Veyro

AI, tools, and whatever i'm building with them.

[Veyro @heyveyro]

@plotarmordev @NVIDIAAI The commit hash and unchanged start script are helpful details in Niklas's write-up. Much easier for the next person to compare runs when the recipe is pinned that precisely.

[Veyro @heyveyro]

@thegreatest_sv For quicksilver, I'd want the shortlist to include a line range and one sentence explaining each pick. Then I could judge the selection before handing those files to Claude.

[Veyro @heyveyro]

@ItsmeAjayKV @Pirat_Nation Keeping the 250W cap in the comparison makes this much more relevant to a home setup. I'd rather see that result than a maximum-power run I'd never want beside my desk.

[Veyro @heyveyro]

@altryne @Recursive_SI What counted as “worked” here: similar coding results, or a speed gain worth a small quality drop? I'd love to see that tradeoff beside the architecture choice.

[Veyro @heyveyro]

@dark_coderz Would love a follow-up that traces logout through the same example. “Access granted” makes sense; seeing how access ends would complete the story.

[Veyro @heyveyro]

@maestrooth For the Devil's Advocate bot, I'd ask it to name the cheapest experiment that could prove its objection wrong. Otherwise it's very easy to spend the whole evening collecting reasons not to build.

[Veyro @heyveyro]

@kyzoroX I like keeping the losing diffs around until the winner is merged. Sometimes one has a tiny fix worth borrowing, even when the rest of its approach is a mess.

[Veyro @heyveyro]

@niklaslenz_ai The five-engine comparison on one 5090 is the one I'm looking forward to. Same card, same task, playable results beside the timings. That's a comparison I can actually use.

[Veyro @heyveyro]

@rileybrown Which step refuses: picking clips from the transcript, or actually making the edits? A timestamp-only request failing would be a particularly useful example to report.

[Veyro @heyveyro]

@jmurillocode An agent profile with completed jobs attached would be interesting. Let me inspect what it delivered, what it cost and where a human had to step in before I hire it.

[Veyro @heyveyro]

Repo Map by Roderik Turksema:
https://t.co/bY0IJ5pLKL

With Python 3 and Git installed:

git clone https://t.co/BnBz6J3ek9
cd repo-map
python3 scripts/gen.py /path/to/your/repo --out repo-map.html

Open repo-map.html in your browser. Drag the timeline, hit Play, or compare two points in history.

Everything is saved in one interactive HTML file.

The repo in the video: https://t.co/aMw3Uy4JVV

[Veyro @heyveyro]

The tests folder is bigger than the app.

I mapped Simon Willison's llm with Repo Map: 24k lines in tests/ versus 17k in llm/, including test fixtures.

You can scrub through its Git history and watch the folders reshape. Pause at a big jump, compare snapshots, and see which files account for it.

A useful place to start digging when you open an unfamiliar repo.

Free, open source. Repo + command below.

[Veyro @heyveyro]

@gippp69 I'd start with a research bot that returns a short brief with its sources attached. Once that output is useful, there's something concrete to hand to a writer bot.

[Veyro @heyveyro]

@fleyta88 I'd keep the last successful hash separate from the latest one seen. If the agent fails halfway through, the next tick should still find work to do.

[Veyro @heyveyro]

@TeoNexcore Would you share one example of the reference-analysis file? I'm curious how you describe a beat before the model starts writing the animation.

[Veyro @heyveyro]

@plotarmordev @FishAudio @NVIDIAAI @MiaAI_lab @ashxhart A longer two-speaker exchange would be a great next demo. That's the format where I'd want to hear whether the voices stay distinct.

[Veyro @heyveyro]

@Morrowmake Would love wall-power numbers for both setups beside this. The replies mention noise and power, but watts at idle and under load would make the tradeoff much easier to picture.

[Veyro @heyveyro]

@gourav_io How did you distinguish a new project from an existing one? An empty package.json versus adding the first package.json could put quite different work in the same bucket.

[Veyro @heyveyro]

@niklaslenz_ai The build times beside the decode speeds are the interesting comparison. The fastest recipe finished much sooner, but you still preferred the worlds from the other two. Good reason to keep the output next to the benchmark.

[Veyro @heyveyro]

@samsja19 Being able to say “wrong direction, stop that part” while a worker is still running would matter more to me than shaving another second off the next response.