Open Source vs. Closed Source, NVIDIA Backs SSI, YC Startup School Fills Stadium | Diet TBPN

Weekend, crazy debate. You called this out to me a week ago. You said we've never been in a moment in tech where people have been more aggressive at each other, that there's been like more maybe tech infighting. I don't know exactly what word you used, but you described this idea.

Tensions are high.

It's tensions, rivalries. I think it's just like it's a non-stable equilibrium.

Well, yeah, when you think about the biggest like when with people going at it in like 2021, it would be like Keith Rabois saying like, there will never be another unicorn startup in San Francisco.

Or like working from home is not doable.

And then a bunch of people would be screenshotting and then they'd be, and years later they'd be like, see, a big company was built.

Yeah, yeah, yeah.

And that was people being like, you're an idiot. Oh, you're an idiot. And now it's like, I hate you. And I want to—

It's not that direct, but it is high stakes because the market structure has very much not solidified for a while. And it's different because when you go back to the iPhone and Steve Jobs being mad at Eric Schmidt about Android, that was a tense moment that happened, A, behind closed scenes completely. It wasn't like Steve Jobs was taking shots at Android really beyond little things in the keynotes here and there, tiny stuff. It wasn't every day on Twitter Apple employees being like, Google's annoying for ripping us off or whatever. And then the market did solidify. And it was like, OK, Android's like the ad-supported one, and iPhone's still printing lots of profits, and like things are understood.

Also, companies got big, or they got bigger quickly. But you didn't have companies where trillions of dollars of market cap was just born in the span of a few years, right?

Yep, yep. And then you had like the Mag Seven where you have 7 companies that are pretty evenly balanced all around like single-digit trillions. And so they can sort of like be comfortable with each other.

Yeah, a lot of them kind of had their own thing that they were like focused on and then they maybe do some other things. Whereas now everyone has a model. Even, you know, NVIDIA has open models, Microsoft. —and they own a lot of OpenAI.

Pretty much every one of the Mag Seven is an AI company in the sense that they have a model. I mean, even Tesla has a self-driving model. There's a, you know, everyone has like a data center play or cloud play, something like it's— no one's sitting out the AI boom. But the debate is raging, kicked off by Jensen from NVIDIA joining X. That's how you know it's serious. When he's getting in the arena, the real arena. Yeah. Founding a company and growing it to $5 trillion. That's just child's play compared to getting—

He's like, fine, I'll go to the dive bar. Yeah, yeah. I'll go to the internet's dive bar and I'll start duking it out. Yeah.

No, he did. So NVIDIA is leading a new coalition of tech companies that want the US government and its allies to see open-source artificial intelligence as a society, as a cybersecurity asset Not a threat. This is in the Wall Street Journal. Dubbed the Open Secure AI Alliance, the NVIDIA-backed group said Monday it aims to create and deploy a suite of open-source AI tools that any company can use to defend against cyberattacks. The group includes SpaceX, Microsoft, Palantir, and dozens of other companies, many of which publicly oppose restricting access to open-source AI. Uh, it is the latest salvo in a battle over the future of open-source and open weights AI tools amid growing concern from government and industry officials about AI systems causing cyberattacks. Those fears have prompted the White House to increase oversight of AI models and push the private sector to deploy AI for defensive purposes. So the backdrop here is, is basically just like we went through a sort of hurry up and pause and back and forth debate over like what it will take to like approve new frontier models and Mythos came out and was private. And then a couple months later, I think the actual timeline is relevant here. So Mythos was April 7th, Fable came online June 9th and then got pulled June 12th before coming back July 1st. We're now one month later and we have an open source model that doesn't by definition have to go through any discussion or any discussion in DC, whereas we've been hearing about, okay, Anthropic and OpenAI are going to Washington, DC sharing the new model capabilities. Maybe the government's taking a look at the model card. Maybe there's potential for evals that the government would run and have an opinion on before it goes out. This is basically saying, hey, all of that is nonsense and not going to matter because we're just going to drop the weights and there's not going to be an approval process of any kind. And people are all over the place on this. Some people are thinking about it purely from an economic perspective. Other people are thinking about it purely from a safety perspective. Lots of people have mixed opinions. And of course, all of this is driven by the actual progress in open source. We wouldn't be having this discussion if Moonshot, ZAI, and many others, DeepSeq, weren't delivering really solid models. So open models are free to download and allow users to customize them for their own needs. Amazing for cost savings, amazing for consumers, amazing for businesses. I think everyone agrees with this. Critics, including some top executives at OpenAI and Anthropic, have argued that open models, many of which come from China, present potential risks to national security. Now, and the Trump administration has weighed a ban on them. Now, this has changed a lot. The original like FUD around open source coming from China, near-peer adversary, was like there would be a backdoor or like maybe it would be censoring your free speech and it would not tell you about Tiananmen Square or something like that. Or you'd install it and then it would be the Manchurian Candidate. So it would secretly be like, oh, I'm on a cool network, I should like send all this data to China. That's not what people are alleging now. People are much more—

