Credit: © Stefanut Sava's Images via Canva.com
Ineffable Intelligence has handed the title of cofounder to six people who did not found it. Four come from Google DeepMind, one from InstaDeep, and one from the Seattle venture firm that wrote the company’s first cheque. The London startup is ten months old, has no product, and carries a $5 billion valuation.
Jeremy Kahn reported the hires for Fortune on Monday, in a story the magazine ran as an exclusive. Ineffable confirmed them and declined to comment further. All six had updated their LinkedIn profiles to show the new roles, which is how the appointments became visible at all.
The four from DeepMind are Chris Apps, Wojciech Czarnecki, Lasse Espeholt, and Junhyuk Oh. Alexandre Laterre arrives from InstaDeep, the London company BioNTech bought in 2023, where he ran research. Heather Gorham was a partner at Flying Fish, one of Ineffable’s earliest backers.
The AlphaStar band, back together
Apps, Czarnecki, and Oh all worked with David Silver on AlphaStar, the system that reached grandmaster level at StarCraft II in 2019. Espeholt built weather models on Google’s MetNet. Ineffable is not assembling a general-purpose research staff. It is rebuilding one specific team from one specific era of DeepMind, the era before the lab reorganised itself around shipping Gemini.
That is the thesis made into an org chart. Silver argues that large language models cannot reach superintelligence, because they only ever recombine what humans have already written. His answer is reinforcement learning: agents that learn by acting inside simulated worlds and discovering things no human wrote down. Ineffable raised $1.1bn on that argument in April, the largest seed round Europe has recorded.
The hire that says the most
Oh will run reinforcement learning at Ineffable, and his recent work points somewhere specific. In October 2025 he was joint first author on a Nature paper that Silver also signed. The team used machine learning to discover a new reinforcement learning rule, rather than designing one by hand. They called it DiscoRL, and it beat every existing rule on the Atari benchmark.
The paper’s closing line is the part that matters commercially. The authors suggest that agents may soon find the algorithms for advanced AI themselves, out of their own experience. Researchers would stop writing them by hand. Every frontier lab is now chasing some version of that idea, usually filed under recursive self-improvement. Ineffable has hired one of the few people with a published result in it.
TNW has covered the neighbouring bets. Silver’s former colleague Richard Sutton left Keen in July to start his own reinforcement learning lab. Recursive Superintelligence raised $650mn in May to build AI that improves itself. The thesis is no longer eccentric. It is crowded.
An investor who joined the company
Gorham’s move is the odd one, and the best documented. GeekWire’s John Cook told that story in April. She approached Silver in early 2024, after he published a paper arguing that knowledge drawn from human data was nearing its ceiling. Flying Fish wrote the first cheque roughly two years later.
Managing director Frank Chang told GeekWire the firm had spent years building relationships in London’s ex-DeepMind circle. Much of that came through its 2020 investment in the Seattle startup Phaidra. Gorham now handles compute, fundraising, and operations at Ineffable. She also helped secure the Google Cloud deal announced in July, according to her LinkedIn profile.
What DeepMind keeps losing
The hires land on a lab that has been leaking senior people all year. Beatrice Nolan reported for Fortune in August on the scale of the drift. DeepMind’s share of research and advanced-engineering hires across Europe, the Middle East and Africa fell from 49% in 2022 to 2023, down to 18.6% in 2025 to 2026. The figures come from Zeki Data, a UK analysis firm, which shared them with the magazine.
Zeki also tracked the ratio of arrivals to departures in those roles. It fell from roughly 12 to one in mid-2023, to about two to one this quarter. Chief scientist Jeff Dean, Oriol Vinyals, and Quoc Le left on the same afternoon in August. Google DeepMind did not respond to Fortune’s request for comment on the Zeki findings.
Two caveats belong with those numbers. DeepMind still hires more researchers than it loses, globally. And Zeki built its analysis from public information, which means it undercounts by design.
The word doing the work
Fortune put “cofounders” in quotation marks, and the punctuation is earned. Companies House records show a Canadian entrepreneur, George Rose, incorporated Ineffable in November 2025. Silver became a director in January 2026. Rose ceased to be a person with significant control the following month and resigned as a director in July. A company spokesperson said Rose helped Silver get the company running before moving to an advisory role, and that Silver does not consider him a cofounder.
So the man who registered the company is a founding adviser, and six people who arrived this summer are cofounders. The title works as compensation. Equity and status move researchers between labs, and a founding credit costs less than cash. It also recruits the next hire.
There is also an unresolved wrinkle in Fortune’s own reporting. The August piece placed Czarnecki as chief technology officer at a company called Fundamental. The September piece has him leading Ineffable’s science team. Either he moved twice inside a year, or one of the two descriptions is out of date. TNW has not independently confirmed which.
What has to be true
Reinforcement learning wins wherever the rules are explicit. Go, chess, StarCraft II, and Atari all supply a clean signal for what counts as winning. Nobody has yet shown the method holds when the goal runs open-ended and the reward stays ambiguous. That describes almost every useful task in the world.
Ineffable now has the people who built the wins inside the games. The bet costs $5 billion and rests on the claim that the method travels outside them. The first test is not a product launch. It is whether the seventh, eighth, and ninth researchers of this calibre decide the title is worth having.
Get the TNW newsletter
Get the most important tech news in your inbox each week.
