Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

CONFIRMED: LLMs have indeed reached a point of diminishing returns

Nov 10, 2024 - garymarcus.substack.com
The author, Gary Marcus, has long warned about the limitations of scaling in artificial intelligence (AI), arguing that simply adding more data and compute power without fundamental architectural changes would not yield infinite improvements. Despite receiving criticism from industry figures like Sam Altman, Greg Brockman, Yann LeCun, and Elon Musk, Marcus maintains that the truth is now emerging: scaling is reaching its limits, as evidenced by recent comments from venture capitalist Marc Andreesen and industry editor Amir Efrati. Marcus argues that the high valuations of companies like OpenAI and Microsoft, based on the belief that large language models (LLMs) will evolve into artificial general intelligence, are misguided and potentially economically disastrous.

Marcus criticizes the media and tech influencers for perpetuating hype around LLMs and ignoring the principled limits of these models. He warns that the U.S.'s AI policy, driven by this hype and the assumption of endless returns from LLM scaling, could leave the country vulnerable if adversaries invest in alternative approaches. Marcus believes that while LLMs will not disappear, their role may be smaller than anticipated, and achieving reliable, trustworthy AI may require a return to the drawing board. He expresses hope that the market is finally recognizing the truth of his warnings, which could pave the way for real progress in AI.

Key takeaways:

  • The author, Gary Marcus, has long argued that the current approach to improving AI by simply adding more data and computational power, without making fundamental architectural changes, is unsustainable.
  • He believes that this approach, known as 'scaling', will not solve issues such as hallucinations or abstraction in AI, and that the industry is now starting to recognize this.
  • Marcus warns that the high valuations of companies like OpenAI and Microsoft, which are largely based on the assumption that continued scaling will lead to artificial general intelligence, are based on a fantasy and could lead to a financial bubble.
  • He criticizes the media and tech influencers for not listening to scientists and for glorifying the hype around AI, and warns that the US's AI policy, which is largely based on the assumption that returns for scaling will not diminish, could prove to be a massive mistake.
View Full Article

Comments (0)

Be the first to comment!