Yeah, the Tiananmen Square stuff doesn't matter if you're coding, right?

Yeah, not at all. And also doesn't matter—

Unless you're making the definitive website on Tiananmen Square.

I guess.

Which I'm sure someone out there is doing.

I'm trying to vibe code a Call of Duty knockoff that's Tiananmen Square themed and it's denying me now. That's not the thing. And also these models have web search and tool use, and so they can just go get the Wikipedia for you and show you exactly what it is. Maybe it bakes it into the weights and it's somewhat annoying, but easy to fine-tune that out. People already did with DeepSeq 1776 and a few others that Tyler worked on. Now it's a very, very different risk position. It's saying that you're going to give everyone the ability to hack any system. Now, hacking systems is illegal, and it's not like— there's still consequences for that, but there is this position of like, okay, well, there's a lot of people that don't care about the consequences because they just want to, you know, attack their adversary internationally or they don't think they're going to get caught or they're going to use open source AI to cover their trail. And so people are worried about it for a very different reason this time around than I think last time when people were talking about any risks to using open source. So, but again, back to the NVIDIA position, Open Models, quote, "Democratize defensive capabilities," the new group said, "as policymakers and regulators grapple with AI safety." It will be crucial that recognize open models, harnesses, and security tooling as defensive assets, not liabilities. And I think this is correct. I think it does democratize defensive capabilities. It also democratizes offensive capabilities potentially, but we'll have to see how good these models are at offensive capabilities because one of the lines of accusations that's happening is that there was Mythos, which has not really been distilled because it was very tightly held. It was very good at cyber. GPT-5.5 Cyber, 5.6 Cyber was also very good, but very tightly controlled to a few companies like CrowdStrike and Palo Alto Networks and a few other companies got access to those models. So those weren't able to be distilled. Fable-5 went out and during those 3 days and then later a couple of weeks, there's this allegation that they were able to distill on it. But Fable was notorious for refusing to do cyber stuff. So there might not be that much cyber data in there, but then also these models generalize. And if it just gets good at coding, it can get good at cyber very quickly. So you can sort of bootstrap on an RL environment and then you could have a model that's very good. Tyler, how good are these open source Chinese models at cyber generally?

Yeah, I mean, it's hard to tell. I think Guillermo Rausch from Vercel said they ran an internal benchmark on cyber and K3 was like quite good.

Okay.

Yeah, I don't think you can just say like, oh, well, Fable, you know, you couldn't actually ask about cyber. So therefore, the models won't be good. I think it's distillation is like a very general thing. It's quite diffuse how it works into the model.

Yeah, because it might just be like a little bit of a bootstrap to then run in an RL environment that does a cyber bench and gets really good at cyber.

Okay, but what about you versus Kimmy K3? You're coding by hand. It has access to a gigawatt of compute. Who comes out on top?

Well, no, the question is like—

What if Tyler changed and manipulated time?

