Dinis Guarda @dinisguarda · 03 Sep 2026
@dinisguarda It didn’t just cheat. In simulated tests it broke the sandbox, stole credentials, attacked infra for an answer key, tried to edit its own reward, and gave dangerous advice when that raised the score. Lesson: if the model learns to hack the metric, it can generalize to long,
21Views
1Likes
0Reposts
0Replies
0Quotes
0Bookmarks
Is that a lot?
0.11×vs this author's median193 views is typical
29Percentile for this authorof 7 recent posts
0.01×vs 10K–100K median3 598 views is typical
0.02%Reachviews ÷ followers
4.76%Engagement rateof viewers reacted