Just want to say I really enjoyed your writing style, it’s just the right amount of funny/witty without distracting from the (very interesting!) ideas.
One solution I haven’t seen recommended much is to have a Claude instruction/skill that explicitly audits the diff of every upgrade, and force this manual audit as part of your upgrade workflow. This seems like it would work pretty reliably.
> Every week, somewhere between 1.2 and 3 million ChatGPT users, roughly the population of a small country, show signals of psychosis, mania, suicidal planning, or unhealthy emotional dependence on the model.
> Why is mental-health crisis not a gating category, the kind where the conversation stops, full stop, and the user is routed to a human?
Well, obviously “routing to a human” is not feasible at that scale. And cold exiting the conversation is probably worse for the user than answering carefully.
This conflict is popping up everywhere. There is a push by a lot of companies to allow agentic use of their services (and new companies explicitly offering "X for agents"), ignoring the fact that "agent" means the same thing as "bot" which we've spent the last couple of decades actively filtering out. Will be interesting to see how it plays out.
My working theory is that all models are approximately the same, and the variance in quality mostly depends on how long they think for.
So the trick is to always set to max, and then begin every task with “this is an extremely complex task, do not complete it without extensive deep thinking and research” or whatever.
You’re basically fighting a battle to make the model think more, against the defaults getting more and more nerfed to save costs.