Skill Issue:代码智能体与自动化研究

Skill issue: code agents and autoresearch

安德烈·卡帕西 Andrej Karpathy · No Priors 播客 · 2026-03-20 · 约 67 分钟 · 原视频 ↗

本期速览 · Overview

代码智能体还差在哪，以及通往自主研究那条「绕圈」的路。

Where coding agents fall short, and the loopy road to autonomous research.

嘉宾自去年 12 月起未写一行代码，全权委托给智能体。
Speaker hasn't typed code since Dec, delegates all to agents.
瓶颈从打字速度转向指挥智能体的技能。
Bottleneck shifts from typing speed to skill in directing agents.
自动研究移除人类环节，能发现人类遗漏的改进。
Auto-research removes humans to find improvements they miss.

自去年 12 月以来，我没写过一行代码；所有工作都委托给智能体。
Since December, I haven't typed a line of code; I delegate everything to agents.
瓶颈不再是打字速度，而是你指挥智能体的技能。
The bottleneck is no longer typing speed but your skill in directing agents.
自动研究证明，将人类移出循环可以发现人类忽略的改进。
Auto research shows that removing humans from the loop can find improvements humans miss.
大语言模型智能参差不齐：在可验证任务上出色，但还在讲多年前的烂笑话。
LLMs have jagged intelligence: brilliant at verifiable tasks, but still tell the same bad joke from years ago.
行业应重构为智能体优先的工具，用 API 取代应用，智能体作为粘合剂。
The industry should refactor for agent-first tools, where APIs replace apps and agents are the glue.
开源模型落后前沿几个月，但这有利于去中心化。
Open-source models lag behind frontier by months, but that's healthy for decentralization.

自去年 12 月以来我没写过一行代码，全交给智能体了。
I haven't typed a line of code since December; I only delegate to agents.
大语言模型就像天才博士生和十岁小孩的结合体。
LLMs are like a brilliant PhD student combined with a 10-year-old.
模型并未跨领域泛化智能；讲笑话的质量毫无提升。
Models are not generalizing intelligence across domains; joke quality hasn't improved.
前沿实验室正在积极自动化淘汰自己的研究人员。
The frontier labs are actively automating away their own researchers.
我们应该为智能体写 Markdown，而不是为人类写 HTML 文档。
We should write markdown for agents, not HTML docs for humans.
互联网上不可信的智能体集群可能超越前沿实验室。
An untrusted swarm of internet agents could outcompete frontier labs.