MIT CSAIL 实验 GPT-6 Astra 直接修改神经网络权重
MIT CSAIL 团队报告初步实验，称 GPT-6 Astra 能在不编写训练循环的情况下解读并修改神经网络权重。在 Qwen 模型测试中，微小的权重调整显著提升了数学与指令跟随得分，但该发现尚处于早期阶段。
报道正文
An MIT CSAIL team reports early experiments in which GPT-6 Astra interpreted and modified neural-network weights without writing a training loop. In one Qwen test, the authors say a small weight change sharply improved math and instruction-following scores; the findings are preliminary.
Rikiya Takehi and colleagues / MIT CSAIL·623 comments
Rikiya Takehi and colleagues / MIT CSAIL
Andrej Karpathy Builds a Video Library From Hacker News Recommendations
HackerTube gathers 6,454 YouTube videos mentioned by Hacker News readers across nearly two decades. Karpathy describes using AI coding tools to build an alternative browsing experience, with its own recommendations and no user accounts.
Andrej Karpathy·624 comments
Screenshot via Andrej Karpathy
作者发帖说明（HN）：
I thought it would be fun to browse HN as if it were a newspaper. The site updates once an hour. Enjoy!