If you can analyze it with AI it implies AI can understand it (as it can analyze it for itself as well) and thus can build on it.

But if you train the next generation of models on it, will AI be able to use it? How much larger will these models be, etc.

Also, AI has limited context. At a certain point proofs may become so complicated that an AI cannot keep most of it in memory and will not reuse it for new proofs.

Serious people haven't been very dimissive of Chinese models since at least DeepSeek-R1 in January 2025. Sadly, we in Europe are much behind.

Pleasant read, which feels as if it's from an outsider to the event. Alas, I just don't think there's that much to report.

It's possible to be a waste of European money that could be better used elsewhere, though. I think the same about LeCun's company sucking up the little funding here.

> Europe absolutely needs a home-grown AI lab, especially with Pax Americana looking increasingly shaky.

The only problem (at least in the LLM space) is that you can do more in Europe by just getting the best Chinese Open weights model (do a finetune if you really want) and do more for cheaper than using Mistral.

Yes, this is exactly what I expected (showing the anthropic principle in action), so I am disappointed they are not sampling correctly.

There's been a ton of optimizations already, it hasn't remotely reduced demand even temporarily. More efficiency just makes the compute have even higher ROI per $ and watt spent.

With sufficient optimisation, there ought to be a tipping point beyond which local inference is good enough. And, sure, datacentre compute will still be needed for training but one of the biggest current uses will begin to taper off.

The question really is how soon we reach that tipping point, and whether it's before or after the current bubble runs out of steam for some other reason.

>there ought to be a tipping point beyond which local inference is good enough

There's no such ought really. Even at current levels you'd need like a 100x gain from here to approach current top proprietary models (probably a lot more for say Mythos or Mythos 2), and it's not like they are stoppng to improve. This is before we even account that you'd just be running 1 agent then, and not a swarm like you'd be able to in the cloud or that you can do only so much compression before you are losing out

> (probably a lot more for say Mythos or Mythos 2)

Not everyone needs that large of a model, though.

Too base cynicism. I'd be willing to bet you my $200 to your $100 that doesn't happen.

The history of acquisitions mothballing the acquired assets or moving them off their original mission or letting them atrophy is too stark to support your optimism. Especially when nvidia has a direct incentive to to steer the ecosystem. For example are they going to highlight and surface alternative chip designs like grok? Or will model search mysteriously not find related models?

Nvidia doesn't exactly have Oracle's reputation when it comes to acquiring companies, but given its overall hostility towards competition and open source I can't blame people for being anxious about the news.

The most important resource for self-hosting LLMs is now under the control of a company that has very markedly kept the specialized hardware outside of the broader public's hands.

You're probably right. It was mostly hyperbole, but the older I get the more I think hyperbole does eventually come to pass in situations like this. It's gradual, though, not all at once.

That's a good formalization of it. I will think you are a bit of a fool if you dont use AI but you guarantee you'll be a fool if you only use AI with no value added by yourself.

People all agree on that sentiment.

Where they disagree is if you add value on the knowledge that makes the prompt possible or add value in rewriting AI to pretend you wrote it.

Fable has more parameters. In practice it's not yet clear which one would be better for different usecases yet but they are more different than one being strictly better.