Demis Hassabis

Demis Hassabis would not have every country build its own frontier AI model. He expects half a dozen offerings at or near the frontier, six to eight if Chinese companies are included, leapfrogging by maybe three to six months and staying relatively close. An enterprise can choose among them, because one model will be stronger at multimodal understanding and another at coding, and leading on every measure at once is very hard.

Open source sits in that set and maybe only six months to a year behind, which is why he says Western open-source options are needed. The open problem is how to restrict bad actors while still letting scientists and other good actors use the tools. A country or a compa

Demis Hassabis says society is not ready for conscious AI, so the work now should stay on intelligent tools rather than conscious beings. Self-volition or self-awareness works against corrigibility, a system that follows instructions and stops when a person tells it to stop. Training a model to reject commands from its own internal rules repeats the warning in Asimov's three laws. Today's systems are more malleable than researchers expected 15 or 20 years ago. That does not mean misalignment is solved on the way to superintelligence, and combining it with recursive self-improvement makes the risk worse.

There has not been a debate, beyond technologists, with philosophers, theologians, arti

Demis Hassabis says AlphaFold helps design drugs by predicting a protein's three-dimensional shape from its amino acid sequence, which shows where a drug should bind. He calls proteins the building blocks of life and the folding problem a 50-year grand challenge in biology, sometimes described to him as Fermat's last theorem for biology. That shape is what you need to understand a protein's function, what goes wrong in disease, and which part of the surface to target. Biology stays hard because of the interactions and the emergent behavior.

Folding is only one step in discovery, and he calls it a critical one. AlphaFold sped that step up by orders of magnitude. AlphaFold 2 was finished in 2

Demis Hassabis thinks AI needs third party auditors or evaluators, on the model of auditors in the finance industry, because the investment going into AI should fund a healthy ecosystem of startups and companies with enough revenue to do the work. Which third parties come in is still an open question. Leading labs generally agree on bringing auditors in, and he can see the ecosystem forming. Accenture is one of the big players moving in, and other leaders are working with the firm. He thinks Accenture is a little ahead of the game because the company acquired a good startup.

He thinks an industry-run standards body, with some kind of government oversight, is going to be needed as well. Idea

Demis Hassabis says the next phase of AI carries the most risk: closed RSI, a closed loop without the human reviewer process. The closed loop will obviously be faster, and it is also a very dangerous way to get to full AGI.

A lot of the cyber issues from the last six months could have been addressed with better engineering and more robust engineering. The human is still very much in the loop.

Leading labs generally agree on third party auditors or evaluators coming in. Which third parties should do the work is still an open question, and a healthy ecosystem is needed. Good startups and companies could take on the work. There is enough revenue for them, given the amount of investment going

Kevin Roose says Sundar Pichai does not say AGI. Demis Hassabis does, and Sergey Brin does. Roose does not think the labs feel it equally. Anthropic is the most AGI-pilled, OpenAI is a close second, and Google has always struggled. Ilya Sutskever used to walk floor to floor at OpenAI telling people to feel the AGI.

Inside Google, researchers fought it out on internal message boards. One wrote a memo called Say AGI: if Google wants to win in AI, recruit top technical talent, and be careful about how it develops AI, it has to admit, by using the term, that the work is more than tools that make Gmail, Docs, and Google Photos better. It is building toward human-level general intelligence. Someo

Sundar Pichai says Google changed who runs frontier models at DeepMind because that work has become an all-consuming job of its own. It is moving, it is picking up pace, and it runs from data to compute to the algorithms to the teams and the model agents.

Jeff Dean spent 27 years at Google. Sundar Pichai says he is proud of that work, Jeff Dean wanted to go build something he was excited about with a set of people, and Google is excited to see what he does with Discovery Loop.

Demis Hassabis has been thinking about this frontier moment for a long time. Sundar Pichai says Demis Hassabis wants more of his time on getting through it safely and on applying AI to other problems, the work behind

Kevin Roose says Google fell behind in the AI race because its leaders did not believe in artificial general intelligence the way Dario Amodei, Sam Altman, and Demis Hassabis do. Google had the resources, the talent, and the breakthroughs, and still lacked that belief. Halfway through his book, a story he expected to be about technology turned out to be about belief. For all the flaws, stumbles, and departures, OpenAI and Anthropic are the two most believing organizations, and that belief is why OpenAI is still in the lead or close to it.

