Michael Orenstein

Building Electric UI - a high performance, cross-platform user interface toolkit for realtime hardware visualisation and control.

[Michael Orenstein @KillcoDer]

@thsottiaux Let us define our 'temporal completion' requirements, for "I'll be back in 8 hours, have it done by then" (slow) and for "synchronous work" (fast).

[Michael Orenstein @KillcoDer]

@hd_nvim I’m so disappointed everyone is converging on “arbitrary bash, python, sed/awk/cat” instead of “line-number-less, ast based view and manipulation” for such obviously foundational tools such as reading and writing.

[Michael Orenstein @KillcoDer]

@hd_nvim I built a “delete, copy, cut and paste” tool so the agent didn’t output 2N lines to move M tokens. Adherence even with post tool call hooks that detect and scold wasteful lack of tool use was <5%. Primordial tool usage needs to be post trained in it seems…

[Michael Orenstein @KillcoDer]

@fabianstelzer @Crook1Halfway @ElonLikeThings No, everyone does speculative decoding, so you can generate as fast as you can prefill a correct prefix to add. You can pay the time cost once in reasoning then repeat yourself in output for significantly faster output.

[Michael Orenstein @KillcoDer]

@thsottiaux Can I have the ability to define a plugin with a LARK grammar instead of a JSON schema?

[Michael Orenstein @KillcoDer]

@thsottiaux Can you add the ability to have a different model + reasoning_effort for spawn_agents_on_csv too?

[Michael Orenstein @KillcoDer]

@thsottiaux When the model deletes a file and rewrites it, the codex UI should not show a 'full red diff' then 'full green diff' - it should show the actual diff.

[Michael Orenstein @KillcoDer]

@thsottiaux The model seems to have an obsession with 'minimum work, fastest time-to-patch, time pressure' and nothing I can say in AGENTS.md or otherwise can align it to 'optimal and correct implementation'.

[Michael Orenstein @KillcoDer]

@thsottiaux Generally the model is lazy, I struggle to believe these multi-hour-long task claims? I constantly find myself saying "yes continue" - even with ExecPlans, clear instructions "continuing until completion".

[Michael Orenstein @KillcoDer]

@thsottiaux While 'fast' mode is appreciated, for non-time-sensitive asynchronous tasks, a 'slow' mode would be well received. Half the token cost for 'flex' style processing.

[Michael Orenstein @KillcoDer]

@thsottiaux In medium-to-long sessions containing questions, when asked to do an action after a question has been answered, it will perform some file reads then answer the previously answered question instead of doing the action, often with identical responses.

[Michael Orenstein @KillcoDer]

I wrote a bit about constrained decoding optimisations for large language models.

https://t.co/77InoErPaB

[Michael Orenstein @KillcoDer]

@ctoofeverything @jarredsumner `this.constructor.constructor('console.log("oh look RCE", process.env)')()`

[Michael Orenstein @KillcoDer]

@danielkelly_io I would say it's a "human first product" more than it's an AI product. The time travel debugging is excellent. The 'inline watch variables' are ergonomic and significantly faster than console.log debugging. Their support has been responsive and solved bugs I reported within hours

[Michael Orenstein @KillcoDer]

@tmikov I can't believe 'tab' is the default for AI autocomplete, all that muscle memory of language-server suggestions gone, replaced with a 400ms+ wait to read the AI suggestion. I don't see how anyone can get into an 'output flow state', constantly having to read those suggestions.

[Michael Orenstein @KillcoDer]

@tmikov I'm convinced the only reason people think Claude Code is any good is because it _says_ its code is good (sorry, "production ready and literally perfect") and it calls _you_ a genius every third line. It's just weaponised sycophancy.

[Michael Orenstein @KillcoDer]

@AndaristRake Yeah I was working on a compile-time validated 'schema' library, branding number literals / enums with metadata. In 'accessor functions' it's reasonable to select them out with control flow when they're not the same primitive type, but I guess at runtime that's kind of nonsense.

[Michael Orenstein @KillcoDer]

@AndaristRake I have another one!

I'm trying to narrow a union of a literal string and a branded widened string type with a strict equality operator. Is this behaviour expected? (Narrow3 in the playground link)

https://t.co/TquXUKoRAq

[Michael Orenstein @KillcoDer]

@AndaristRake My assumption (before actually spending the day with checker.ts and the debugger, a humbling experience) was any context sensitive types would be deferred, then inferred 'left to right', as deep / nested as required.

[Michael Orenstein @KillcoDer]

@AndaristRake I think I found another one. I'm confused as to why having that dummy 'context' parameter seems to fix the inference, or annotating the ctx in the function before when the inference already gets it right. Does this one seem familiar to you?

https://t.co/EJxYPdqxHc

[Michael Orenstein @KillcoDer]

@AndaristRake Naively I would assume people avoid relying on this inference, given it produces `unknown`.