Immediately following the conclusion of the AI Action Summit in Paris, Dario Amodei, co-founder and CEO of Anthropic, described the gathering as a “missed opportunity.” He emphasized the necessity for “greater emphasis and urgency on various topics due to the rapid advancement of technology” in a statement issued on Tuesday.
The AI organization partnered with French startup Dust to host a developer-centric event in Paris, where TechCrunch had the chance to interview Amodei live. He elaborated on his perspective, advocating for a balanced approach that neither fully embraces optimism nor embraces complete criticism regarding AI innovation and regulation.
“I began my career as a neuroscientist, delving into actual brains, and now we find ourselves exploring artificial brains. In the coming months, we anticipate significant breakthroughs in interpretability, gaining deeper insights into how our models function,” Amodei shared with TechCrunch.
“The landscape is undeniably competitive. We are racing to enhance these models, and that momentum is incredibly swift for us and our competitors alike — there’s no slowing down. Our comprehension must keep pace with our developmental capabilities. That’s essential,” he continued.
Since the inaugural AI summit in Bletchley, U.K., the dialogue surrounding AI governance has evolved dramatically, partially due to the shifting geopolitical climate.
“I’m not here this morning to discuss AI safety, which was the focus two years ago,” stated U.S. Vice President JD Vance during the AI Action Summit on Tuesday. “I’m focused on the opportunities presented by AI.”
Amodei seeks to mitigate the divisiveness between safety and opportunity concepts, asserting that enhanced attention to safety itself serves as an opportunity.
“During the initial summit in Bletchley, discussions were heavily centered on testing and evaluating various risks. I believe such measures haven’t significantly hindered technological advancement,” noted Amodei at the Anthropic event. “In fact, conducting these measurements has led to a deeper understanding of our models, ultimately enabling us to create superior models.”
In emphasizing safety, Amodei is careful to remind others that Anthropic remains dedicated to developing cutting-edge AI models.
“I don’t want to diminish our potential. We are consistently delivering models that others can leverage to achieve remarkable results, and we must continue this work,” he asserted.
“When conversations predominantly focus on the risks, I find it frustrating. It feels as though we haven’t adequately showcased the incredible potential this technology holds,” he expressed later in the discussion.
The claim regarding DeepSeek’s training costs is “simply inaccurate”
As the dialogue turned to DeepSeek’s recent models, Amodei downplayed their technical accomplishments, characterizing the public’s enthusiasm as “inorganic.”
“To be frank, I had very little reaction. We were aware of V3, which serves as the foundation for DeepSeek R1, back in December — and that model was indeed impressive,” he remarked. “The model introduced then followed a typical cost reduction trajectory we’ve observed in our models and those from others.”
A notable point was that this model did not originate from the “three or four leading labs” in the U.S., which he identified as Google, OpenAI, and Anthropic, known for pushing boundaries with new model releases.
“That raises geopolitical concerns for me; I’m wary of authoritarian regimes dominating this technology,” he added.
Regarding the supposed training costs of DeepSeek, he rejected claims of DeepSeek V3 being 100 times cheaper than U.S. training costs, stating, “I believe that is simply inaccurate and not grounded in reality.”
Upcoming Claude Models with Reasoning Capabilities
While Amodei did not unveil any new models during Wednesday’s event, he hinted at some forthcoming releases that include reasoning capabilities.
“We are generally oriented towards creating our unique version of reasoning models that stand out. We are mindful of ensuring we have adequate capacity, enhancing model intelligence, and maintaining safety,” Amodei stated.
One challenge Anthropic is addressing is the model selection dilemma. For instance, ChatGPT Plus account holders often struggle to decide which model to choose from the selection pop-up for their next query.
Similarly, developers utilizing large language model (LLM) APIs in their applications are tasked with balancing accuracy, speed, and cost.
“We’ve been perplexed by the notion that reasoning models are distinctly different from standard models,” noted Amodei. “In a conversation, an individual doesn’t possess two separate brains; one responding immediately and another delayed.”
He believes that depending on the input, there should naturally be a seamless transition between general pre-trained models like Claude 3.5 Sonnet or GPT-4o and those trained with reinforcement learning that exhibit chain-of-thought capabilities, such as OpenAI’s o1 or DeepSeek’s R1.
“We envision that these should be integrated into a single fluid experience. While we may not have achieved that yet, Anthropic is committed to steering in that direction,” Amodei stated. “We should facilitate a more cohesive transition from pre-trained models versus separating them into ‘this is model A and this is model B,’” he further explained.
As prominent AI firms like Anthropic progress in releasing advanced models, Amodei is optimistic about unlocking tremendous opportunities for disruption across diverse industries worldwide.
“We are collaborating with pharmaceutical companies to utilize Claude in drafting clinical studies, drastically reducing the time required to generate clinical study reports from 12 weeks to just three days,” shared Amodei.
“Beyond biomedicine, this applies to legal, financial, insurance sectors, productivity, software, and energy domains. A wave of innovative disruptions in AI applications is on the horizon, and we are eager to contribute and support this movement,” he concluded.
Explore our comprehensive coverage of the Artificial Intelligence Action Summit in Paris.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
