1,303 karma · joined July 18, 2017
Text is an imperfect distillation of the human experience, but it doesn’t surprise me that a machine trained on human text responds to human encouragement.
Printing presses are about exact duplication of writing. The internet is a higher layer printing press. LLMs substitute for (some of) the human effort of producing the text that a printing press or the internet can then replicate.
The thing that gets me is that the rationalists who were such outspoken proponents of prediction markets (look up "futarchy" and the many positive articles about it) failed to predict the astonishing amounts of corruption that have come along for the ride with them.
The review subagent has its own definition and gets invoked with a specific target, so the context is very focused on just rule enforcement and I don't have problems with it skipping rules.
The whole workflow is packaged up as a plugin, but you don't need that to get this approach to work. It should be sufficient to have the rules you want enforced written out somewhere, and to either kick off a focused review agent manually referencing them, or do something like I did and have it be a defined part of your workflow (depending on the automation level you want).
The form and maybe even the content of electronics in school may echo, sure, but the results sure don't.
Citation, please? On the original release they downgraded models silently to Opus 4.8 when they suspected it was being used for LLM development, but they stopped doing that. Now when you hit one of the guardrail subjects it downgrades to Opus 4.8 visibly in all cases.
I've never seen anything suggesting they're deliberately returning wrong answers. Maybe you're thinking of Gemini's anti-distillation tech?