This is a same-stage conversation between Google DeepMind CEO Demis Hassabis and Anthropic CEO Dario Amodei, reunited after a year (the host likens it to 'putting the Beatles and the Rolling Stones on one stage'). The theme is 'The Day After AGI,' but the opening first recalibrates 'how fast we get to that day.'

On timelines the two still disagree. Dario said in Paris last year that '2026-27 will bring models that reach Nobel-laureate level across multiple fields,' and he basically holds to that call: the path is 'make models great at coding and AI research, then use them to build the next generation and accelerate R&D, forming a closed loop.' He says Anthropic already has engineers who 'don't write any code themselves — they only let the model write it and handle the editing and the surrounding work,' and he expects that within 6-12 months models will be able to do 'most, perhaps all' of software-engineering work end-to-end; the rest comes down to how fast this self-improving 'loop' closes (chips, chip fabrication, training time and other links can't be accelerated by AI). Demis holds to his more cautious framing of a '50% probability by the end of this decade': fields like coding and math, where 'output is verifiable,' are easy to automate, but much of natural science must be verified through experiments and will be slower; harder still is the highest form of scientific creativity — 'posing the question/hypothesis itself' — which may still be missing 'an ingredient or two.'

On the competitive landscape, the biggest change over the past year is the 'reversal in the rankings': a year ago, after the DeepSeek moment, everyone thought DeepMind was behind OpenAI; now Gemini 3 has pulled Google back to the top and set off a 'code red' inside OpenAI. Demis says he always believed DeepMind's 'deepest and broadest research bench' could return it to first place, the key being to recover a startup mentality and intensity. On the pointed question of 'whether independent model companies can survive until revenue materializes,' Dario responds with a set of numbers: revenue has grown 10x/year for three years (2023 from 0 to $100M, 2024 to $1B, 2025 to $10B), approaching the scale of the world's largest companies; he stresses that the ones who truly win are companies 'led by researchers, treating hard science problems as their North Star' — something Google and Anthropic share.

On risk, Dario reveals he is writing a sequel to 'Machines of Loving Grace' (the optimistic piece) — an essay about risk ('the upbeat one was easier and more fun to write, so I wrote it first'). He uses a scene from the film 'Contact' as a frame: how humanity survives 'the adolescence of technology' without self-destructing. Specific worries include how to keep highly autonomous systems smarter than humans under control, preventing individual misuse (bioterrorism), preventing abuse by authoritarian regimes (he names the CCP), plus labor-market shocks and 'unknown risks we haven't even thought of.' On employment, the host notes that so far the labor market shows no obvious AI shock (more like a giveback of post-pandemic over-hiring); Demis agrees the near term is 'normal evolution,' with new tools creating more meaningful new jobs, though this year we may start to see entry-level/intern hiring slow. Dario reconciles 'half of entry-level white-collar jobs gone in 1-5 years' with 'AI broadly superhuman in 1-2 years': in between lie 'lag' and 'substitution,' and while the labor force is adaptable (farmer → factory worker → knowledge worker), he worries exponential compounding will 'overwhelm our speed of adaptation.' Both criticize governments and the economics profession for having 'studied this far too little.'

Geopolitics is the toughest stretch. Facing the contradictory policy of 'America being hawkish on China while selling it chips,' Dario insists that 'not selling chips' is the most effective way to buy time, and likens selling chips to China to 'selling nuclear weapons to North Korea for Boeing's profits'; he says that as long as chips aren't sold, the race is no longer 'between the US and China' but just 'between me and Demis' — and that one 'can absolutely be worked out.' Both wish for CERN-style international collaboration and minimum safety standards, and even wish the overall pace were a bit slower so things could be done right — but both admit that, in a reality where 'rivals are sprinting at a similar speed,' an enforceable slowdown agreement is hard to reach.

The closing turns to doomers and model 'deceptive' behavior: from its founding Anthropic has relied on 'mechanistic interpretability' to 'look inside the model's brain,' increasingly documenting and trying to correct bad model behavior. Both are clear that they are 'not doomers' — the risk is real, but as long as we collaborate and use scientific methods it can be controlled; the truly dangerous scenario is 'everyone racing, sprinting ahead without guardrails.' Asked whether they are more optimistic or more worried after a year, Demis says he has done AI for over 20 years and saw both the upside (the ultimate tool for science) and the risk (a double-edged technology that bad actors will repurpose) from the start; the key is 'whether there's time, focus, and the best minds working on it together.' On an audience question about the Fermi paradox, Demis thinks humanity has 'most likely already passed the Great Filter' (the hardest step may have been the evolution of multicellular life), and what comes next is 'written by humanity itself.' Asked what the biggest change will be when they reconvene next year, both answer in unison: the trajectory of 'AI systems building AI systems'; Demis adds that world models, continual learning, and robotics may reach breakthrough moments, and half-jokingly says 'maybe we should all hope it comes a bit slower — better for the world.'

(For AI investors: ① both frontier labs treat 'coding automation + the self-improvement loop' as the main engine toward AGI, with AI coding being the link closest to closing that loop; ② Anthropic's revenue curve is extremely steep, but the capital sustainability of independent model companies remains a key variable; ③ chip export controls on China are a policy through-line Dario repeatedly bets on, bearing directly on the shape of the compute chain.)