The section of this that everyone will scroll past is the only part you can act on today: "None of this requires a new model. All of it is scaffolding, and scaffolding is where most of the available gain currently sits." Five things to engineer if you want a long session to hold: → externalize state before you need it: the context window is working memory, the disk is where the truth lives → write the oracle before the task: with no way for the model to check itself you have not built a loop, you have built a very long guess → bound the blast radius: small reversible commits, so a bad hour gets discarded instead of being load-bearing → force periodic re-grounding: make it re-read the original objective from disk, not summarize what it remembers doing → instrument for silence, not errors: the dangerous failure never throws, it produces confident, plausible wrong output for six hours Drift is never a dramatic turn. It is a hundred small reasonable-looking steps, each justified by the one before it. Bookmark so you do not lose it! Follow @neil_xbt for more AI content!