Do your capabilities homework
RobinHa argues On-Policy Self-Distillation (OPSD) as a safer alternative to RLVR/GRPO: reuse the same frozen model as a 'fake teacher' by giving it extra information in the prompt (hints, worked solutions) and distill the student on-policy so only the erroneous tokens are corrected (example given with a 32 * 5 arithmetic mistake).
Commenters supply technical pushback and extensions: romeo focuses on systems-level verification (Amodo Design’s 24-hour provable-wipe result leaving >100 TB unwiped, proposals to keep SSDs outside inference racks to cut restore time to ~43 minutes, and ideas for verifier-replay or memory-verifier servers over multi-TB/s links), and raises ZKP reproducibility and circuit-safety concerns (cites arXiv 2402.15293 and a 2022 Trail of Bits critique). Samuel notes routing mechanisms (like Kimi K3’s residual-stream routing) predate Kimi K3 (Pagliardini 2024, Heddes 2025). Brendan Long reports that controlling chain-of-thought formatting with realistic discrete prompts is difficult on GPT-OSS-20B (~2% compliance), though soft prompts can work. Caleb Biddulph is supportive and suggests combining OPSD with limits on hint length/divergence (links to SPAR/DemoPSD ideas). Overall the discussion treats OPSD as a promising safety-aligned capability method but flags empirical slop, deployment challenges, and risks when labs mix OPSD with RLVR or ignore hardware/verification realities.
OPSD claimed benefits: reduces 'overselling' to an LLM-judge, pushes student toward specified behaviors (including fuzzy safety norms) without an explicit reward signal, and (author estimate) is ~100x safer than standard RLVR, though not perfect and vulnerable to 'Impossible Knowledge' issues.Open reader