Anthropic co-founder and CEO Dario Amodei is trying to hold two ideas together: AI models should keep improving, and the industry needs a sharper understanding of how they work as that happens.
Speaking with TechCrunch at a developer-focused event in Paris held by Anthropic in partnership with French startup Dust, Amodei described a middle path between unqualified optimism and blanket criticism. His message was that the opportunity around AI depends on making progress in interpretability, testing, measurement, and model development at the same time.
A Push For Urgency After The AI Action Summit
Amodei’s remarks came right after the end of the AI Action Summit in Paris. In a statement released on Tuesday, he called the summit a “missed opportunity” and said that “greater focus and urgency is needed on several topics given the pace at which the technology is progressing”.
That concern shaped the discussion at Anthropic’s own Paris event. Amodei said his background as a neuroscientist now maps onto the work of understanding AI systems. He said, “I used to be a neuroscientist, where I basically looked inside real brains for a living. And now we’re looking inside artificial brains for a living.”
He added that Anthropic expects progress in interpretability “over the next few months,” describing it as work aimed at understanding how models operate. But he framed that progress as a race, because model capabilities are advancing quickly across Anthropic and other companies.
“Our understanding has to keep up with our ability to build things. I think that’s the only way,” Amodei said.
Safety And Opportunity Are Not Separate Tracks
The tone around AI governance has shifted since the first AI summit in Bletchley in the U.K. At the AI Action Summit on Tuesday, U.S. Vice President JD Vance said, “I’m not here this morning to talk about AI safety, which was the title of the conference a couple of years ago. I’m here to talk about AI opportunity.”
Amodei’s position is different. He is not presenting AI safety and AI opportunity as opposing goals. Instead, he argues that safety work can become part of the opportunity itself.
At the Anthropic event, he pointed back to the U.K. Bletchley Summit, where there were discussions around testing and measurement for various risks. In his view, that kind of work did not meaningfully slow the technology. He said that measuring models has helped Anthropic understand them better, and that better understanding ultimately helps produce better models.
This is an important distinction in his argument. Amodei is not calling for the industry to stop building frontier AI models. He said, “I don’t want to do anything to reduce the promise. We’re providing models every day that people can build on and that are used to do amazing things. And we definitely should not stop doing that.”
He also said that when people focus heavily on risks, he gets frustrated because, in his view, “no one’s really done a good job of really laying out how great this technology could be”.
DeepSeek Raises A Geopolitical Concern
The conversation also turned to Chinese LLM-maker DeepSeek and its recent models. Amodei said his own reaction was “very little,” because Anthropic had already seen V3, the base model for DeepSeek R1, back in December.
He described V3 as an impressive model, but said it followed a “very normal cost reduction curve” that Anthropic had seen in its own models and in others. What stood out to him was not just the model’s technical profile, but where it came from.
Amodei said the release was notable because it did not come from the “three or four frontier labs” based in the U.S. He named Google, OpenAI, and Anthropic as examples of labs that usually push the edge with new model releases.
That led him to a geopolitical concern. “I never wanted authoritarian governments to dominate this technology,” he said.
He also rejected claims about DeepSeek’s supposed training costs. On the idea that training DeepSeek V3 was 100x cheaper than training costs in the U.S., Amodei said, “I think [it] is just not accurate and not based on facts.”
Claude Models And The Reasoning Question
Amodei did not announce a new model at the Wednesday event, but he did discuss where Anthropic is heading. The company is working on its own approach to reasoning models, with attention to capacity, smarter models, and safety concerns.
One issue he highlighted is model selection. People using ChatGPT Plus, for example, may not know which model to pick from a model selection pop-up. Developers using large language model APIs face a related problem when they try to balance accuracy, speed, and costs.
Amodei questioned the idea that ordinary models and reasoning models should be treated as entirely separate categories. “If I’m talking to you, you don’t have two brains and one of them responds right away and like, the other waits a longer time,” he said.
He described Anthropic’s preferred direction as a smoother transition between pre-trained models like Claude 3.5 Sonnet or GPT-4o and models trained with reinforcement learning that can produce chain-of-thoughts (CoT), such as OpenAI’s o1 or DeepSeek’s R1.
“We think that these should exist as part of one single continuous entity,” Amodei said. “We should have a smoother transition from that to pre-trained models — rather than ‘here’s thing A and here’s thing B,’” he added.
Where Anthropic Sees AI Applications Going
Amodei closed with a broad view of where stronger AI models could matter. He said large AI companies such as Anthropic will continue releasing better models, and he expects that to create opportunities to disrupt large businesses across industries.
He gave one concrete example from pharma. “We’re working with some pharma companies to use Claude to write clinical studies, and they’ve been able to reduce the time it takes to write the clinical study report from 12 weeks to three days,” he said.
He also pointed to legal, financial, insurance, productivity, software, and energy as areas where AI applications could expand. His conclusion was that there will be “a renaissance of disruptive innovation in the AI application space,” and that Anthropic wants to help and support it.
The through line is clear: Amodei sees the future of AI as both a building challenge and an understanding challenge. For Anthropic, the argument is that better models, stronger measurement, and clearer insight into model behavior have to advance together.