3 ms·
> to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked That's likely coming from the 3:1 rati
by storus 8mo ago
> to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked
That's likely coming from the 3:1 ratio of linear to quadratic attention usage. The latest DeepSeek also suffers from it which the original R1 never exhibited.
- nl 8mo agoThere is no way you can diagnose this like that. Correlation isn't causation and much more likely is a common source of reinforcement training data.