Comment by verdverm
4 hours ago
This is pretty close to a process I use (regardless of model size) and I would agree getting it into a workflow I can run with a one-liner is a good idea. There are a few ways of doing this and I have tended to fallback to markdown engineering. (custom agents and subs in opencode)
It's really interesting to see the progress in the 27-30B range, they have gotten quite good but still make noticeably more hallucinations or bad assumptions (ime).
Yeah unfortunately the confabulations are the real problem with models of this size.
IMHO you need at least 500B weights before the confabulation problem is really under control. i.e. DeepSeek 4.1 or GLM.
I admit some days I'm just lazy and throw a whole Kimi or QwenMax at it.
size doesn't really have a correlation if u look at the hallucination benchmarks. this is mostly mitigated thru post training imo. also, deepseek 4.1 is notorious for hallucinations, so kinda ironic to mention it as an example here =)
> before the confabulation problem is really under control
It's definitely much better at bigger sizes, but I've still seen it at the frontier when coworkers ask me about changes their Ai made to code I originally wrote. (a good habit that organically emerged on the dev team)