as incredible as they are, let’s not forget that LLMs are, ultimately, token-generation machines. two places where this really shows: writing: - they can, and will, write incessantly - they’re built for continuation; extend their budgets, and they’ll go on forever - it’s our job to moderate them, and to know when to stop; aka judgement the beaten path: - everything in context conditions what is generated next - as a session unfolds, the conditioning accumulates, and the space of plausible continuations narrows - this makes it surprisingly difficult to turn a mature session against itself - major corrections, adversarial audits, radically different framings; all become harder to access - the model gets “wedged” in a kind of “basin”: increasingly biased toward a small fan-out of future trajectories - effectively, it becomes “obsessed” with just a region of the solution space - you can sometimes unwedge it by substantially perturbing its context, but at that point you’re better off starting fresh than fighting an uphill battle (literally!) in other words: long sessions buy coherence, but they impose major path dependence once a session takes a turn you don’t like, scrap it. or rewind it to a known-good state, and proceed from there.