Basil

[Basil @BasilMakesRagu]

Spent some quality time with GPT 5.6 Sol today. What a breath of fresh air. I miss this.

[Basil @BasilMakesRagu]

@natemcgrady hunter2 is not bad. You should give it a shot

[Basil @BasilMakesRagu]

I've been sitting on a bunch of Context Bonsai changes because I've been heavily focused on day job stuff. I really need to get my automated maintenance system just up and running on a cron.

[Basil @BasilMakesRagu]

@GeoffreyHuntley The compaction use case is similar to what I'm talking about with Context Bonsai. It lets the LLM call a tool that replaces chunks of its own context with a summary and keywords, breaking some cache points, but also then having every step after that have fewer tokens in context

[Basil @BasilMakesRagu]

@theo Don't pull the viewport to the bottom of the scroll when new content arrives in the buffer. Seriously - I can't believe you even let that one through.

[Basil @BasilMakesRagu]

I've been using Claude Code's fork subagent for some specific tasks recently. I've found that it frequently does more than just what it was asked. It will continue working as though it wasn't the fork. You end up with mystery commits and edited files in your codebase.

[Basil @BasilMakesRagu]

@theo I didn't say that. I don't know why you think it would be worse, or why you couldn't provide both. It's also possible to provide the tokens/sec affordances without token by token streaming. I do know the way t3 does token-by-token streaming is the absolute worst version I've seen

[Basil @BasilMakesRagu]

Sonnet 5 > Opus 5

[Basil @BasilMakesRagu]

@mitchellh This is the kind of thing that was ingrained into me because I started doing work with sonnet 2.5. Funny that Opus 5 seems to have regressed on this dimension and does the obviously wrong thing more than the prior generation

[Basil @BasilMakesRagu]

@endingwithali Because I want to read the PR in the actual platform others will see it on. I don't trust the LLM's .md draft

[Basil @BasilMakesRagu]

@mattpocockuk I usually start with full LLM-driven for the PoC, do some fine-tuning, then ask the LLM to convert what it can to scripts and leave the LLM to judgement tasks. And then the LLM tries to make everything a script and I remember how bad they are at judging :)

[Basil @BasilMakesRagu]

@JeremyKrak @QuinnyPig @theo You sound like someone who has never built skills/plug-ins/workflows quick feedback on how your changes affect the model's behavior is essential

[Basil @BasilMakesRagu]

@theo I was just shocked at how sensitive you are to ux bugs in other products, yet yours has embarrassing flaws. And I *know* you know about them because your videos show you do the exact action that T3 fails at (reading the buffer while the LLM streams the next response)

[Basil @BasilMakesRagu]

@QuinnyPig Oh, goodness. There's a ton of people in Seattle who pay attention to your work

[Basil @BasilMakesRagu]

@mattpocockuk That's welcome. I didn't actually know what it used CONTEXT.md for. Every time it offered to make one, I thought it was ridiculous since we already have tons of other docs in the system and a README.md. The new name makes it clearer

[Basil @BasilMakesRagu]

I *really* hope the next generation of LLMs fix the "Y not X" problem it's so tiring to chase that down and the LLM isn't great at finding it during adversarial review.

[Basil @BasilMakesRagu]

We're at the point where LLM models are "smart enough" that we can use them to build things we wouldn't have imagined before. Instead of making them smarter, I think we'd get much better results if we focus on acheiving this level of intelligence at lower cost and higher speeds.