Kayvane Shakerifar

@manthanguptaa Basically an RLM with REPL?

LLMs are really good at writing code, so why are we giving them 100 different tools instead of just giving them code execution? This idea came up in a conversation, and it just made sense and felt like it was right in front. It feels like a much cleaner way to structure things. Instead of turning the context window into a dumping ground of raw outputs, you let the model write code, process the data, and return only what actually matters. You are not just making things cleaner, you are likely saving a lot of tokens as well. The model only sees the results it needs instead of parsing through noise. This becomes even more obvious with things like web search or scraping. HTML is mostly garbage, and pushing all of it into the context is just inefficient. Filtering it through code first makes far more sense. I haven’t tested this deeply yet, but it’s interesting to see Anthropic leaning into a similar direction. Feels like a strong validation of the idea. Intuitively, this should improve latency, cost, and accuracy by turning the LLM into more of a controller than a processor.

This bug also not great UX

@jxnlco - I’ve seen you asking around for feedback on codex. I’ve been using it alongside CC for about 6 weeks now and it’s now my preferred tool of the 2. Here is some feedback

This is slightly different to subagents because you can see and interact with the sub-worktrees / agents in a different thread. Whereas a pure subagent within the same thread wouldn’t really allow for the same kind of refinement as you could achieve if you spun up a new thread

Simple idea but often in a plan you might come up with 2 different paths to implement an idea, if you could split into worktrees from the main thread, and see the outcomes in the child threads, you could prompt the main thread on what you want to keep / import

I’m not sure this is possible through a skill but what I’d really want is a workflow where I can branch new worktrees out from a thread then ‘reduce’ the best one / combination of features into the main thread. All the primitives are there for it to work and it would be amazing.

Aside from the archive thread action would it be useful to also have an archive & remove worktree action on a thread

I saw an attempt at worktree cleanup recently added but that was kind of brute force crazy, it removed all worktrees including the ones I was actively working on, thankfully it was easy to restore them.

Worktree bloat can happen quite quickly specially with automations, I have loads of daily actions that kick off new worktrees for short lived tasks, would be a nice small QOL if you can set up the automation with an expected TTL for the worktree

@jxnlco Pydantic

what are the ai, oss libraries like this in 2026? please nominate them below

@ryanvogel Modal

going to benchmark sandboxes i have - cloudflare - vercel - daytona - e2b.dev - exe.dev - sprites anyone else i should try

@samuelcolvin @mitsuhiko I think the excellent name choice is going over everyone’s head

Fuck it, a bit early but here goes: Monty: a new python implementation, from scratch, in rust, for LLMs to run code without host access. Startup time measured in single digit microseconds, not seconds. @mitsuhiko here's another sandbox/not-sandbox to be snarky about 😜 Thanks @threepointone @dsp_ (inadvertently) for the idea. github.com/pydantic/monty

@chrisalbon The codex app has a really nice diff section where you can comment on the files themselves in the app and push those comments back to codex. It’s feels similar to the IDE experience but focused. Switching to the IDE from the app is an integrated 1-click. I’m a big fan

Just saw the new codex app. I have just been using vscode with Claude in the in-app terminal. But all these new apps (conductor, codex, cursor) have this new paradigm where you basically aren’t looking at the code at all. Has everyone switched to this new paradigm? Is it hype?

@mervenoyann @pcuenq I’ve wanted this for so long, built one myself but will check it out 🙌🏼

we just shipped daggr, a new library to build complex AI workflows 🤗 it's a breeze to code and debug apps, and visualize the workflow itself 🙌🏻 try it out and let us know what you think!

@charles_irl Dflash too? 👀

when you finally get specdec, asynchronous scheduling, and flash attention 4 to work together

@vamsibatchuk @GoogleAIStudio This is so so cool, well done 💪🏼🙌🏼

Ever wondered where words come from? 🗺️ built this app on @GoogleAIStudio called 'Wanderword' to map the evolution of language through time and space. It uses Gemini to trace linguistic roots and D3.js to animate the geographic migration of words through history. …derword-141284551734.us-west1.run.app

@jxnlco Waiting for ruff to release custom rules feature so Incan do this in python - until then looking at semgrep custom ast rules

ai coding - you could be writing more lint rules One of my big takeaways from working with Vignesh is that, while oftentimes I will add style preferences to the agent files, Vignesh, on the other hand, will actually have the AI write a new ESLint rule and just turn on pre-commit hooks. I'm curious if folks are doing the same. What do you do?

@jxnlco 100% agree

codex plans get 10x better if the word 'spec' is in the input not sure why

@deanimatedmonk @Vignesh_ey @rive_app Companion iOS app that turns your plants into tamagochis, they can prompt you to feed them!

@Vignesh_ey @rive_app I actually thought about OLEDs early on, but they’re very limited compared to what I can do with Rive on the web. You’d need a different, much lighter visual language for a pot display. Still, the idea of a “pet plant” feels promising, worth exploring.

Made this plant persona (I call him Tiny) using @rive_app 's data binding, GPT and some hardware (ESP32 and a few sensors like touch, mositure, hum, light, temp). Most of the logic is on my client side rn. But will try and see how far I can take the logic part with scripting (They keep on building so fast!)

@adamdotdev The usage limits on codex $200 plan are waaay more generous than claude code

It’s wild to me how many people don’t realize how much smarter 5.2 is than opus, it’s not even close. My guess is this is entirely based on the Claude max plan token subsidy