qudat

I really like this idea and have been experimenting with it over a week or so.

I think there’s an opportunity to use an AST diff system for code forges where you don’t present the user with line diffs in the UI — or at least not as the first diff the user sees.

I firmly believe code review should happen in your editor.

I'm still thinking this through but I was arguing this position to colleagues to some shock: LLM's are a race-to-the-bottom and frontier models will not be able to afford to work on coding specific models (or coding features at all) in the very near future.

27B is already really good at coding-specific tasks. Fundamentally, there is little innovation on the core architecture: LLMs are all designed essentially the same, with minor differences in how they are trained. They are all feed-forward multi-headed attention models; it doesn't matter if it's a 4B model or a 1T model, that's just scale.

Further, the frontier models cannot afford to innovate: they have to scale as quickly as possible to "beat out" their competition. The frontier models fundamentally will not create the next "attention is all you need" monumental jump in AI.

Frontier companies are stuck on scale with zero capacity to innovate. You cannot point capitalism at "basic science research" and expect any ROI. This is a known reality. Innovation is much more indirect and a "random walk" style of knowledge acquisition.

Finally, these LLMs are quite literally designed with a human-in-the-loop, and we do not give ourselves enough credit for how well we ourselves tool-call. We are doing a lot of heavy lifting to make these models useful and you cannot simply remove us from the equation without also removing ourselves from the training pipieline.

I built https://zmx.sh to make it easier to interact with your terminal sessions programmatically. 1 window = 1 session which might feel like a negative but it makes programmatic access easy and agents can use it just by pointing it at the zmx help command. Basically, an agent just needs 2 commands (run and write) for full control and the commands are synchronous so you don’t need to do any polling.

And a labeling action which requires `pull_request_target`: https://github.com/actions/labeler#create-workflow

These types of features are not worth it and need to be removed from the marketplace.

Literally the model “mythos” is being marketed towards finding these exact type of bugs used for exploitation. I really don’t understand the argument: are agents not good at findings memory management issues? What’s the gap?

My guess is it won’t be worth it to focus specifically on coding models once local small models work just as well or within range. That will naturally close the gap even more

Yep. I’m approaching the same problem from a different angle: writing code fast means you aren’t being thoughtful about the features you’re building. I started realizing that after I had kids and spent more time thinking about code than writing it and it really improved the quality of my work: https://bower.sh/thinking-slow-writing-fast

Nah. These agents are getting easier and easier to run local. Have you tried Qwen 3.6 27b? It’s insane what it can do compared to its size. Like 100% vibe small projects if you manage context properly.

These models are a race to the bottom just like compute.

I don’t understand: just use an agent to find all memory leaks and segfaults. I don’t get the argument if you are gonna vibe code anyway.

With unlimited tokens make it a lint rule or auto formatter.

We take a slightly different approach for https://pico.sh -- no automatic subscription, but we charge for an entire year. It's great for us because each sub is a year and if someone truly isn't using our services then it'll quietly drift into the background for the end-user.

The entire concept that we need to cater CLIs to agents at all should tell us how far away they are from being “junior devs” or “an intern” and I reject the premise.

A lack of structured output has never been a blocker for agents to work, that’s a traditional coding problem.

“Write good help text and error messages” is just good design which is self evident.

I don't get it, it looks like they are copying data to the sandbox filesystem why would that impact production data? Because the agent can re-upload the file to s3?

Wild example as it has been corrupted by VC money as well. I wouldn’t touch deno either.

Interesting analogy. Through the lens of a data driven design, all we are doing is taking data in shape A, transforming it to shape B, and then sending it somewhere else.

I think if you look purely at the numbers, the real reason TUIs are popular is claude code, everything else is background noise compared to it.

What originally got me excited to build TUIs was the concept of delivering apps over the wire via SSH. SSH apps resemble a browser in that way: no local installs required.

It's a major reason why I enjoy hacking on https://pico.sh -- deploying the TUI requires zero user involvement.

Fundamentally it cannot be much better than how well we can write the spec and then validate the results.

It’s always gonna be a multi shot process. And it can already write code good enough. That’s no longer the bottleneck.

Further, Qwen 27b is such an incredible masterpiece for coding and it can run on consumer hardware today. Anthropic/OpenAI are gonna give up on coding models very soon. There’s not gonna be any money in it when you can run your own local model for significantly cheaper.

Qwen27b is not SOTA but the value is insane. You can basically use it for small tasks and then route harder problems to opus or sonnet and boom you’ve said a lot of money.

Interesting idea. Tangentially related I’ve been using my local agent to interact with remote shells via zmx, described here: https://bower.sh/zmx-ai-portal

The use case is different but this article strikes some vague similarities around an agent API to remotely execute commands.

Thanks for the callout: we’ve been reimagining code forges by making them irrelevant with tools and tiny services like: https://pgit.pico.sh (static site generator for git) and https://pr.pico.sh (pastebin for git collab)

They are still a WIP but it’s on our roadmap to continue to improve.

how do you provide the base url?

Agreed! Over at https://pico.sh we are chugging along and having a blast. Profitable but at a scale that is manageable for us. Cheers

I didn’t think pi supported local models?

I just want sonnet 4.5 but I also want it to be fast which is trickier

I'm guessing part of it is design philosophy: mawww doesn't want splits embedded in the editor, for example.

Totally agree. Nextjs is a vehicle to sell their PaaS, every other feature is a coincidence.

What’s worse is vercel corrupted the react devs and convinced them that RSC was a good idea. It’s not like react was strictly in good hands at Facebook but at least the team there were good shepherds and trying to foster the ecosystem.

I’ve been experimenting with letting a local agent manipulate my remote servers using https://bower.sh/zmx-ai-portal

What’s exciting is how simple cli tools can be so impactful to dev workflows

Hello. I couldn't find a great way to interact with remote terminal sessions, especially ones that required multiple hops through SSH / containers so I adapted zmx to integrate with local code agents.

- https://zmx.sh/

- https://github.com/neurosnap/zmx

- https://youtu.be/CV3skPYHP4Q

Happy to answer questions

This is the key. These frontier model companies are funneling all of their time and resources into scaling, how could they possibly be researching the next phase of AI? Once scaling hits the limit the money is gonna dry up.

AGI is not gonna come from these companies

Check out: https://github.com/neurosnap/zmx

It uses libghostty to restore scrollback history and terminal state.

A similar post with more emphasis on validating changes: https://bower.sh/thinking-slow-writing-fast

Hard agree.