The output undoubtedly improves when looping LLM output back into the model at inference-time, but there is a limit to this and it is still bounded by the acume

The output undoubtedly improves when looping LLM output back into the model at inference-time, but there is a limit to this and it is still bounded by the acumen of the underlying model. You cannot just recurse these models with tooling and compute to e.g. solve new physics.