The AI debate that started with one explosive claim
1. The starting point: The discussion began with an All-In Podcast conversation about Anthropic’s confidence and the claim that it might eventually become the only private company in the world. Gavin BakerDealroom has a profile for this one. Try Dealroom → said he had heard a version of this from several people he trusted, and that he would advise Dario AmodeiDealroom has a profile for this one. Try Dealroom → not to make such a claim. Sholto DouglasDealroom has a profile for this one. Try Dealroom → said the claim was completely false. He also argued that Dario’s public comments about AI risk had made the story sound believable to some people, even if Dario had not actually said it.
2. Gavin’s view: Gavin is concerned that the most pessimistic view of AI is becoming the dominant public story. He thinks that could lead to bad policy and make people underestimate the possible benefits of advanced AI. He also supports open-source models that people can run themselves, rather than a world controlled by one or two companies. In his view, a wide choice of models is safer and better reflects the fact that people and societies have different needs.
Gavin agreed with Sholto’s distinction between cyber and biological risks: cyber attacks are often easier to defend against because defenders can use large amounts of computing power to find and fix weaknesses, while biology may give attackers more ways to cause harm. Gavin said this was probably the strongest case for regulating advanced AI. He also said the US government should prepare for serious risks, while keeping open the possibility that rapid AI progress could improve living standards and help solve major problems.
3. Sholto’s view: Sholto made the cyber-versus-biology distinction, but stressed that it is not a simple rule. In biology, an AI system might greatly improve a person’s ability to understand or design harmful pathogens, but real-world experiments, materials, and testing could still slow them down. In other words, even a very powerful model may not be able to improve itself or change the world instantly: it may still face practical limits outside the computer.
Sholto also argued that governments should prepare for major changes in employment without pretending that those changes have already happened. He would not support immediate intervention while employment remains strong, but thinks governments should develop plans for different scenarios — including one in which jobs such as data entry and call-centre work disappear quickly. His point is that people are more likely to cope with disruption if there is a credible plan before it arrives.
4. Dario’s view: Dario joined the exchange to address two issues. On regulation, he said the choice is not simply between putting AI entirely in the hands of one company and allowing unrestricted access to powerful systems. The harder question is how to control serious risks while still allowing competition, innovation, and broad public benefit.
On the tone of his public statements, Dario rejected Gavin’s suggestion that he talks about AI in a disproportionately negative way. He said his major writing has given roughly equal weight to the risks and the benefits, even if interviews about safety naturally focus more on the dangers. His response also showed his preference for discussing these questions openly, rather than leaving them to rumours or private conversations.
5. David’s view: On 17 August 2026, David SacksDealroom has a profile for this one. Try Dealroom → responded to Dario’s post and broadened the debate around regulation and centralisation. Sacks argued that the concern is not opposition to all regulation, but the risk of regulatory capture: rules shaped by incumbent companies and agencies in ways that raise barriers for challengers. He criticised proposals for a federal pre-release approval process for frontier models, comparing the possible result to a “DMV for AI”, with long queues and accumulating requirements that could disadvantage the United States relative to China. He also argued that a powerful frontier lab such as AnthropicDealroom has a profile for this one. Try Dealroom → would be well placed to navigate and influence such a system, potentially reinforcing the concentration of power it claims to manage.
6. What they agree on — and where they differ: The four agree that advanced AI could bring very large benefits and very large risks, and that different risks need different responses. They also agree that the public conversation should be serious and conducted in good faith. Gavin, Sholto and Sacks all warn that excessive centralisation or poorly designed regulation could entrench incumbents, while Dario focuses on finding a middle path: reducing the risks of powerful systems without handing control to a single company or shutting down useful competition.
The main differences are about emphasis. Gavin focuses on openness, competition, and the danger of allowing fear to shape policy. Sholto focuses on technical detail, real-world bottlenecks, and preparing for economic disruption before it becomes a crisis. Dario focuses on balancing serious risk reduction with competition and public benefit. Sacks focuses on regulatory capture, the incentives of frontier labs, and the risk that centralised oversight could reinforce the position of the largest incumbents. They are not arguing about whether AI matters. They are mainly arguing about which risks should receive the most attention, how quickly they might arrive, and how much precaution is appropriate.
Read more: The All-In Podcast discussion · Gavin Baker’s opening response · Sholto Douglas’s reply · Gavin Baker on cyber, biology, and regulation · Sholto Douglas on fast takeoff and practical limits · Dario Amodei’s response · David Sacks’s response
