Anthropic CEO Dario Amodei sat down for a five-hour podcast interview with Lex Fridman, where they discussed the future of artificial intelligence. Despite evidence to the contrary, Amodei believes that scaling up models is still a viable path towards more capable AI.
Amodei’s approach involves increasing both the amount of compute used to train models and the size of models themselves. He notes that “probably, the scaling is going to continue” and there’s some magic to it that hasn’t been fully explained on a theoretical basis yet.
However, Amodei also acknowledges potential challenges, such as the growing cost of AI computing and data limitations. He expects companies to spend billions of dollars on clusters to train models next year and hundreds of billions by 2027.
Amodei is also candid about the unpredictability of even the best models, citing a “whack-a-mole” aspect where pushing one thing causes other issues to arise. Despite this, he anticipates that Anthropic or a rival will create a superintelligent AI by 2026 or 2027, exceeding human-level performance on various tasks.
The implications of such a development worry Amodei, particularly with regards to economics and power concentration. He’s concerned about the abuse of power, but remains optimistic about his ability to address these issues as CEO.
Source: https://techcrunch.com/2024/11/13/this-week-in-ai-anthropics-ceo-talks-scaling-up-ai-and-google-predicts-floods