前 Anthropic 预训练研究员 Jacob Coxon 因路演幻灯片离职
一个前预训练研究员亲口解释自己为啥离职，还晒了内部数据：代码任务算力效率两年多一直在超指数加速。如果你关心模型训练成本和算法进步速度，这条一手信息挺难得。
前 Anthropic 预训练研究员 Jacob Coxon 公开表示，公司 IPO 路演中的一张幻灯片是他决定离职的重要因素。该幻灯片展示了自 2024 年 7 月以来代码任务的预训练算力乘数曲线，采用对数刻度绘制，而曲线仍在持续向上弯曲，意味着算法效率的增长快于指数级。Coxon 的判断是，这种超指数级的算法进步意味着 Mythos 现在只需当初所需算力的一小部分就能完成重新训练。他将此概括为：递归自我改进已经出现在 IPO 路演的推销材料里。
对数刻度上的向上弯曲说明进步速度本身在加速，而不只是线性增长。Coxon 离开的同时公开这一细节，也把 Anthropic 内部对算力效率提升速度的评估摆到了公众面前，投资人版本的研究叙事与一线研究员的实际反应之间出现了分歧。
Ex-Anthropic pretraining researcher Jacob Coxon says this slide from the Anthropic IPO roadshow was a big factor in his decision to quit.
It shows the pre-training compute multiplier for code since July 2024, on a log scale (!), and the line keeps bending upward.
His take: superexponential algorithmic progress, and Mythos could already be retrained with a fraction of the compute it first needed.
so tl;dr Recursive self-improvement is now part of an IPO pitch.
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Anthropic, Jacob Coxon, pretraining, compute efficiency.
Source attribution: curated from kimmonismus (@kimmonismus). Cross-source corroboration: 11 independent sources have covered this development (including orange.ai, Gary Marcus, 小互), confirming industry-wide signal strength.
