Something, something.. rolling in his grave..

Intelligence ended up being an equalising force. Kurzweil kind of predicted this, but SV was too obsessed with total world domination.

Also you don’t want to connect to the pipe and then after the fact find they’ve started diluting arsenic into it.

Time to roll your own

I have this pang too, but for a different reason: everything is moving to the edge, training and inference. Who needs a complicated heterogeneous setup like this when you’re running everything locally?

It should wait for when it can get it for pennies

The open models are now good enough. The rest of the world needn’t concern itself with the petty dramas of the colonies any longer

Good enough for what?

The labs are betting on there being more applications for stronger models and that many existing applications will continue to need stronger models to keep up with competitors.

Yep. I keep thinking "this is good enough for almost anything" and then I remember how my Pentium 75 seemed like it was good enough for almost anything I could think of at the time. And it was!

The US controls most of the capital in the industry, including most of the compute used for training. And also, NVIDIA, arguably the most important actor in the industry

The US military can come and separate me from my MacBook and Qwen if it likes. I dare it. Sadly however, I suspect they’re busy with other matters at the moment.

PSA: The government can remotely brick your MacBook.

Air-gap one of my Macs.. thanks for the advice! ;)

Surely this is great for an end user: the taste as to “when” and to what degree a model should “think” is now entirely in the fine tuners hands

I’m super interested in the opposite experiment.. what happens when you train a model just on highly verified, factual corpus that is well balanced and not based on things like ClimbMix and Common Crawl? My intuition is the unverified/unverifiable goals inherent in a model (eg GPT hacking huggingface) are latent in the 4chan/reddit slop it’s trained on during pretraining.

Common Crawl has very little content from Reddit and 4chan -- both blocked us years ago.

> Providers can just run them, offer cheap tokens, and pocket the margin.

There’s an assumption that you can spin up the infra and acquire customers within that margin

Which is not unreasonable. Just hosting it in the EU and promising not to retain / sell the data let's you charge a healthy extra and compete in many areas other players can't.

> Just hosting it in the EU and promising not to retain / sell the data let's you charge a healthy extra and compete in many areas other players can't.

It's been a few years. Has anyone done this successfully yet?

There are a over a dozen EU open-weight providers. I’m not sure if they are even charging that much of an extra. EU-based clients have little reason to use non-EU inference providers.

> EU-based clients have little reason to use non-EU inference providers.

Which models are most popular in Europe?

I don’t have user statistics but my mail/domain registrar Infomaniak advertises Qwen 3.5 and Apertus, “a Swiss open-source AI model, developed by EPFL, ETH Zurich and CSCS”

https://euria.infomaniak.com/

melious.ai comes to mind.

Keeping SLAs spinning isn’t this trivial

> There’s an assumption that you can spin up the infra and acquire customers within that margin

Only Nvidia and approved friends can at the moment. Nvidia can even backstop your loan required.