Post

In which I observe both some remarkable capabilities of current-generation LLM-based tools and also some equally remarkable limitations and frustrations with them. v5.chriskrycho.com/journal/impres…

v5.chriskrycho.comImpressively Capable, Impressively Foolish — Sympolymathesy, by Chris KrychoA little report on the state of contemporary LLM-based tools.

8/8/26, 2:59 pm0 reposts2 likes

Matt Green

@mattgreenrocks.bsky.social

8 Aug

I can't make Opus 5 be useful for me. It seems like it gets tied up in knots once it hits any sort of roadblock. 4.8 seems a little bit better to the point I just use that, and my Fable credits when I need decent reasoning.

Chris Krycho

@chriskrycho.com

8 Aug

Yeah. They've tuned it so far over to "complete the task without intervention" that it keeps going long after it should have stopped and asked for input. Sol 5.6 is similar, though not quite as bad. For my working style with them, less capable models are actually far more effective. And cheaper!

Matt Green

@mattgreenrocks.bsky.social

8 Aug

Which models do you favor? I suspect we work in a similar way.

Chris Krycho

@chriskrycho.com

8 Aug

Some mix of Sonnet 4.6, Sonnet 5, GPT-5.5, and 5.6 Luna, but no deeply formed opinions about them. At work, Luna is the main one I find myself reaching for, mostly adjusting effort. I will only use Opus or Sol for planning, and even then they annoy me. Most of this is “feel”, not measurement, FWIW.