What's this about?
Chinese AI commentator OrangeAI argued that the last six months of model progress concentrated almost entirely in agentic coding (driven by RLVR), while other abilities stalled or even regressed — he says English writing has gotten harder to read, calls the Opus 4.7→5 line "failed experiments," and asks for new RL approaches beyond RLVR. dotey echoed a related practical read of which models are worth using for what.
What exactly is RLVR?
RLVR — reinforcement learning from verifiable rewards — trains a model by rewarding it when its output can be automatically checked as correct. Code that passes tests and math with known answers are perfect for it, which is why coding ability has shot up. OrangeAI's worry is the flip side: optimizing hard for what's checkable may be crowding out or degrading the fuzzy, unverifiable skills like natural writing.
What's the community saying?
- The regression complaint: heavy users echoing that newer models over-think, burn tokens, and write less naturally than earlier checkpoints — some saying Fable/older versions felt better before heavy RL.
- dotey's practical stack: he mostly uses Claude Fable 5 (reliable for real technical discussion), GPT-5.6 Sol (cheap grunt work), and Opus 4.6 (best for writing), finding Opus 5 caught between too-expensive and not-smart-enough.
- The bigger question mirrors Mollick's optimism from the other direction: is RLVR a dead end for general capability, and what replaces it?