The LLM reasoning models behave strikingly similar to superscalar out-of-order execution processors with decomposition, verification, and error correction steps.

Moreover, the LLM reasoning models are reliably consistent solving the same task with the same prompt using different variables. This can be demonstrated.

Not everything has to be deterministic to be useful. Nonetheless, understand how LLM models can be applied and be useful will help a lot of people to be less frustrated and spend less tokens.