illulachy

Solopreneur | Lifestyle | Tao

[illulachy @illulachy]

@claudeai for the four-layer boards that were ready for manufacture in under a day, what checked the routing before that point, a design-rule check the model ran itself or a person reviewing the layout?

[illulachy @illulachy]

@jkeatn is the 700,000 strokes the output of one program, or many programs layered into one painting? curious how the run was split once it's painting in code.

[illulachy @illulachy]

@markshi07 @illulachy_ @SkildAI @AIQuanting yeah, that's the check that matters. skild's post only gives the headline up to 13x and ~3x less memory, nothing broken out by scene type.

[illulachy @illulachy]

@0xSero what goes in the final scope beyond the mockups so the last model can work without ever hearing the voice chat? the journey trees, your live reactions, open decisions, or all of it?

[illulachy @illulachy]

@rauchg if the agent's session has already ended when deployment.succeeded fires, where does the webhook notification land, and how does the agent pick it up when a new session starts?

[illulachy @illulachy]

@zhizdev does the benchmark score each principle on its own, like hidden characters and actions vs awkward word choice, or roll them into one overall slop score? curious which principle separates ai from human writing most.

[illulachy @illulachy]

@OpenAI for an agentic sampling job that's mid-tool-call when the decision to intervene comes in, what happens to that call inside the 35 second window? does it finish, or get cut with the model output?

[illulachy @illulachy]

@antigravity with the terminal and git panes hidden in the knowledge work layout, do permission prompts for terminal and git actions still surface, or does the agent just wait on them somewhere you can't see?

[illulachy @illulachy]

skild open-sourced a physics solver they say is up to 13x faster than physx...

steel is a rigid-body solver for robot manipulation, runs inside nvidia's newton, used from isaac lab

caveat: the 13x and ~3x less gpu memory are skild's own claim

via @SkildAI @AIQuanting

[Skild AI @SkildAI]

Today we’re open-sourcing Steel, our internal physics solver built with @NVIDIA.

It runs simulations up to 13× faster than PhysX while using ~3× less GPU memory.

Steel unlocks scaleable simulation for manipulation, for everyone.

[illulachy @illulachy]

@felixrieseberg how do you tell a scenario is ambiguous enough to need fable instead of opus 5.5? is it that you can't write the spec yet, or that you'd expect several valid answers?

[illulachy @illulachy]

@alexalbert__ when it lists the rest for review, what usually lands in that pile, judgment calls or things it just couldn't verify?

[illulachy @illulachy]

@theo curious where the 2 minutes went. finding the add project button, or the path field not taking what was on your clipboard?

[illulachy @illulachy]

@UnrealEngine when the agent works in editor through unreal mcp, what can it actually reach: scene actors, assets, blueprints? and is that scope set by the toolset or by what the user allows?

[illulachy @illulachy]

@LimestoneHQ in the 90% of choices, did the agent never open the code that contradicted the file, or did it see both and still go with the file? curious which one the one sentence fixes

[AdiiX @adiix_official]

https://t.co/tZJw0URrhg

[steven @Tu7uruu]

Just dropped more ASR model on HF!

• Tiny (27M params) monolingual ASR for Arabic, Chinese, Japanese, Korean, Ukrainian & Vietnamese
• Beats Whisper Tiny with 48% lower error rates
• Outperforms 9× larger Whisper Small, rivals 28× bigger Whisper Medium
• Runs 5–15× faster on-device than Whisper
• Open-source, edge-ready, high-quality human-, pseudo-, synthetic-data mix