Chain of Thought is a scaling trap. the next wave is latent reasoning (Coconut / HRM / RecrusiveMAS)... but then we hit the black box wall. Where does BDH fit? [D]

Read a long piece on the future of LLM reasoning that makes a provocative claim: Chain of Thought is a useful hack but we've started to confuse a readable trace with the actual computation. All in all, "generating text is not the same as thinking." There are two practical problems here: Faithfulness: CoT style traces can decouple from what the model actually "did." u can get plausible steps with a wrong answer, or messy steps with a right answer (so the trace isnt a reliable audit trail) Systems cost: Autor