twaldin

4 ms·

1. 1.▲
   GEPA prompt optimization: Claude Code Haiku +20% solve rate on new bugs(tim.waldin.net)

   3 points by twaldin 6mo ago|0 comments
2. 2.▲
   by twaldin 6mo ago

   Yea this makes sense to me; if they were just analyzing student commit history, but 'hiding & executing code' is conceptually dangerous, facilitates their surveillance, even if they might say "It's just git commit &a
3. 3.▲
   by twaldin 6mo ago

   its realistic to assume anthropic (and other frontier labs) have already gobbled the entire surface of the public internet into training data, plus so much more llm-generated traning data, that if you have public works and a model has enoug
4. 4.▲
   by twaldin 6mo ago

   Sort of, I'm hesitant to say surveillance, since its less of 'running spyware' or something similar and more 'tracking student commit history'; where it gets weird is this section of the paper: In our system, the Ma
5. 5.▲
   by twaldin 6mo ago

   idk over my testing, glm-5 inside opencode beats all other agents head to head
6. 6.▲
   Purdue University CS240 Class over 50% of students 'caught' using AI on homework(old.reddit.com)

   4 points by twaldin 6mo ago|5 comments
7. 7.▲
   by twaldin 6mo ago

   If anyone is interested, the professor, Jeff Turkstra, wrote a paper called 'Tracking Large Class Projects in Real-Time Using Fine-Grained Source Control' https://turkeyland.net/research/encourse.pdf and it&#