Post

1. Opus 4.1 was the Sonnet 3.7 for me, given how it "knew" how and when to use tools without me needing to bonk it via prompt every time like other models. 2. I need to do the set tasks, run it via LLMS, go for walk, then work in the afternoon.

Indeed. I've been quite surprised by Qwen 3.8 27B. I am seeing that for the things I use LLMs for - which is often scanning through academic publications or searching the web, and occasionally small coding projects like my personal website - this could be good enough.

I've been using small/cheap/local models for gates on PRs to ensure that my frontier co-operator sessions don't forget important things, and it's boss as hell. it's just critical that you approach their use with evidence/certifications

Thank you. Any time I’ve tried to search YouTube for any useful info on using LLMs it is at worst (and most often) just some AI slop written bullshit. At best, it is still not particularly useful info.

Do you have a place where you outline your current suggestions for what models to use where? Like for planning you'd use models in this order, for code review you'd use models in this order, for writing code, etc.

There’s a discord. Everyone there has slightly different workflows and it’s more about learning how *you* work and the ways to integrate the models. Generally I don’t use the big models unless I’ve hit a problem. I like multiple reviewers with different models/guidance.

If you have limited budget GLM is cheap and open weight. For me I don’t one shot anything. Plan feature, break into stages, review after each stage, test constantly. Most of the code I generate is tests. I’ve had implementations be more than 50% tests. Tedious for humans, robots do not care.

I know I'm on the discord server and I bug him and other folks with questions. I just thought this general information might be useful more broadly.