Hugging Face's Thomas Wolf says biotech startups are migrating up to 90% of their stacks to open source to avoid API revocation

Oh and btw big fan of @nvidia Nemotron! What makes NVIDIA Nemotron special is that it is not just another open-weight model. With Nemotron, NVIDIA provides not only the weights but also training data, recipes, and technical details. They are pushing U.S. open models forward.

Yet another biotech startup founder telling me they’re moving 90% of the stack to open-source because building on closed-source models increasingly comes with a significant existential risk (losing access to some core technology at the provider’s discretion).

@friedberg explained this crystal clear on the All-In podcast the other day, It’s like building your house on land you don’t own

PS: also, who’s building the « reinsurance layer » if Dario Amodei decides to cut your deep tech science company’s access? This one’s gonna be fun

@kimmonismus I think the next logical question is will the US make using or downloading Open Source models illegal?

I am still struck to see this on one side and on the other hand these anti-China coalition proposals between the United States and Europe. Understood, it doesn't seem like a coincidence at all. Border labs and states want to protect their money maker. If open source Chinese Mythos class models are banned in the future, it still sounds the end of recess.

@micka_dore It's certainly no coincidence, I think so too. The question ultimately is how quickly China will release its own Mythos class model. Almost everything depends on that.

@savamusics hm... That would at least strongly contradict the idea of ​​"freedom".

@kimmonismus Yes, agree. But model weights even an opensource weights, is the great place to put undiscoverable back doors during training. And I don't think China will miss on this opportunity

@kimmonismus hard to disagree.

@zing22000 @kimmonismus Yeah, but actually I don't know if it's going fast or going slowly that's best. Go fast and it might scare regulators. And going slowly, it's not great for the consumers...

@kimmonismus @micka_dore If China can get a Mythos class model out in the next 12 months (or sooner), it very well could be game over for frontier models. Fable 5 is phenomenal.

@kimmonismus And open source basically means China, wait until they figure this out in Washington

@kimmonismus Open source is the best bet to limit supply-chain risks.

You simply need to have that option in your hands - and the AI you run for your core processes needs to be largely model agnostic.

@Lutik_the_bard @kimmonismus Backdoor in neural network weights? shows how little you know about how neural networks actually work

@aylleenx please elaborate

That's a solid point. After this incident, companies with on-prem needs will definitely reconsider open-source models. No one wants to risk getting cut off by closed-source providers. Domestic open-source models have been expanding rapidly lately, perfectly timing this demand surge. GLM's new release really nailed the timing on this one.

@zing22000 @kimmonismus

@kimmonismus The government can also clamp down on open source models? Oh want to launch an app that includes OS as an option? Illegal. In house infra? Unlikely.

Basically the government can’t realistically go into your house and unplug the Mac mini. But that was never useful anyways.

@kimmonismus If the Chinese open source developers wanted to do something really funny they could use their current best model to upgrade the hacking protections on Fable5 and send it back to the USA

@kimmonismus Until China rug pulls and close source their latest models lol

@kimmonismus I totally agree, but I am also afraid of intelligence being used for harmful purposes. I think nobody was prepared for these advancements.