跳到正文
原文
The Decoder:AI News· Matthias Bastian·· 2 天前AI 评分45

前研究员Coxon称Anthropic只需一小部分算力即可重新训练Mythos模型

Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

文章要点
  • 1前Anthropic研究员Coxon称Mythos并非一次性跃升,如今只需当初一小部分算力即可重新训练
  • 2他入职约四个月后离职,称Anthropic IPO路演的一页图表是关键原因,曲线在对数坐标下仍陡峭爬升
  • 3该图展示2024年7月至2026年10月代码预训练算力乘数,Anthropic称数据点为内部估算仅供示意
  • 4他称算法改进接近可被自动化,主张明年暂停前沿模型研发,认为暂停只会损害OpenAI与Anthropic
决策要点

一名离职研究员借Anthropic自家IPO路演图表宣称前沿模型训练成本正加速缩水,这是内部人的个人解读而非可独立验证的披露。

  • 影响:若其判断成立,前沿能力的复制成本会持续下降,企业自建与采购模型的预算假设需要重算;Anthropic只承认图中数据是内部估算,据此做投入决策应保留折扣;关键研究员因路线分歧离职,也是评估供应商稳定性的一个信号。
  • 参考:可考虑把算法效率带来的训练成本下降纳入模型选型的敏感性测算,区分路演数字与独立验证;值得跟踪前沿实验室的人员变动与公开分歧,作为供应商稳定性评估的一项输入。
  • 依据:核心依据是Coxon对Anthropic路演图表的个人解读,Anthropic口径称图中数据点为内部估算仅供示意,暂无独立验证,暂停研发的呼吁属其个人主张。
正文

Jacob Coxon, former OpenAI and Anthropic researcher who brought AI fears into the mainstream, says a slide from Anthropic's IPO roadshow was a major reason he quit.

Coxon says a slide Anthropic released for its IPO roadshow was a key reason he left the company after only about four months. The chart shows how much compute Anthropic needs over time to train its models, and the axis is compressed so that a 10x and a 100x increase barely look different.

Yet the curve still climbs steeply, which Coxon says shows that algorithmic efficiency is improving not just quickly but at an accelerating rate. "It's an insane plot and I'm surprised they made it public," he writes.

Coxon argues that Anthropic's Mythos model, a major capability jump in the spring, was part of this accelerating trend. "Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed," he writes.

This slide from Anthropic's IPO roadshow shows the "Pre-training compute multiplier" for code over time. The y-axis is logarithmic. Former Anthropic researcher Jacob Coxon says the steep climb from July 2024 to October 2026 points to superexponential algorithmic progress. Anthropic notes that the data points are based on internal estimates and are for illustration only. | Image: Anthropic

He says algorithmic improvements are "on the verge of being automated," which is why he wants frontier model development paused next year so competitors can catch up. A pause, he writes, would "commoditize human-level intelligence. It would only harm OpenAI and Anthropic."

Coxon previously trained AI models at OpenAI and, most recently, Anthropic. He drew wide attention in September with public warnings about the risks of advanced AI systems, echoed by other protagonists in the industry. One of their fears is that advanced, self-improving AI could get out of control.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

来源:The Decoder:AI News · the-decoder.com