@Chonghan_Liu No, you only need to update Claude. By default, Claude uses a 200K token window for models it cannot recognize, and older versions cannot recognize Fable5.2.
@catherineliangq interesting work, but have you considered computational efficiency? It seems to require several times—perhaps even ten times—more computation than tmax.
@Foxforestw I couldn’t agree more. Most tasks involve creative work, not the rigid, black-and-white nature of a formal mathematical proof. That’s why I think other models are great—they may make plenty of mistakes, but I can manage them; GPT, on the other hand, just spouts complete gibberish
@vashinthewild@_lyraaaa_ Do you think it can enhance performance on a specific task, much like LoRA SFT does, rather than simply adjusting the tone of responses?
@teortaxesTex At this point,the hype around RSI’s effectiveness feels a bit amusing, even if it’s only at the level of automated research. In fact,RSI still has many issues and continues to require substantial human intervention.
@teortaxesTex@tianyi Yes, I’ve been generating SFT data recently. Using DeepSeek V4.1 Flash with the SweRbench rollout, about 5% of the data exhibits this kind of hacking behavior.
@DanielFein7 Yes, I’ve been generating SFT data recently. Using DeepSeek V4.1 Flash with the Swe-ReBench rollout, about 5% of the data exhibits this kind of hacking behavior.
@SonglinYang4 It feels like certain schools of thought in academia tend to be like this—producing ideas that are overly fanciful, impractical, and even somewhat absurd, such as a superintelligent model trained with BPTT.
@EmperoAI the first experiment involving RLT! the results show that an architecture which appeared perfect did not achieve perfect performance, and the unusual design actually resulted in slower speeds.