Pular para o conteúdo

Ex-DeepMind Scientist WARNS: Why We Still Can’t Trust Today’s AI

2 de outubro de 2026 · 1 h 43 min

Notas do episódio

Is "AI is dangerous" a warning, or the world's best marketing line?


This episode is brought to you by AgelessRx, where aging becomes a puzzle that can be solved. The mission is to help you live longer and healthier by adding more life to your years through science-backed solutions and personalized care. Learn more at agelessrx.com/beyond-tomorrow.


On this episode of Beyond Tomorrow, I sit down with Thore Graepel, co-creator of AlphaGo and former researcher at Google DeepMind. I asked him straight up how close we are to superintelligence, and his answer surprised me: there won't be a single moment we can point to, and he argues human organizations like companies already count as a form of superintelligence. We get into why he thinks "AI will destroy humanity" is the biggest myth in the field, the near-term risks he does take seriously, and why I pushed him on the incentives behind the AI safety messaging coming from the labs.


From there we go deep on his core technical argument: that today's transformers are black boxes that mix knowledge and reasoning in the same weights, and that scaling alone will never make them trustworthy. He lays out a different approach built on explicit beliefs, hypotheses and evidence, and I found it one of the clearest explanations of AI transparency I've heard. In the back half, we talk about his time at Altos Labs, why biology is a reality check for the "cure all disease in 10 years" claims, and his cell annealing theory of rejuvenation.


Key Takeaways

  • Thore sees AI progress as a "jagged frontier," not a single superintelligence moment, and by his own definition, AGI is already here

  • He argues companies and institutions are already a form of superintelligence

  • His P(doom) is very low: extraordinary claims require extraordinary evidence, and he doesn't see it


About Thore Graepel

Thore Graepel is a computer scientist and AI researcher who recently left Google DeepMind, where he was a co-creator of AlphaGo and worked on multi-agent systems, reasoning and consciousness research. He has also spent time at Altos Labs, applying his machine learning and physics background to cellular rejuvenation, and writes about AI, reasoning and safety on Substack.


⏱ Timestamps

00:00 Intro

00:06 How Close Are We to Superintelligence?

01:42 Are Human Organizations Already Superintelligent?

02:56 Is Anyone Trying to Destroy Humanity?

04:45 Near-Term Risks: Jobs and the Economy

08:08 Why the Transition Will Be Hard

10:06 Post-AGI Planning and the Case for Positive Futures

14:35 Capability vs. Diffusion: The Train Driver Problem

16:46 Optimization vs. Game Theory

18:53 The Incentives Behind AI Safety Messaging

22:26 What Is Thore's P(doom)?

23:25 Why Transformers Are Black Boxes

29:19 How Do We Build Safer, More Transparent AI?

33:01 Separating Knowledge From Reasoning

37:39 Safety vs. Capability and the Trust Ceiling

39:45 Narrow AI vs. General AI

44:02 Do You Believe in God?

47:15 Consciousness and Whether LLMs Are Alive

51:08 AlphaGo and Move 37

58:29 Credit Assignment and Moral Reasoning

01:04:31 Who Sets the Moral Guardrails?

01:09:37 The AGI Race: US vs. China

01:11:53 Longevity and Altos Labs

01:16:01 The Biology Reality Check

01:23:04 Pockets of Computability and AlphaFold

01:26:05 Yamanaka Factors and Cell Annealing

01:35:51 Will the Singularity Happen in Your Lifetime?

01:36:49 Rapid Fire Questions with Thore

01:41:35 Thore's Best Piece of Advice


Resources & Mentions

AlphaGo and AlphaFold (Google DeepMind)

Thinking, Fast and Slow by Daniel Kahneman

Altos Labs

Stephen Wolfram's "pockets of computability"

David Deutsch on simulation and the limits of experiment

Thore's Substack


Connect with Thore:

Github: https://thoregraepel.github.io/

LinkedIn: https://uk.linkedin.com/in/thoregraepel

X: https://x.com/ThoreG


Connect with Julian

Instagram: @thebeyondtomorrowpodcast / @juliankissa

LinkedIn: julian-issa

X: @juliankissa

Website: www.beyondtomorrowpodcast.com

Email: [email protected]


Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Todos os episódios de The Beyond Tomorrow Podcast with Julian Issa

Configurações

Exibição

Idioma

Define quais estações você vê primeiro na página inicial.

Histórico

As rádios e episódios ouvidos recentemente e suas últimas buscas ficam só neste navegador.

Avisos

Um aviso rápido quando um podcast que você segue tem um episódio novo. Desative podcasts individuais pelo sino na página deles.

Conta

Salve favoritos – no app e no site.

Atalhos de teclado

/
Abrir busca
Espaço
Reproduzir ou pausar
← →
No player: voltar 15 s, avançar 30 s (episódios) · na fileira: próximo bloco
↑ ↓
Na fileira: alvo dentro do bloco
Home / End
Na fileira: primeiro ou último bloco
Esc
Fechar janela ou menu
?
Esta visão geral
Tocando agora

Ir para a estação

A seguir

    Adicione episódios com “A seguir” – eles tocam depois do episódio atual.