antoineMoPa

> Because this is very much a part of what ML engineers do.

Do you realize most people aren't ML engineers?

I even built tiny neural nets myself and never played with the weights directly post-training. The original post was eye-opening for that reason. I knew it was possible to visualize weights, but as far as I understood, it was not really something I could edit with any success. Now, it seems possible without a phd just by asking another LLM. This is fun!

I don't get it. Because python was used, it's supposed to not be cool? IMHO, working with raw weights with Python is still working with raw weights.

I kind of fear this would cause agents to get lost in previous outdated context, ultimately wasting time and tokens.

Neat, Thanks!

I've been "coding" a couple of apps with agents + rust recently and I realize how terribly huge the target/ folder can grow. Then, you run cargo clean, get a clean disk but compilation takes forever again. I routinely run out of disk space.

Love the experience in general except that part.

Not yet fully baked, but I am planning to add a woodworking mode to claydash with natural cutting UI - For now I have a couple of wood materials and there is the ability to extrude + cut faces à la sketch up, which should work for simple projects.

The local version has a MCP server.

see https://claydash.com/

I tried to embody this ideal when I managed a team, but realized shielding too much meant less connection of members with the wider org and too much dependency on me for context and such. Overall, this meant we were a lower value team for the org.

I built some tools that lets me view diffs and summon agents to execute changes. You can probably vibe code that in a few minutes.

It's like corporate lingo for computers.

That's such a simplification. Canada has a respectful relationship with EU whereas the leader of the neighbors to the south has repeatedly hinted at annexation of Canada as a 51th state. How can you negotiate trade deals respectfully with this neighbor?

Currently the 2 major providers have outages, maybe we should not silo our knowledge work behind statistical text models?

I mean, yes, but some governments may be tied to using European alternatives by regulations and some individuals may prefer not relying on US / China models for various reasons (could be political / personal preferences).

egui is very neat indeed. With agentic coding, Rust UI dev is as easy as electron dev without the bloat. Electron was relevant when we had to manually code stuff.

I don't find myself needing to open a code editor anymore. I ask agents to code and review using some custom review tooling I had agents build. I bother less and less about code details.

I prototyped some rather complex internal library with it, taking advantage of the free period. It was cool while it lasted, but I wouldn't pay for it. I'd be able to do the same thing with cheaper models and more time + involvement from me. I'd work a bit longer, but would also have more insights about how the project works. It would have cost ~160$ for one day. Hard to justify using it as a daily driver. I did appreciate its concision + willingness to go an extra mile to make sure things work, but all of this could potentially be ported to smaller models with the right training setup I assume.

In my experience, this is a more frequent issue. At least 1 time per day, I hit play hoping to play Spotify in a browser tab or some radio tab and for some reason it opens Apple Music instead. Sometimes it could be an issue on my side, ex: the tab is dead or not even opened.

But whatever, the experience is bad: I have to wait for the Music application startup time, then click the context menu and select "Quit Apple Music". It feels like being forced to watch a product ad. Opening Apple Music is never what I want. Imagine if pressing shift opened TextEdit by default, that would be silly. Or doing CMD-v where you can't paste would automatically pop up some random app.

I feel like no machine response is a correct UX pattern in this case. The absence of sound playback would indicate to me that I need to do something else to play sound.

There are multiple different sources that the user might want to start playing. Browser tab A/B/C (example: web radios), a music application or music service website in a tab that's not even opened yet (eg: spotify), the last video tab they opened (ex: youtube).

Whatever is the last thing that was paused should play IMHO. If nothing was paused, it should do nothing. Else, you open a pandora's box of possibly wrong choices that the user then has to close.

It cannot assume which media player I want to use, so the best course of action is to do nothing.

Skipping most reviews and waiting for disasters to happen.

Abusing the words "canonical" and "normalized".

How am I expected to (vibe) code without spotify? Taking the afternoon off.

magit is pretty useful! In particular, I use M-x magit-blame for blaming a file line by line and jumping to previous versions, M-x magit-status to stage individual files, M-x magit-log-buffer-file to know the file's history. Very helpful when tracking down bugs.

I get the best output when I work on a single project or 2 entirely different projects at the same time. Example: work + a side quest or prototype where agents cannot overwrite each other's code. I can add a third stream of real world productive tasks, example: do the dishes.

I think in terms of UI, codex has become superior probably due to its sandbox concept. It can complete tasks reliably without intervention whereas claude would have stalled asking permissions (unless you run it in --dangerously-skip-permissions, which I won't do). I don't know about you, but sitting in front of my laptop just to continuously press enter is depressing, I'd rather do other tasks in the meantime, which is exactly what I can do with codex.

So combined with the issues claude has been experiencing in the last couple of weeks (see https://news.ycombinator.com/item?id=47660925) - codex has become a better option. Claude has been unable to complete simple tasks and got lost in 5+ minutes of exploration too many times and I just called quits.

Claude has become pretty slow and in the last couple of weeks, so I switched to codex. Codex code can be quite sloppy, so it's worth doing multiple review passes.