Anthropic CEO Dario Amodei warns about ‘racing’ to understand the AI ​​because it is stronger

Immediately after the end AI Action Summit Anthropick co-founder and CEO Dario Amodei in Paris named The incident “missed opportunity”. Added that “Taking into account the pace of the technology, it is necessary to actual and relevance in several issues” Announced on Tuesday.

AI has held an advanced event at the start of French in Paris Dustand TechCrunch, found the opportunity to meet Amodani on stage. The event described the line of thought, and accordingly, AI defended a third way in the topics of innovation and management, nor pure criticism.

“I’ve been a neurocientist, here I looked inside the real brains for a living. So we will make some interesting progress in the field of residence – really started to understand how models work.”

“But it is definitely a race. It is a race between making the models stronger, and you can’t slow down to us? … Our understanding should continue our ability to create things. I think,” he said.

Since the first AI Summit in Bletchley The tone of the discussion around the AI ​​control in the UK has changed significantly. Partially related to the current geopolitical landscape.

“This morning, I am here to talk about AI security, which is the name of the conference a few years ago,” US Vice President JD Vance said on Tuesday at the EU Action Summit. “I’m here to talk about the OI opportunity.”

Interestingly, Amodiii is trying to prevent this antagonization between security and opportunity. In fact, he believes in a centralized center have an opportunity.

“In the original summit, there were many discussions about the test and measurement for the BLetchley Summit, various risks. “If there is something, to make such measurements, helped us better understand our models that help them produce better models.”

And every time I focuses on a number of Amodei security, he also likes to remind the anthropy that the frontage of the frontline.

“I don’t want to do anything to reduce the word. We present the models used to build people every day and do incredible things. And we must not stop doing it.”

“People talk a lot about the risks, I’m a kind of frustration and say,” Oh, man, no one really did a good job, “he said.

Deepseek’s training costs ‘not just exactly’

When the conversation changes China LLM-Maker DeepSeekRecent models, Amodei has lowered technical achievements and said the people felt that the reaction was “inorganic”.

“Honestly, my reaction was very small. The department saw the V3 of the base model for R1 in December. And it was an effective model,” he said. “The model released in December was a curve of reducing normal costs in our models and other models.”

What is what is that the model “Three or four border laboratories” in the United States as some border laboratories that push the envelopes from three or four border laboratory “to the highest model releases.

“And this was a matter of geopolitical concern for me. I never wanted authoritarian governments to dominate this technology,” he said.

As for the probable training expenditures of DeepSEEK, DeepSeek rejected that the V3 of DeepSeek V3 was 100x in the US training expenditures in the United States, “I think (this) is not only accurate and based on facts,” he said.

Claude models upcoming claude with indisxi

If Amodei did not declare a new model in the event on Wednesday, the company closed the approaching releases – and yes, it contains some thinking.

“In general, we are aimed at trying to think of better models. We have enough opportunities for the models to be smarter and concern.

One of the issues he tried to solve the anthropy is the choice. If you have a Chatgpt Plus account, for example, you can be difficult to know which model you choose in the model selection pop-up for your next message.

The Great Language Model (LLM) for its own applications is the same for developers using the API. They want to balance everything between accuracy, response and expenses.

“We have been a little confused by the fact that there are normal models and reasonable models and each other are different and said they are different from each other.” “If I talk to you there are two brains and one of them responds immediately and the other is waiting for a longer period.”

According to him, depending on the entrance, Clod 3.5 Must have a smoother switch between Sonnet or GPT-4O and reinforcement learning and among models, and it is Openai’s O1 or Deepseek R1.

“We think that these should be available as part of a single sustainable body. And yet we can not be there, but the anthropic really wants to act in this direction,” Amodei said. “We have to make a more trained transition to the models prepared from us – more ‘something here and here is something here,” said.

As large AI companies such as anthropic continue to release better models, Amodei believes that each industry will open some great opportunities to break large enterprises in the world.

“We were trying to use a clone to write clinical studies with some Pharma companies and they were able to reduce the time needed to write a clinical education report for three days in 12 weeks,” said Amodiei.

“There are biomedical, financial, finance, insurance, productivity, software, energy around the energy. I think – mainly, we want to help it,” he said.

Read our full coverage Artificial Intelligence Activity Summit in Paris.