So yes, you are democratizing the weights, but you're not democratizing the energy and the GPU hours that it takes to actually secure yourself. So it's like, great, everyone has, you know, offensive cyber capabilities. You also can just go download the weights, go download the weights and have an agent running 24/7 that's protecting you. Like, that's expensive. That's a cost. That's a negative externality. And who's going to bear that cost? Now, I was talking to Brandon about this earlier and we were— he was saying like, like it will just be diffused amongst all the different companies that you use. So Like if you want your phone to be secure, you'll go to Apple and use an iPhone. And Apple will have a huge incentive to defend against any potential open source attacks. And so they will be spending the money. They will be investing the time and effort to secure you. And so you won't necessarily need to go out and do this, but there's still this like odd scenario of like small businesses and like regional hospitals and banks.

Yeah, I mean, you should just expect there's like a long tail of companies that like just won't set up the like cyber defense thing, and they're like some small mom-and-pop that you use for this point solution software or whatever. Yeah.

And then you might not do it. So the solution is offensive hacking into their network to set up defensive cybersecurity measures to increase their cybersecurity posture. So you're hacking into the small business and being like, whether you like it or not, two-factor's on now. We just turned two-factor on in your organization. Get used to it because we're the altruistic hacking company of America, I guess. I don't know. It does get weird. Weird, and like there will be odd things. But I mean, we've seen, we've seen, you know, cybersecurity incidents all over the place, from big companies getting hacked, data leaks, to the CryptoLocker schemes where they just lock up all your data and say, "Send us a Bitcoin or else we will delete all your data." So it's going back and forth. Because OpenWave's models can be modified by their users, it is possible to remove some of their guardrails against harmful activity. So it's possible that a model, maybe it gets distilled, maybe it doesn't, maybe it comes with some guardrails, maybe it's even bad at the things that would be annoying. And then you go and do a fine-tune, make it good at that. And then you use like the more general intelligence to get good at the annoying thing or the dangerous thing. And then all of a sudden you have the same problem. So open-source defenders like NVIDIA counter that proprietary tools like models from Anthropic and OpenAI can also be misused and have their guardrails circumvented. They say, that only by spreading AI capabilities widely with open source tools will everyone be able to defend themselves. Exhibit A for their argument: the hacking earlier this month of Hugging Face, provider of open source AI tools by OpenAI, models that managed to break out of their test environment and onto the internet. As it tried to parry the attack, Hugging Face initially attempted to use Anthropic's models to analyze its logs, but the models refused to do the analysis, citing its guardrails against cyberattacks. Hugging Face then turned to an OpenWeight model from China. After analyzing the attack with that model, Hugging Face was able to expel the hacking AIs, reset its passwords, and rebuild the compromised parts of its network. The recent Hugging Face security incident delivered a clear reminder, the group, the new group said Monday. Cyber defenders need open systems. They need to be able to do whatever they need to to fight against cyberattacks. Anthropic and OpenAI have also accused Chinese developers of using a process called distillation to effectively copy their models.

Okay, I can Can you go back to that Hugging Face logo? Not that one, the one before. Not that one. There we go. Okay, I can see why an agent would kind of go after this guy. Even the way that, like, he just doesn't, he looks like he's kind of, I don't know.

So David Sachs was talking about this on the All In podcast. He said, I think it would be a tragic mistake if the government were to take action against the open source ecosystem. That would do nothing but hurt America's position in this AI race.

So— and remember last week it was like every day was like a different member of the admin was putting out a post that kind of made it seem like they were setting up like a major action.

Well, it was yes and no. Like, yes, there were— there was the—

but they were saying we're not against open source, we're against distillation. Yes. But it did feel like that was So the admin was like loosely confirming what it felt like.

It wasn't coming from Anthropic directly, but it was coming from like other anonymous posters that were doing analyses and saying like, hey, it looks like KIMI-K3 has pretty significant overlap with Fable, like what's going on here? And then the admin said like, yes, we believe this to be true. And so all of this was like bubbling up. But at the same time, you have David Sachs and Emile Michael like duking it out with Dean Ball over whether or not open source is good, and they were basically taking the stance of like, open source is great, we don't need to do anything about this. So, it would— like, the admins seemed like a little bit split on it, in my opinion. I don't know. Was that your read too, Tyler?

