Dario Amodei, CEO of the $19 billion AI startup Anthropic, doesn’t think humanity is in immediate danger from the evolution of artificial intelligence models and tools. But he has a problem with the justification some of his peers use for risk-taking.
Speaking at an AI conference in San Francisco on Wednesday, Amodei argued that AI models are and always will be mere chatbots with limited capabilities. He was particularly criticized by Marc Andreessen, a famous venture capitalist and champion of limitless AI, who has dismissed concerns, arguing that AI is just math. “Limiting AI means limiting math, software and chips” Andreessen tweeted in March.
This logic does not hold, according to Amodei; because everything in the world can be classified as mathematics, he said.
“Isn’t your brain just math? A neuron fires and sums up calculations, that’s math too,” he said across the stage. Eric Newcomer’s Cerebral Valley lecture on wednesday “We shouldn’t be afraid of Hitler, it’s just math. The whole universe is mathematics.’
Amodei, a former OpenAI vice president who left in 2021 to found rival firm Anthropic LLM, is among a group of AI executives who are outspokenly wary of the technology. potential risksfrom rugged models to bad actors. The CEO supports some regulation of the AI industry and Anthropic even co-sponsored a controversial California bill to do so, viz. after all, vetoed.
It has Andreessen, his VC firm invest Billions of dollars in AI companies, including OpenAI and Elon Musk’s Xai, are on the other side: an AI “boomer” that demands unfettered development of AI technology by individual companies. “The ‘regulation’ of AI (mathematics) is the basis of a new totalitarianism,” the VC wrote last year. He also called AI security critics or doomers “a worship“. A representative for Andreessen declined to comment on Amodei’s criticism.
Amodei admitted at the conference that today’s AI models are “not smart enough … not autonomous enough” to pose a serious risk to people. But he noted that the technology is evolving rapidly, with AI “agents” capable of acting autonomously on behalf of a human command. As these nascent tools come to the fore, Amodei said the public will gain a deeper sense of what AI is capable of and the potential harm.
“People laugh today when chatbots say something unexpected,” Amodei said. “But we’re going to have to do a better job of controlling the agents than that.”
Are you a tech company employee or someone with an idea or tip to share? Contact Kali Hays securely the signal at +1-949-280-0267 or at kali.hays@fortune.com.
