Newshot
Account

Former Anthropic researcher 考克森: Retraining Mythos today would take only a fraction of the original compute

Source: IT之家 Published: AI 中文 训练AnthropicComputeMythos
AI summary

A researcher who previously worked at OpenAI and Anthropic and drew widespread attention for publicly warning about AI risks says a slide shown during Anthropic's IPO roadshow was an important reason for his departure. He argues that the Mythos model released by Anthropic this spring was not an accidental breakthrough but a node in a super-exponential growth trend, and that retraining Mythos now would require only a small fraction of the compute used initially.

Translation(model translation, not the source's own text)

Former Anthropic researcher 考克森: Retraining Mythos today would require only a small fraction of the original compute

IT之家, November 14: Researcher Jacob · 考克森, who previously worked at OpenAI and Anthropic and drew widespread attention for publicly warning about AI risks, said that a slide shown during Anthropic's IPO roadshow was an important reason that prompted him to leave the company.

考克森 worked at Anthropic for only about four months. One of the reasons he decided to leave was a chart in the company's IPO roadshow showing changes in compute demand for AI model training. The chart used compressed axes, so that even a 10x and 100x increase in compute demand looked roughly similar.

Even so, the curve still rose sharply. 考克森 believes this shows that algorithmic efficiency is not only improving rapidly, but that the pace of improvement is still accelerating. "This chart is truly astonishing; I didn't expect they would make it public."

考克森 believes that the major performance breakthrough of the Mythos model Anthropic launched this spring was not accidental, but the result of accelerating gains in algorithmic efficiency. He wrote: "Mythos is not an accidental breakthrough, but a node in a superexponential growth trend. Now, they can retrain Mythos with only a small fraction of the compute used for the original training."

考克森 points out that algorithmic improvement is about to become automated, and therefore advocates pausing frontier AI model development next year to give competitors a chance to catch up. He believes this would turn human-level intelligence into a widely available commodity, and that the only parties harmed would be OpenAI and Anthropic.

考克森 once led AI model training work at OpenAI, and his most recent job was at Anthropic. In September this year, he publicly warned about the risks that advanced AI systems could pose, drawing widespread attention, and other important figures in the industry have issued similar warnings. They worry that advanced AI capable of continuously self-improving could eventually escape human control.

Related articles

  • 黄仁勋 resigns as CEO of 英伟达, potentially adding uncertainty to Anthropic's "largest IPO in history" next week
  • Anthropic issues multiple outage notices in succession: Microsoft service disruption affects Claude
  • Claude sees a large number of errors, Anthropic's official notice says it is investigating
  • Anthropic IPO subscription demand about 6x: plans to raise up to $64.4 billion, target valuation about $1.4 trillion
  • OpenAI's latest private funding round of $15 billion, pre-money valuation of $1.25 trillion
  • Warning about AI risks while developing advanced models: five AI giants including OpenAI, Anthropic, and Google sued, asked to halt ChatGPT, Claude, Gemini

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