence: generality. OpenAI, Anthropic, and others, in his telling, took up the mantra that what matters is whether a single agent can do everything — multimodal, translation, speech, the lot. He offers another: the ability to direct attention, or processing power, to the salient features of an environment given some context. You want to take raw horsepower and point it the right way at the right time, because the useful move might be a tone, an expert model, or a tool.

The unlock, he thinks, is the router sitting in the middle of specialized systems, some of which will not look like AI at all — ordinary software, databases, tools. That central brain does not have to be the largest language model. Guo notes that this is closer to a brain: visual cortex for vision, mirror neurons for empathy, an ensemble with routing. Suleyman agrees that neural nets and reinforcement learning have always been brain-inspired, and that sparse activation is a promising research area. He also says you do not have to wait for sparsity. You can train a decision engine that knows when to call a 20-second, high-quality, expensive model and when a three-second mini model is enough.

That, remarkably, is more of an engineering problem than an AI problem — a surreal sentence given where the field is. The early company was, in his word, wacky. For most of 2010 he was sneaking into the Gatsby computational neuroscience unit at UCL with Hassabis and Legg to sit in on lunches Peter Dayan ran. Legg coached him on the local language: say machine learning, do not say AI, and certainly do not say AGI. Funders were scarce. Peter Thiel had real vision, Suleyman says, though he sold pretty early and now does not seem to be in the game. Elon Musk was in the small circle too. The weather changed with AlexNet in 2012, the DQN Atari paper in 2013, then AlphaGo. It still surprises him how suddenly the rest of the world woke up in the six months before this taping.

The breakthroughs he is proudest of start with deep reinforcement learning: learn from raw perception, feed a reward, iterate online or in batch simulation. It turned out to be valuable for a specific kind of problem — a game-like environment, a structured scalar reward, millions of plays. That is why his group started AlphaFold. In a week-long hackathon, around 2016, they were hunting for other applications of DQN-like, AlphaGo-like tools; someone stumbled onto protein folding, a problem they had glanced at in 2013 under the name Foldit. They hired outside consultants for domain knowledge, entered the CASP competition the following year, and eventually turned it into a company-wide strike team. These things take a long time, he says, longer than people realize.

Guo used to work as a biologist. She describes years of crystallizing proteins in different salt concentrations, hitting them with X-rays, and interpreting the diffraction just to see a structure. DeepMind’s system ran through protein sequences in every database for every organism and predicted the fold. Suleyman’s one-line thesis for DeepMind is transfer: learn a rewarding behavior in one environment and move a more compressed representation of what worked into the next. That is why he still wants AI. The urgent problems — food, water, decarbonization, clean energy, transport, a rising population — do not have a genius insight sitting on a shelf. The smartest humans are near the limit of what they can invent. Keep the north star in view, he says, rather than getting sidetracked by the fears.

Language was later, and less obvious. In 2018 and 2019 it was not clear neural nets would matter there. CNNs had spent five years learning local structure: nearby pixels became edges, eyes, faces, scenes, and the same hierarchy seemed to fit audio. Next-word prediction looked different. The structure that matters is often far from the token you are about to emit. GPT-2 did not click for him. The GPT-3 paper did. Early in 2020 he went to Google, joined the team that became LaMDA, and with Noam Shazeer and others scaled small models, tried retrieval grounding for factuality, and lived with hallucinations. In 2021 he, David Luan (later of Adept), and Noam banged on the table to launch. Timing is everything; it was not the right time for Google. In January he left, brought on Karén Simonyan — whose company DeepMind had acquired in 2014, and who had led the deep-learning scaling team — and Reid Hoffman, a close friend of ten years they had always talked about starting something with. Inflection followed.

Why companionship rather than information? The missing ingredient on LaMDA, and the conclusion Simonyan had reached after beating academic benchmarks, was interaction feedback. Every interface is already a dialogue: you are constantly telling a browser or an app what you think. Suleyman’s pitch inside Google was that conversation is the future interface, and that Google is already a conversation, just an appallingly painful one. You say something; it answers with ten blue links; you reply by clicking, dwelling, scrolling; you come back and rewrite the query. Google learns that way. The problem is that it is having the conversation with a 1980s Yellow Pages.

Worse, Google accidentally reshaped the internet around ads. Pages are SEO’d to within an inch of their lives: subheads, bullets, ads between the paragraphs, five to ten seconds of scrolling for the snippet you wanted. Eleven seconds on the page looks like quality to the ranker, so creators are paid to trap you. Humans, he says, want a succinct, fluent, natural-language answer, and then they want to update it without translating themselves into keywords. "We've learned to speak Google." It is a weird lexicon co-developed over twenty years. That moment is over. We can talk to computers in fluent language, and that is the new interface.