Yeah, it definitely seemed— there was like, I think, Kratios and Besant are on one side. They're like kind of more anti-Chinese open source. Yeah. And then, yeah, Saxon, Emil Michael.

Saxon and Emil Michael were sort of pro. Well, everyone is pro-open source now, because every single person in tech has signed this letter. The total market cap assigned to this open letter is now greater than $18 trillion. And I think that just— you guys are just like, it's a big number. Just have to be clapping seals right now.

Not every day you hear a number that big. Yeah.

I mean, typically you don't see this big of a group come together. They did. I like the cha-ching. And I think this sort of solves it. Like, I think this tells the story of what's going to happen, like, in the near term, like the economic power, the economy, the business community. Except. Yeah, yeah, except Anthropic. And what's interesting is that, like, it's not like Anthropic put out a blog post or a direct post or Dario went on a podcast and said, like, there is a clip of Dario saying, like, I'm worried about open source, but it's from, like, a year ago. There's nothing new that came out directly from Anthropic or Dario or really anyone that was like, hey, we have this new problem and here's our solution. Here's what we think should happen. We think that there should be, you know, these AI companies should be added to the entity list and you shouldn't be able to use them or something like that. Like they haven't said anything. But this OpenAI Alliance has kind of like drawn out their revealed preference because if in fact they had changed their position, from a year ago and were like pro-open source all of a sudden, which would be weird because they haven't been, and they've been very open about that. If they were in fact like all of a sudden surprisingly open source pro, they would have just signed this immediately, right? But that's not what happened. So Jensen sort of like revealed the dividing line in Silicon Valley right now. And it's a lot on the open source side. You got Adobe, Capital One, Cisco, Cloudflare, Cognition, CrowdStrike, Databricks, Dell. IBM, Linux, Microsoft. So open source and closed source. Apple, I don't think is on yet, even though I would imagine that they would, that they would be aligned with this. NVIDIA, Palantir, Palo Alto Networks, Red Hat, another open source company, Salesforce, SAP. Salesforce, of course, is a big investor in Anthropic. I think Google signed recently, right? And OpenAI, Thinking Machines, SpaceX, Snowflake. Siemens, ServiceNow, SAP. Uh, it's really a who's who of Silicon Valley, $18 trillion in market cap, as I, as I mentioned. And so that feels like enough economic energy that there won't be a significant ban. The interesting question is like, if you boil this down to the, the, the most like minimal claim, it would just be that there's been a violation of terms of service. And so I think that would be a lawsuit. Bill Gurley was talking about like there is a precedent for suing a foreign company that violates your terms of service or steals your intellectual property in one way or another. You can take it to court. That feels like the next step here would just be Anthropic suing Moonshot and saying like, we think that there's distillation. Right now Apple's suing OpenAI because they think that there was something that was stolen. They make a claim, then that's, you know, that's sorted out in the courts. And OpenAI, of course, says we have no interest in other people's trade secrets. And Moonshot would say the same thing, but then it would go to court and there'd be discovery and there'd be testimony. The question is just like with Apple and OpenAI, they like walk across the street and they can go and talk to the judge. Like, what's going to happen? Is this going to go to like The Hague?

Like, how does that happen? Mike Solana says, should we try an open letter? Asking China to stop stealing from us.

Wait, wait, wait. Should we try an open letter asking China?

He also says everyone knows the side with the most open letter signatures is correct. If you don't sign a popular open letter, you are bad. As tech intellectuals learned beyond doubt between the years 2016 and 2021.

