UX in my Audi Q5 (2024) is terrible. With two phones in the car you never know which one is connected and whose google maps is being currently displayed. And then come the buttons designed with a contempt for a driver. I recently had to change a flat tire which is a story in itself. German engineering is soooo different these days.
I really wish that cars were legislated to have documented APIs/canbus. It would be great to be able to load an app that set my car up the way I like it, instead of having to change a bunch of settings every time I start it (EV mode on above 10% battery, eco mode accelerator mapping, single pedal driving all the way off. Every. Single. Time.)
Re Germans: I’m not sure it’s a new thing. I can remember trying to uninstall a seat in a 90s BMW and wondering how they had managed to make something that could be accomplished with 4 bolts into something so complex.
Yeah, it was probably patched. It could reason novel problems only of you ask it to pay attention to some particular detail a.k.a. handholding..
Same would happen with the the sheep and the wolf and the cabbage puzzle. If you l formulated similarly, there is a wolf and a cabbage without mentioning the sheep, it would summon up the sheep into existence at a random step. It was patched shortly after.
I’m not sure ‘patched’ is the right word here. Are you suggesting they edited the LLM weights to fix cabbage transportation and car wash question answering?
Absolutely not my area of expertise but giving it a few examples of what should be the expected answer in a fine-tuning step seems like a reasonable thing and I would expect it would "fix" it as in less likely to fall into the trap.
At the same time, I wouldn't be surprised if some of these would be "patched" via simply prompt rewrite, e.g. for the strawberry one they might just recognize the question and add some clarifying sentence to your prompt (or the system prompt) before letting it go to the inference step?
But I'm just thinking out loud, don't take it too seriously.
There was a Tacker Carlson interview with Sam Altman where Tacker probed him on Balaji's murder and Sam quickly got confused and disoriented. Make your own conclusions.
Saying he got confused and disoriented is and interesting conclusion to make of that interview. He was defensive from the onset and even went combative when Carlson continued down a specific line of questioning, which he allegedly did at the request of the victim's family.
> 2. Centralized venv storage — keep .venvs out of your project dirs
I do not like this. virtual environments have been always associated with projects and colocated with them. Moving .venv to centralized storage recreates conda philosophy which is very different from pip/uv approach.
In any case, I am using pixi now and like it a lot.
I like it. Enjoyed having it with Conda, was sorry when it was lost with uv. Been a pain to search my projects and have irrelevant results that I then have to filter. Or to remember to filter in the first place. The venvs may be associated with the projects, but they're just extraneous clutter unless there's actually something to be done directly on them, which is very rare.
One problem I have on my work machine is that it will do a blind backup of project directories. Useless .venv structure with thousands of files completely trashes the backup process. Having at least the flexibility to push the .venv to a cache location is useful. There was (is?) a uv issue about this similar use case (eg having a Dropbox/Onedrive monitored folder).
thats my biggest problem with uv, i liked the way pipenv did it much better. I want to be able to use find and recursive grep without worrying that libraries are in my project directory.
Pip doesn’t have any philosophy here. It doesn’t manage your virtualenv at all, and definitely doesn’t suggest installing dependencies into a working directory.
Putting the venv in the project repository is a mess; it mixes a bunch of third party code and artifacts into the current workspace. It also makes cleaning disk space a pain, since virtualenvs end up littered all over the place. And every time you “git clean” you have to bootstrap all over again.
Perhaps a flag to control this might be a good fit, but honestly, I always found uv’s workflow here super annoying.
Disagree—better to have space allocated in each project where they can be easily deleted at once. Rather than half hidden in your home folder somewhere with random names and forgotten about.
If for some rare reason you wanted to delete all venvs, a find command is easy enough to write.
Virtual environments have been always associated with projects in your use case I guess.
In my use case, they almost never are. Most people in my industry have 1-2 venvs that they use across all their projects, and uv forcing it into a single project directory made it quite inconvenient and unnecessary duplication of the same sets of libraries.
I dislike conda not because of the centralized venvs, but because it's bloated, poorly engineered, slow and inconvenient to use.
At the end of the day, this gives us choice. People can use uv or they can use fyn and have both use cases covered.
> and uv forcing it into a single project directory made it quite inconvenient and unnecessary duplication of the same sets of libraries.
Actually, uv intelligently uses hardlinks or reflinks
to avoid file duplication. On the surface, venvs in different projects are duplicate, but in reality they reference the same files in the uv's cache.
BTW, pixi does the same. And `pixi global` allows you to create global environments in central location if you prefer this workflow.
EDIT: I forgot to mention an elephant in the room. With agentic AI coding you do want all your dependencies to be under your project root. AI agents run in sandboxes and I do not want to give them extra permissions pocking around in my entire storage. I start an agent in the project root and all my code and .venv are there. This provides sense of locality to the agent. They only need to pock around under the project root and nowhere else.
This is actually the feature that initially drew me towards uv. I never have to worry about where venvs live while suffering literally zero downsides. It's blazing fast, uses minimal storage, and version conflicts are virtually impossible.
Do you only work on projects individually? Without project-specific environments I don’t know how you could share code with someone else without frequent breakages.
pixi is a general multi-languge, multi-platform package manager. I am using it now on my new macbook neo as a homebrew _replacement_. Yes, it goes beyond python and allows you to install git, jj, fzf, cmake, compilers, pandoc, and many more.
For python, pixi uses conda-forge and PyPI as package repos and relies on uv's rattler dependency resolver. pixi is as fast as uv (it uses fast code path from uv) but goes further beyond python wheels.
For detail see [0] or google it :-)
There is a good chunk of overlap but mise predominately pulls from github releases artifacts/assets and pixi uses conda packages. While mise can use conda packages, the mise-conda backend is still experimental. I don't think github releases or conda packages are better than the other, they both have tradeoffs.
Pixi is very python focused, it's both a tool manager and a library dependency manager (see uv/pip). Mise considered library dependency an anti-goal for a long time, while I don't see that on the website anymore I haven't seen any movement to go into that space.
Or by hours spent in the office
reply