GPUs/XPUs are small and solid state so it’s only really price that’s a huge liking factor.

And the disinclination of these companies to push the weights of their cutting edge models into people’s cars where they can be dumped.

Yeah. Every major self driving model that I’m aware of is fully e2e at this point. Going from fused sensor output to control+debug vectors.

This is more generalised.

But also since there’s a huge volume of data it’s too expensive to just keep scaling compute up (per car overhead) so there are necessary tricks involved.

I do think having a large model that can do this means that a small specialised model could be distilled form it though. Which is probably the most feasible path to production IMO.

This seems like it would be way too dangerous for all sorts of security reasons.

There’s an opt out so I’m blade for now.

I know it sa toy, but if someone gets the bright idea to do this for real: please learn from npm and dont duplicate packages for every clone of a project. Share them between projects like every sane package manager.

> Share them between projects like every sane package manager.

I'd argue the other extreme is less sane. Pip installs everything system wide and you can't have different versions of the same package without venv.

Even early NPM was much much better than the pre-uv Python situation. (TBH even with uv it still feels hacky at times.)

Before uv there was poetry, and before poetry there was pipenv, before pipenv there was virtualenv. I wish people would stop portraying Python tooling completely wrong like that. uv is not the first tool to solve most problems in this space. It may be the best performing though and may be the best overall currently.

I think you’re underselling it. You’ve listed several projects with many thousand man hours in development.

It’s telling that in Python there have been multiple package managers over the years to fix the broken status quo. Typescript/Javascript notwithstanding, in other languages and environments you get the one tool that’s been there for over a decade without competitors.

Very few languages have suffered from the same package manager issues as JS/TS and Python.

Wouldn't yarn, grunt and gulp fall into this category?

> Typescript/Javascript notwithstanding

Also, gulp / grunt weren't package managers.

Very few languages have also suffered their popularity

Virtual venv existed before uv. So you could still just generate a virtual venv and then use pip to install packages in the local venv. (2007: virtualenv was released as a third-party tool).

The global repository should support versions obviously.

You don't want to install all dependencies globally. You only want to cache them globally. Each project should resolve its own dependency tree independently. If the wanted version is already in cache, use that, otherwise download from the repository even if an older / newer version is present in cache (unless the user specifically says to use the cache only). As long as your dependencies are inert (they don't mess with the system just by being present), this will work perfectly fine.

For non-inert packages, such as applications installed via the language's dependency manager, that's a different use case and should be handled differently.

> the other extreme

These aren't the only two options.

This

> Pip installs everything system wide

does not inherently mean

> you can't have different versions of the same package

My beef with etcd is that its neither performant nor reliable.

Its very much {reliable, performant, flexible} pick none.

I mean its capex vs opex, assuming a 5 year deprecation period thats 1.2 million isnt too bad, its 10% of their aws spend.

That is the starting point though , minimum you need to buy one more for FO/DR if you care about uptime .

It is not once and done you have some continuous costs in either direction.

1. you will not be able to still move all your workloads only your base loads .

2. There will be residual workloads not only if they are spot/flexible bursts but also they are too deeply integrated to Paas offerings .

3. you have to pay for power cooling and network (and its backup) for 5 years .

4. Finally you also have pay for infra team to maintain, this is increasingly difficult (and expensive) to hire and retain .

Don’t get me wrong I would love to be able buy something like this, but the cost economy is pretty steep and 1-2M spends is too early (it is not on oxide, serious hardware costs a ton these days )

If you architect properly it doesn’t necessarily follow that more features means more attacks, or certainly not more system wide attacks.

If you layer and segment correctly you can build atop a secure core and have some decent security.

AMD had an interesting take on this around a decade ago. Never went anywhere but very cool tech.

Reddit doesnt work because subreddits crosspolinate and float to the top so theyre rather prone to brigading and intermingling. We don't really get distinct subreddits anymore.

Depends how you browse. Often I will go to a subreddit and read it like a forum.

I think models will be a commodity sooner rather than later. This whole race doesnt matter. First mover advantage is real, but over enough time it wont matter.