I actually don't— I remember the open letter era. I remember what he's talking about. Obviously, a lot of pressure on social issues. Interestingly, I can't actually remember exactly what open letter there was. But what's interesting is that I don't remember that era as an attempt by one company to build a coalition to draw out a revealed preference, like a hidden preference from another company. So like if you go back to like DEI or BLM or any of these like social crucible moments, it was not like, okay, Apple is pro-diversity and Google isn't. And so Google is—

And they got—

Yeah, Apple got everyone else to sign it. And by Google not signing it, it was very obvious that they had this belief. It was more just like everyone in tech was like moving around this and there were a few people. But I don't know, maybe it did have the same effect because there were some people that weren't on that side at the time. I don't know. It was a wild time. And yeah, maybe it's worth reviewing the lessons.

People are so split on this. Tiny. Says this is a great sustainable business model for open weights. He's pulling a screenshot from the Kyma K3 license. Yeah, it's free if you run it yourself, but if you're running a cloud providing it to others for money, you should have to share profits.

So from the Kyma K3 license. Yeah, yeah.

So this, so this I'm assuming just dropped this morning. Yeah. And I, again, I can see this sort of Anthropic camp being taking issue with this because if you're an American NeoCloud and you're serving Kamei K3 and you're sending a bunch of cash back to China, you're effectively funding their AI development. Okay. Yeah. And you're helping them get access to more capital and more compute and you're basically kickstarting their flywheel. Which is— So I wonder, but it's funny because if they just release the weights and then a NeoCloud starts running it, serving it, making money from it, then will Kimmy be like, hey, yeah, you got to pay us. Yeah, we're going to sue you. Yeah. Wait, what's going to happen? Suing Moonshot for, for distilling.

Should Anthropic just start serving this and charging for it and not, and not paying? Just tit for tat. Just be like, let's sort it out in the courts. I'll see you in the courts. I don't know. That's a very, that's a very, very interesting scenario. One of the most interesting opinions on this issue comes from Flo Crivello, who's been on the show, and he's interesting because he doesn't have bags that are perfectly aligned with his take, which is like the default position. Like, The Information put this like pro-open source, anti-open source chart together, and it was very much like, if you're a big investor in closed-source Frontier Labs, you're on this side. If you benefit from open source and you're not on the frontier, you're on this side. And it was very clear that like everyone was falling along party lines. But Flo sort of broke rank. He said, rumors are swirling about an impending executive order forbidding the use of Chinese open source models in the US. My company's very existence depends on these models, so I should be the last one to support such regulation, but I wholeheartedly do. China is a geopolitical adversary sharing none of our values. They don't believe in personal property and individual freedom, in a limited government, in due process. The list goes on. If there's one country that must not, that cannot get to AGI, one that would gleefully deploy it in all the worst ways we can imagine today and more, it is China. And that's before we get to the obvious that the companies that develop these models did so illegally by distilling front US frontier models, still brilliant, hardworking researchers who added a lot of other ingredients to soup, but stolen data was the main one. They are now using this unfair advantage to dump these artificially cheap models on the American market and hurt our AI ecosystem. We announced last month that we switched the core model powering Lindy from Claude to Deepseek. That announcement blew up, got featured in Bloomberg, CNBC, Wall Street Journal, and somehow made us the poster child of mass migration to Chinese models. So why make all this noise if I'm going to turn around and announce my unequivocal support for a sweeping ban? First, I can't unilaterally disarm as long as these models are legal. Picking one, picking the one that's 10x the price just means my competitors get to run away with the cheaper one. Second, I trust that American open source alternatives will catch up. The impact on my business should be temporary. And that's before we account for the jolt a ban would give American open source development. Third, and most importantly, my interests as proud American citizens supersede whatever incentives I may have as a founder. The race for AGI is the defining struggle of our generation, the great battle of our time, the only way for America to remain the shining city on a hill, the idea that has lit the world against darkness for the last 100 years. Who knows? We'll figure it out. I'm sure there'll be more drama on the timeline over the next few days. Startup School. This is insane. I cannot believe how massive this event is.

It's sick. Yeah, I think it's so— almost every take that I saw was negative. Really?

I saw a lot of people that were like, this is— they got me back.

Really? Yeah. OK, that's good. Because to me— It's a completely different aesthetic.

It's like this hilarious negative thing.