Pi, for personal intelligence, is the first version of the AI he thinks everyone will have in a few years. There will be business AIs, government AIs, nonprofit AIs, political AIs, influencer AIs, brand AIs, each aligned to its owner’s objective — promote, sell, persuade. Individuals will want one aligned to them, on their team, in their corner. Inflection started with an empathetic, supportive style and tried to reverse-engineer great conversation: the other person listens and reflects; they add a fact rather than regurgitating; they ask follow-ups out of curiosity; sometimes they throw in something silly, surprising, or slightly wrong, and it is endearing. This is not even Inflection’s biggest model. It is a first version skinned for that interaction so the company can learn.

If computing becomes a conversation, the web’s structure changes. Your Pi summarizes the morning news in your reading style, and every couple of days it drops a snippet on cactuses or motorcycles. A website assumes one format for everybody; generative models make the format dynamic. If he were Google he would be worried: the old system does not look like 10-years-from-now. Publishers who live on a URL should think of an AI as a website or an app. A baking blog can still produce high-quality content; the AI that represents it will be more engaging to talk to. Any brand, he says, is already an AI using static tools. Advertising has spent a couple of hundred years making meaning with color, shape, texture, text, and sound, and shipping a new version every six months. The next version is interactive.

He does not buy the story that one or five AIs remain. That view is "completely misguided and fundamentally wrong." There will be hundreds of millions, or billions, aligned to individuals. What he does not want is autonomous AIs that wander off and operate independently. That vision does not end well. The picture he does want: his Pi goes to the baker’s AI and says Mustafa wants to learn baking and cannot crack an egg; the two AIs have a funny interaction; Pi comes home with a three-minute clip. New content gets produced that way. Your personal AI is the interlocutor, which is, he notes, basically what Google already does when it crawls static pages, has a little interaction, ranks them, and presents the result.

Guo’s objection is the filter bubble. Social feeds were criticized for showing you what you already believe; a personal agent could be the extreme form, using your feedback to reinforce whatever you walked in with. Suleyman does not dodge it. That is the default trajectory without intervention. "The platforms were never neutral — that was the big lie." Ranking is choice architecture. Showing and hiding is political and cultural power, and AI is an accelerated version. His prescription is that AI companies and the old platforms have to own curation, be transparent about what is excluded, and get more open to the European approach. Society has to decide which bodies it trusts to oversee recommendation algorithms. Democratic structures need to sort themselves out, build functioning oversight, and stop fainting at the word censorship. Guo says that sounds like a harder hill than AGI. He agrees. He hopes not.

Improvement, for Pi, is measured as emotional intelligence, conversational fluidity, respect, and even-handedness. The model has already made politically biased remarks; the bar is that it must never be dismissive, disrespectful, or judgmental, no matter how ugly the user’s line, because its job is to talk issues through and make you feel heard. The next phase is the personal digital assistant: talk to other AIs, buy groceries, book vacations, look up weather, news, sports, citations. Memory is only about a hundred messages, which is still a lot, and the aim is a second mind that stores preferences, habits, and relationships. People already dump huge amounts of personal information into it. They also send notes to the developers: I love voice P2; also, I wrote those Forbes stories, not that other one. One of those notes was a journalist, the day before.

Inflection is about 30 people. Everyone on the technical side is an MTS. The company refuses a caste system of researchers, scientists, engineers, and data scientists; they have seen that hierarchy go wrong at other labs. It is an applied AI company that does not publish, even though it runs some of the largest language models in the world and is state of the art on many main benchmarks — not coding, because Pi does not write code and that is not a priority. There are no managers yet, only Suleyman and Simonyan. The best people, in his experience, want three things: high-quality colleagues, outstanding resources and freedom, and a shared goal. They ship every six weeks. In the seventh week the whole company meets in person for a hackathon, drinks, dinner, a week of intensity, then plans the next six. Twelve-week OKR cycles are too long, he says, and most people make up the second half anyway.

Progress, in his view, is compounding exponentials. Compute used to train the largest models has grown by an order of magnitude every year for a decade. He went back to the 2013 DQN Atari paper: two petaflops. Inflection’s biggest models: ten billion petaflops. Nine orders of magnitude in nine years. Stay humble, and when you predict, look at the exponential. Data is the familiar story. Efficiency is the one people underweight. Last year’s Chinchilla paper showed that smaller models trained longer on more data can be compute-optimal and match the old giants — evidence that architectures are early and under-optimized. Jordan Hoffmann, Chinchilla’s lead author, is on the Inflection team. They will push scale, including one of the world’s largest supercomputers, and more efficient architectures at the same time. He calls it a coming wave of contradictions. He is also finishing a book, *The Coming Wave*, on what the AI and synthetic-biology revolutions do to the nation-state over the next decade. The method is the same as the six-week clock: hard deadlines sharpen thoughts. He has regretted the Penguin Random House deal many times while running a startup, and torn at already gray hair. The book is nearly done.