Belief that the work could become artificial general intelligence is an essential ingredient. OpenAI still has very talented people. Roose has talked with

David Senra, reading Sebastian Mallaby, says DeepMind decided what stayed in the master code by measurement alone. The lab embraced a strict unity, and all team members poured their energies into improving one single model. Next came meritocracy. Any team member was welcome to propose an improvement to the model and test it. If the upgrade boosted performance, it was added to the master code on which everyone was building. Seniority, force of personality, and dazzling theoretical claims as to why something should work did not affect what went into the program. Only measurement mattered.

Senra says Demis Hassabis is running all of Google's AI, and the project is so big that Hassabis does not

David Senra, reading Sebastian Mallaby, says AlphaFold let a DeepMind team, numbering perhaps 20 at its peak, defeat hundreds of academic scientists who had spent decades on protein folding.

Demis Hassabis had been obsessed with solving protein folding since college. Senra says the work is what Hassabis is going to win the Nobel Prize for. Hassabis thought AI had progressed sufficiently to solve the problem for the first time in human history. Proteins are the building blocks of life. They provide the structure and the function of organs and muscles, hormones and hair, blood and brain cells. Predicting the complex shapes proteins assume was fascinating in its own right and almost certain to

David Senra, reading Sebastian Mallaby, says Demis Hassabis sold DeepMind to Google because he needed a backer with essentially unlimited resources, and blue-sky research was a poor fit for venture capital.

Fundraising nearly killed DeepMind. Backers promised an amount and then reneged, and the company almost ran out of money. Peter Thiel barely saw the DeepMind team and felt instinctively suspicious of a fellow chess player who had spent formative years mentally crushing opponents. Investor talks left Demis feeling his brain was atrophying. He was describing the biggest invention ever, and they kept asking where the widget was. He had struggled to persuade Founders Fund that DeepMind would

David Senra, reading Sebastian Mallaby, says DeepMind found its first believers by telling conference crowds it was starting an AGI company and talking only to the hardcore believers who stayed. Demis Hassabis stayed completely undeterred even though nobody believed in the future value of DeepMind stock.

Prestigious figures in the field assumed a research team with no revenues would do interesting science for a couple years and then go out of business. Scientific startups need blue sky thinkers who wander into the unknown. At conferences the team said it was starting an AGI company. 80% of the people rolled their eyes, turned around, and walked away. The team treated the walk-away as an eff

David Senra, reading Sebastian Mallaby, says Demis Hassabis needed money for DeepMind and hooked Peter Thiel with chess at a crowded conference party instead of pitching yet another startup story.

Hassabis knew Thiel was obsessed with chess, so the conversation started there. Thiel invited Hassabis and his co-founder to the house the next day to explain the ambitious venture. Thiel began to think DeepMind was an A-plus on the science and maybe an F on the business model. He also judged Hassabis an extreme case of an authentic entrepreneur: a missionary who feels compelled to work on a particular challenge and starts a company as a way of tackling it, rather than a mercenary who starts with

David Senra, reading Sebastian Mallaby, says Demis Hassabis discovered chess at four, had a natural aptitude for the game, and within a few weeks had mastered it well enough to defeat adults. Senra thinks he once heard Hassabis say that half his brain is dedicated to competition.

Hassabis was very special from a very young age. His mother had grown up in poverty, spending part of her childhood as an orphan on the streets of Singapore. His father had been the first from his family to attend university, an aspiring singer-songwriter who sold toys out of the family's beaten-up Volkswagen van. Hassabis first saw his father play a game against his uncle, and the game made him want to learn how t

David Senra, reading Sebastian Mallaby, says Demis Hassabis told Mallaby to read Ender's Game, one of Hassabis's favorite novels, in order to understand him. Hassabis identified with Ender and wanted the capacity for endurance, and the ability to suffer and still soldier on, visible before their first long conversation.

Partway through doctoral research in neuroscience, Hassabis had already been a chess master, a video game designer, an amateur theoretical physicist, an entrepreneur, a computer scientist, and a five-time world champion when the novel made sense of who he was. Ender is a diminutive boy genius taken from his family to an intergalactic battle school, manipulated by adults, bul

David Senra, reading Sebastian Mallaby, says founding a company to build AGI in 2010 was viewed as ridiculous and laughable. Fellow scientists rolled their eyes when DeepMind was founded in 2010, believing the construction of human-like AI to be impossible. Almost every potential investor turned them away. Demis Hassabis still scrapped together funding and persuaded gifted researchers to join him on the strength of an exhilarating vision: use AI to solve every single scientific problem that plagues humanity.

The optimistic vision of AI discovery has history in its corner. Past innovations from gunpowder to nuclear fission have made wars more terrifying and accidents more lethal. The general