Post #1490302

2026-03-05 18:05 UTC

@bilboed @khaled Thanks. No, I haven't played with those yet. Although, most concerns raised equally apply to Open Source models I think.

Replies (2)

@sri@mastodon.social 2026-03-05 18:07

@behdad @bilboed @khaled I think once you change LLMs things get dicey because you have to add additional prompts to Qwen to make the same output. In my own experiments changing LLMs is like changing out to a new human.

Open ##1490304
@bilboed@bilboed.tech 2026-03-06 06:08

@behdad @khaled well... I was thinking about the smaller weights ones too, where the likelihood of it spurting out something it saw in training verbatim is drastically lower (if not impossible). As opposed to something it "learnt" the same way humans do (I.e. the high-level concept). I'm not touching Claude (or other frontier model services) for coding because of that risk (along with other ethical reasons). The smaller ones won't one-shot answers as much.... but their reasoning and handling of information you provide in context is getting better week after week.

Open ##1490310