Post #1619362
2026-02-22 21:51 UTC
@ju so I've experimented with FIM in the past (cue December). Copilot auto-completed them in an Ok-ish fashion. It was a hit or a miss. I tried small local models to see if it works (spoiler alert, 1.5B qwen-coder model on my local machine just started nonsensical completion every time I tried it). I don't have a very powerful machine. So I can't run more powerful models, but the fact that copilot didn't work that well for me, speaks enough. [1/]
Replies (1)
-
@innocentzero@mastodon.social 2026-02-22 21:54
@ju I'm definitely not into agentic coding (I want to be fast, not lazy). Maybe I'm not very good and very fast at describing what I want from the LLM, but FIM should infer context well enough because my code is decently structured. If it takes so much input to create such mediocre output, then from an economic standpoint alone, it's not favourable, ethical concerns aside. Not to mention, right now all the companies are riding the VC funding wave. They still haven't proven it's sustainable. [2/]