AI companies should disclose trade-offs in opaque reasoning, one post argues

A post urges caution about AI designs that could greatly reduce or eliminate reliance on chain-of-thought reasoning. The author says too little public information was available about Astra’s architecture and training changes for an informed scientific discussion of the trade-off between performance and monitorability. They call for companies to release relevant evidence and publish policies so the public can discuss those choices before companies pursue such designs. The author says they have helped write a proposal for this disclosure, but are unsure whether it would be enough to avoid the most concerning architectures.