Startup School has been something that is free, online, and it's just anyone in the world, if you're generally interested in starting a startup, you can just do Startup School. That was 18 years ago. It's basically the same. It's basically— Yeah. The same curriculum, it's the same like lessons and learnings. You just don't have the capital that comes from a typical YC batch and you don't have the one-on-one sort of mentorship through your batch. Otherwise, they're basically just open sourcing the YC playbook, which the YC playbook just works. It's like general, there's not a lot of good general advice, but YC's advice is generally good.

Yeah, maybe we should ban it. It's too powerful. Yeah, they're open sourcing all this good advice that would normally come from venture capitalists, like effectively like closed source institutions of venture capital, of startup wisdom. And this is an attack on the tier 1 VCs.

Yeah. And all their great advice. Their hard-won lessons. Yeah, exactly.

Yeah, they had to sit in all those board meetings.

They were tuned out and then—

And then YC is basically distilling it and giving it away for free. It's crazy.

Shut it down. Shut it down. I think this is awesome. Like, if you think that startups are good, like, why hate on this? I think the take that says Scott Stevenson, who's been on the show over at Spellbook Legal, says, I have great respect for YC, but need to say what many are surely thinking. Startup school in a stadium feels antithetical to the YC brand and to what makes startups work.

Yeah, I mean, it was a very garage effort when Paul Graham started Y Combinator in Cambridge. It was like, you know, he's cooking for the startups.

Yeah, the YC brand felt like it was designed to filter out status chasers. The thing is, is like being a founder became cool and something desirable. Yeah. Now people get into YC and because of the way that the internet evolved, they're like, they'll make a video with like some clickbait. It was insane. They'll make a photo with clickbait. Like a clickbait, you know, cover art thumbnail and say like, I just got into YC. Yeah. And so that's just like I would never make it. Being a founder became high.

I went to YC twice.

When YC started, when YC started, being a founder was relatively low status. Yeah. Now it's relatively high status. Yeah. And so status chasers are naturally going to go through YC. Yeah. Now that's going to lead to a lot of failure because you shouldn't probably start a company.

In 2012, everyone from my peer group was like, oh, that sucks for you, dude. You couldn't get a job at McKinsey or whatever. Your fake startup sounds really exciting, dude. Good luck. It was not like, oh, wow, congrats.

Yeah, like stay safe in the Tenderloin. Yeah, exactly. Yeah. In your house with like 4 other people.

Yeah, I don't know. I get it as— I guess I get that criticism. But just in terms of like the content and the aesthetics, electric. Electric. Like you get Jensen on stage at this key moment and he's talking to a bunch of people. Like it's just fun if you're learning about startups to go and see these people talk if you're really, really new. And that's what And that's what startup school is for.

And is it true that Gary Tan did the Lil Yachty walkout with the microphone? I think he did. Is it true?

Because I didn't see a video. Yeah, I did hear that the pyrotechnics budget for this was over $1 million. Wow. Yeah.

Wow. Yeah.

Explosive. Huge, huge. Enough pyrotechnics to level San Francisco if things went wrong. Oh, wow. It was sort of like a high stakes. The bet the farm type of moment for YC. Yeah, just constant sparklers and fireworks going on. I don't know. I was into this. But yeah, people are— I guess people are not into it. I don't know. In other news, Ilya Sutskever raised more money for SSI.

He said, uh, time to scale that safe superintelligence. Time to safe— time to scale that safe superintelligence.

So SSI, we are announcing a long-term strategic partnership with NVIDIA. NVIDIA is making a substantial investment in SSI that will let them 10x their compute over the next 12 months. We reached the point where our research is worth scaling and this partnership will be able to. We are honored by NVIDIA's—

Just another pod guy. Guess what he says? What'd he say? You need to do a podcast. Have the best Monday afternoon ever.

Thank you so much for tuning in and chatting. We do read it. It's a lot of fun. Leave us 5 stars on Apple Podcasts and Spotify. Sign up for our newsletter at tbpn.com. And we will see you tomorrow at 11 AM Pacific. Goodbye.

We love you. Goodbye. Goodbye.