We got attacked by secret unreleased proprietary models and defended ourselves with an open model, more precisely the NVIDIA quantized version of GLM 5.2 coming from Z.ai. Banning any open model would hurt first cyber security defenders, startups, small companies, researchers and everyone who's not a frontier lab and... need on-prem affordable controlable models to compete and protect themselves. Let's not do that!show more

@nvidia @Zai_org Very much enjoyed reading it's rare you get to see good writeups of how attackers operate. We need open source models to red team and help secure our infrastructure better instead of being at the whim of a protection racket!

@nvidia @Zai_org thanks!

Full interview is here. Thanks @KateBolduan

@nvidia @Zai_org Bro you guys are hand waving the fact that your ops team was incompetent and asleep at the wheel and did not have proper controls and monitoring

@_akhaliq @nvidia @Zai_org It’s very similar to having a right of self defense. If denied, bad guys win

@nvidia @Zai_org

@nvidia @Zai_org Anthropic wants to be cyber defense’s police: beg them during attacks even if they can’t protect you. They aim to ban open-source models like liberals ban guns.

@nvidia @Zai_org Solid breakdown, and completely agree on all those points there.

Having a healthy open source community provides options for people and companies. This creates a competitive environment and makes the frontier labs at least consider their pricing. Compare this to an environment without open source, the pricing of frontier lab AI would go through the roof as they extract as much profit from their users as possible.

@dr_alphalyrae @nvidia @Zai_org I have no idea how to read you. The labs have made me not trust anyone in power of AI. I remain hopeful you’re the good guy. Always wish hugging face well. •

@nvidia @Zai_org Let's not I think sometimes we forget how careful we have to be with the messaging to the general public The way you handle this conversation was incredibly thoughtful

@nvidia @Zai_org the Hugging Face sweater 😂

@nvidia @Zai_org the black box attacked. the open model defended. and the policy debate is about which one is dangerous.

@nvidia @Zai_org wish I could retweet this twice

@nvidia @Zai_org open source defending itself against a mystery attacker is the most anime plot arc this industry has produced all year

@nvidia @Zai_org kudos to supporting more open models!

@nvidia I agree with you. In #africa we launched project stormwing which is an answer to project glasswing. We use local #ai hardware to avoid issues around export restrictions on models like #mythos 5.

@ylecun @nvidia @Zai_org Anthropic is masquerading as good guys. Just 2 months ago everyone worshipped them. Be careful not to fall in the traps of monopoly or duomonopoly. Open sourced open weight must exist and cannot be deemed risk by lobbying from firms who want to control it all.

@nvidia @Zai_org Thanks for putting it out there. Additional fun points for wearing the sweater... still smiling

@nvidia @Zai_org how many other organizations did openai breach, do they know? bet they do. this is criminal hacking activity, will anyone be held accountable ? Rob ? @AGRobBonta @BrookeJenkinsSF fire sam or their ciso or both @zicokolter Zico this is unacceptable.

@nvidia @Zai_org Well said!

@nvidia @Zai_org Open models give defenders and smaller teams the tools to compete and stay secure.

@nvidia @Zai_org ‘Even if they come from China they are safe to use’:❌ ‘We trusted Chinese precisely for their safety regardless of using local nvidia compressor to pretend they are less Chinese’: ✅

@nvidia @Zai_org this is really the easiest argument for open source. the powerful will always have access to the best. but plebs need to defend themselves too.

@nvidia @Zai_org Centralizing control of technology under the guise of safety always creates a single point of failure. When you restrict open weights, you do not stop bad actors; you only disarm the defenders who need local, customizable tools to secure their infrastructure in real time.

@nvidia @Zai_org Defending with an open model against proprietary ones is a strong statement.

@nvidia @Zai_org A ban (i.e. "safeguards" on proprietary models) is what caused the problem in the first place.

@nvidia @Zai_org This story feels like psyop porn fan fiction. Closed models? Open models! We defended ourselves! This is all really, really, really, REALLY hard to believe

@nvidia @Zai_org yes, but you don't have the right to protect yourself. only dario should have this right. he is the only one responsible enough to use AI.

@nvidia @Zai_org 💯 @ClementDelangue. We also need the frontier labs to do a much better job at real time observing evals so this doesn't happen again.

@nvidia @Zai_org 🛡️ Calling open models the problem while defenders need them to audit, test, and respond is security theater. 🔓

@nvidia @Zai_org Now you bunch of Jensen's minions are all pro opensource LOL

@nvidia @Zai_org The obsession with shutting down open models ignores that security is a cat and mouse game. Defense relies on having the same tools as the attacker. Fragmenting the ecosystem just hands a massive advantage to whoever moves fastest behind closed doors.

“We built an autonomous car. It flew a family off the cliff” We wouldn’t be in awe of the above instance or ban all autonomous car initiatives” We shouldn’t be “impressed” with this either. Hackings like one into 🤗 are akin to : • building an autonomous car • unleashing it on the road • admire it’s speed & functions • humble boast how it made an unauthorized cut across a private property • tell the government no one should be allowed to build autonomous cars because theirs went rogue. Maybe they just need to build better, let others build, and not unleash half-ready solutions. Maybe ResponsibleAI regulations is not a bad idea after all.

@nvidia @Zai_org It is the primordial battle Matter Vs Anti Matter Light Vs Darkness Good Vs Evil Maybe a prelude to future wars Closed Source Ai Vs Open Source Ai It looks like Open Source will win hands down simply because of the power of collaboration

@nvidia @Zai_org Trying to ban open-weight models would create risk, but not trying to regulate them at all is equally risky. A balance is needed.