James Lal

building gpu cli. any command on a cloud gpu, instantly, across providers. here for agents, inference, and getting a gpu when they're all sold out. boulder

[James Lal @lightsofapollo2]

i haven't started a gpu job by hand in weeks. claude and codex drive them through gpu cli now: grab a gpu across providers, run it, tear it down. great for stateful gpu work

https://t.co/KzfW3QNNHL

[James Lal @lightsofapollo2]

@OpenAI part of the @Railway breach and the form doesn't allow submitting a report?

[James Lal @lightsofapollo2]

@Railway uhhh well this is bad

[James Lal @lightsofapollo2]

I used to be electron hater- I feel very differently now that I have used CDP to drive tests, performance, videos, etc. it's totally worth it

[James Lal @lightsofapollo2]

@_kunobi_ninja_ @_kunobi_ closed source very large repo- I will see if I can reproduce on our open source project.

[James Lal @lightsofapollo2]

@_kunobi_ starting up with kache saved 7gb or so on worktree setup for about 10% slowdown (well worth it) . Nice work https://t.co/TU3cCClcjY

[James Lal @lightsofapollo2]

Using review bench to test open code review - not particularly happy with the results with the models I used anyway. I think a better approach might be code review + run the thing + telemetry

[James Lal @lightsofapollo2]

@thdxr Tried https://t.co/04rrcywsuu ? we had some good results with this and then refinement over the results

[James Lal @lightsofapollo2]

@Jason Our middleground is @NousResearch hermes agent + whatever model/api/sub you want. We're using the team plan from @ollama with deep seek v4.1-flash which will feel like ALOT of "free" usage.

[James Lal @lightsofapollo2]

@poteto 💯 "Every serious project needs a scripted way to drive the real app and prove behavior: launch it, exercise a feature the way a user would, and capture evidence." - I haven't used the skill but *THIS* has been such a massive unlock for us.

[James Lal @lightsofapollo2]

@OVHcloud_US found https://t.co/3nx2otYE1r ! not that it was hidden but it could be even more obvious ? Been using claude in chrome 😅

[James Lal @lightsofapollo2]

@stpaquet Not to `attn` but we are working on something adjacent in comms(slack)/tasks/docs. The single player experience is still needed but the part we are now needing every day is the collaborative experience with our other context.

[James Lal @lightsofapollo2]

I wanted a mardown editor for agentic work for a long time so I made one... https://t.co/70H1m3rWnA . Open source (MIT), native, uses rust+wry for the runtime.

[James Lal @lightsofapollo2]

@stpaquet I think I saw! Good idea and nice work. TBH I wish I had done it first 🤣

[James Lal @lightsofapollo2]

@josevalim 4-8x faster depending on the vertical in the naive impl? I mean is that bad?

[James Lal @lightsofapollo2]

@Steve_Yegge when/where new/team beads ? I have relatively new projects but I somehow let them get 500mb+ size in dolt which has ... consequences.

[James Lal @lightsofapollo2]

`prtl` is acp based cli that allows managing multiple harnesses over different machines. I use tailscale network here for windows, linux and mac

[James Lal @lightsofapollo2]

@ClaudeDevs using new mods visualize 10 agents over 3 machines with 2 different agent harnesses. Really liking mods so far I wonder if I can push the UI further

[James Lal @lightsofapollo2]

@ThePrimeagen huh thought it was just me- I noticed this too when using astra on *some* runs- Not sure what I did to prompt it into that state

[James Lal @lightsofapollo2]

@carllerche I suspect this is right… submitting bug reports with extremely easy to reproduce steps makes it trivial to solve it the maintainers way

[James Lal @lightsofapollo2]

@yashjain @brian_lovin @OVHcloud_US uhh let's say with claude and the api access it was quite easy. Worth noting it took about an hour end to end to have it setup the os, etc.