I don't have a prompt per-say.. but recently I have managed to ask certain questions of both openai o1/o3 and claude extended thinking 3.7 that have spiraled way out of control. A simple high-level architecture question with an emphasis on do not produce code lets just talk thru this yields nearly 1,000 lines of SQL. Once the conversation/context gets quite long it is more likely to occur, in my experience.
The only model I've seen so far that doesn't end up going crazy with long contexts with Gemini 2.5 pro, but tbf I haven't gone past 700-750k total tokens so maybe as it starts to approach the limit (1.05M) things get hairy?
I tried Gemini Pro 2.5 with a lot of context: all the documentation for a system and several papers of interest, then asking it to use the system tom implement the proposed solution in the papers. The total context was over 500k words, so with usual estimates probably over 700k tokens.
The answers started out ok, but fairly quickly it seemed to loose track of the mid-stuff in the documentation, insisting on using one concept instead of another even when I explicitly told it not to. Full attention on 1M context is not really feasible (I don't believe that Google actually stores upwards of 1T of data just for my query), and there are various ways LLMs use selective attention. I'm not sure if Google has published anything on how they do it?
