Ben Davis

I already did my reaction vid to GPT-5.6, but wanted to do another one going deeper into how I use it (and next gen models in general) 1. which reasoning level to use 2. "loops" (subagents) 3. computer use 4. issues with the model to watch out for

being pro-opensource doesn't mean you have to be anti-proprietary it's fine for proprietary options to exist, they usually invent the category and make it possible for the opensource option to emerge the only thing that matters is the opensource option is allowed to exist

Bad, bad, bad. Opposite of the reaction needed from the USG in light of the Chinese gov't deciding to go the other direction. Not exactly the type of stuff you expect from an admin that preaches about "accelerating innovation" and "free markets"

It's probably pretty obvious what I'm trying to do with this, will have a lot to talk about next week depending on how it all ends up going... github.com/davis7dotsh/sv…

I made a custom cli for svelte called "svelte-utils" U setup a daemon on one of ur machines, then use that as the "host" that hosts the svelte playground and has the check/best-practices cmds built into it. Then u can setup ur other machines to have the "svelte-utils ___" cmds run through that server. These are the exact same cmds that are part of the svelte mcp (which are great btw). The only reason I forked it to run it locally myself is because I set it up so that it will save the data from all the cmd runs (svelte components) Trying to start building up more and more data from my work, and also being really diligent about saving and backing up all my coding agent sessions. Highly recommend u do the same, I'm running a couple experiments rn that are going way better than expected...

@MichaelArnaldi U get summaries at least in pi that arnt great, im mostly working off of diffs at this point with that model

@davis7 Are you getting reasoning traces with GPT 5.6? I don't and it's pissing me off big time

When u have a model go off and build something u understand there's an absurd amount of upside in just reading some of the reasoning traces and diffs as it goes U don't even need to read much, I've found that just by looking at 10 lines I can instantly tell if it's doing it right or not. And now that the frontier has gotten absurdly steerable, u can just tell it what's going wrong and how to correct then it just kinda will

@Nateemerson It’s rly just my favorite parts of the big harnesses with minimal system prompt tbh

@davis7 I’m tempted to fresh install all of this and main it for a weekend just to really inject some psychosis

Using pi to orchestrate other coding agents is quickly becoming my favorite way to work gpt-5.6-sol on medium reasoning for the main thread I talk to, then firing off cc/codex/pi whenever I have things that make sense for them Computer use? codex. Planning/copy/architecture? fable. Long running grunt work? codex Quick changes/testing? pi. It's so good...

@mwolson that's a very good way of putting it actually

@davis7 Reminds me of an image diffusion model as well, turning static into a final copy with enough review passes/tests

In the process of building out my Pi extensions system I've been reading a ton more code Swung pretty hard to not bothering reading much for a couple weeks while trying to push the limits of 5.6 (lots of really good and really bad results), but what's been interesting is how massive of a difference just a couple extra prompts makes when u know what ur doing The functions for my file search extension (fs and rg) were "using effect v4", but also not really, it was the classic bastardized try catch hell but this time it's in an Effect.tryPromise inside a Effect.runPromiseExit in a normal fucking async function (which sucks, a lot) I just kinda read through all of it, copied one function into pi, told it why it sucked, told it to have a subagent look into effect's node platform package, and to use generator syntax, then after like 2 more turns got actually good code The end result is much better code than I would've hand written, but also much better code than the model would've written without direction Fable is the closest to solving this, a lot of why it feels magical is that without having to give this extra direction it's way closer to what I actually want (but still not fully there yet, although I do think eventually it will get there) I still think the pottery analogy is the best one rn: u start by having it make the full thing and end up with something decent, but not great. Then u slowly reshape it over and over again. The main difference is now ur reshaping with the model, which does big chunks instead of line by line

@_ChrisCovington opencode is better for most people. I just have very strong opinions on this stuff so benefit a lot from doing it hyper custom (which pi is way better for since it's more minimal out of the box)

@davis7 is there an advantage to pi over opencode for the main driver? other than pi has less batteries included, they both seem just as adaptable, am i missing something speicfic?

@jetpackjoe_ I have but it usually just doesn't matter. It's really just like 1 or two sentences I need to see to instantly tell which direction it's going and then I'm set. Although I do like this idea, haiku is way too dumb at this point, but luna might be great for it...

@davis7 Have you tried sicking Luna or Haiku on it to summarize?

And to be clear, there are a TON of things that I now have which I don't care about the code for, that are extremely useful, things I've made in the last week like this: - my token tracker which shows my sub limits + usage across all agents on all my machines - cli that wraps the YT api for my hermes agents to get live data - fork of vibe proxy to load balance my subs and hide my emails - svelte-utils cli which hosts a version of the svelte playground + check/best practices tool on my network that all my machines can use so I can easily review and collect data as my agents go - custom themes for my pi setup (not every extension needs to be giga robust) - shownotes - the software that we use while recording the podcast to send messages/links/notes - better git status cli that shows me open prs and repo info

The best setup I've found for this approach is zed + pi. Perfect for easily reviewing and directing changes as the model makes them. Then t3 code on a linux box for the massive long running jobs or little things I don't care about the code for

@theo It's currently unusable lol

You can get 5.6-sol to write great code, build great systems, it's an incredible model. But u have to control and direct it a lot...

These two generations are the best visual depiction of Fable vs 5.6-sol I've seen It's also the best way I can explain "big model smell". Big models don't vomit all of this extra llm-y garbage into the UI This same difference happens in the code the models write

Honestly all the labs (with one very obvious exception) have been great lately. OpenAI is always great and the Chinese labs are super pleasant to work with as well