Sifted Podcast on YouTube

Radical Ventures' Aaron Rosenberg on DeepMind's alumni wave

Inside the talk

Aaron Rosenberg spent six years at Google DeepMind, three of them as Demis Hassabis's deputy, and now runs Radical Ventures' UK office. He talks to John Thornhill about Radical's $1B growth fund, the DeepMind staff leaving to start companies, and the London neo-labs he has backed.

  1. Chief of staff at McKinsey, then deputy to Demis Hassabis

    1:46
    Career

    He joined DeepMind's office of the CEO about nine years ago, when the lab had around 500 researchers. He opened offices, set up the ethics and safety group and worked on commercialising research.

  2. Left DeepMind for Radical in 2023

    3:27
    Career

    Six years in: three as Hassabis's deputy, three as head of strategy and operations. Radical was set up in 2017 to invest only in AI, around Geoffrey Hinton's lab in Toronto.

    His old McKinsey boss, Dominic Barton, joined Radical at the same time
  3. Offices in Toronto, San Francisco, New York and London

    5:59
    Radical

    Most of the portfolio is in North America. In Europe, Radical looks at Paris, Munich, Switzerland, Stockholm and Israel, the cities with strong technical universities.

    Rich Kotite joined the London office from Accel this year
  4. $1B+ first close for the Breakout Fund

    7:19
    Radical

    The largest venture fund in Canadian history, announced in September. It is for late rounds: AI companies stay private longer and now go public at valuations above $1 trillion.

    Early cheque: Discovery Loop, from Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le
  5. An investment boom and a bubble at the same time

    8:44
    Thesis

    A lot of capital is chasing few deals, so prices rise. Radical works from a thesis years ahead and tries to invest before others believe in a company.

    “not too early, but just early enough” 9:49
  6. Radical invests in every layer of the AI stack

    10:04
    Thesis

    Compute (Crusoe), chips (Etched), grid management (Emerald), models (Cohere, Reka, World Labs), biology (Latent Labs) and applications (Slingshot, for mental health).

    World Labs sold to AMD for more than $8B
  7. Buyers want AI that is neither American nor Chinese

    11:35
    Thesis

    After 6 to 12 months of upheaval in the US, companies want providers a government cannot switch off. Cohere, where Radical was the first investor, merged with Aleph Alpha and is backed by Schwarz Group.

  8. AlphaFold won a Nobel Prize but was not obviously valuable

    13:32
    DeepMind

    Colin Murdoch, formerly chief business officer of Google DeepMind and president of Isomorphic Labs, has joined Radical. At its free founder masterclasses he explained how DeepMind built value from research in steps, through partnerships.

  9. Often the first call for DeepMinders thinking of leaving

    15:09
    DeepMind

    He acts as a sounding board, sometimes for staying. For those who go, he explains a funding market that is hard to read from inside a lab.

  10. 500 → 6,000 DeepMind staff when he joined, and today

    16:32
    DeepMind

    At 6,000 people some attrition is normal. Gemini work is moving to Mountain View, and London staff are deciding whether to follow or start their own company.

  11. London neo-labs each bet on one research path

    18:24
    Neo-labs

    World models (Emulate, Odyssey), diffusion language models (Revolution Labs), reinforcement learning (David Silver's Ineffable). Google DeepMind now runs like an assembly line for Gemini, so researchers leave to get resources for other ideas.

  12. 40+ neo-labs have raised about $40B between them

    20:53
    Neo-labs

    Radical's count, worldwide. The bar for each next round is high and some consolidation is likely. Most work on approaches the big labs have deprioritised for their large transformer models.

  13. European neo-labs can become global champions

    24:00
    Neo-labs

    The risk is that research is unpredictable. A lab that misses state-of-the-art results cannot raise its next round.

  14. What Radical looks for in a founder

    25:36
    Thesis

    Hassabis is his model: technically exceptional, able to explain the work to any audience, commercial, and quick to hire around his own gaps. Radical turns down most neo-lab pitches; a contrarian view backed by evidence gets attention.

  15. $50M seed round for Inherent, out of stealth in May

    29:31
    Neo-labs

    Radical and Index Ventures co-led it. The London lab treats its own organisation as an experiment, on the model of electrification: factories only got more productive once they were redesigned around the new power.

    CEO Tantum Collins, the first person Rosenberg knew at DeepMind
  16. Inherent staff narrate their work to an AI agent

    32:12
    Neo-labs

    Employees wear headsets so Faraday, the company's agent, gets the context a colleague would have: Slack, email, conversations. The aim is teams of people and machines working as one.

    “a team of centaurs” 33:43
  17. Non-compete reform should mean more DeepMind spin-outs

    36:08
    DeepMind

    Restrictive covenants have held back staff who leave to start companies. Without them, Rosenberg expects more companies at the earliest stage and founding teams that grow faster. Inherent lobbied for the change.

  18. $50M Series A for Latent Labs in 2025

    37:15
    Neo-labs

    Founder Simon Kohl led protein design on AlphaFold. Its models now design proteins from scratch that bind a chosen target and have drug-like properties, and it signs deals with pharma companies.

  19. 70+ former DeepMinders in his London alumni group

    40:07
    DeepMind

    The Deep Alum WhatsApp group meets in person, most recently when Oriol Vinyals was in town. Rosenberg uses it to check founders' technical claims with experts.

  20. Quick-fire round

    44:27
    Career

    His tools: Cohere North agents for meeting briefs, Gemini, Claude for coding, Wispr Flow for dictation. Most talented founder: Hassabis. The startup he would join: Inherent. Biggest miss: no answer.

20 moments from the 48-minute talk. Each time opens the video on YouTube at that point.

More top stories