alt Hacker NewsThis isn't true, there's a limit to how much feed-forward processes can correct themselves and well-poisoning of context is far too common.
Try using an LLM to rewrite an LLM output without the slop (vs asking for no slop to begin with) or sandboxed subagents that critique a parent's draft.
There is absolutely a step-function improvement in quality but: 1) not everyone wants to explode their cost by adding extra calls 2) this can't just be "trained in" to a system as obviously they have attempted this but the technique still provides an uplift.
If you can do adversarial agent review, or use an LLM to fix the slop and it works, then technically you already have an LLM trained which knows what to do. You can just use them to improve the other model. The issue is that we don't have these at the moment.