The whole industry is going to need to be regulated and I think the industry needs to regulate themselves. That's the key to this. We need to have a set of tests that Google, Microsoft, Amazon all agree to. Elon, hey, these are the things we should test and they should self-certify each model before asking the government which doesn't understand the models to certify them. The industry should have an industry certification like they do for countless other things. I've talked about the MPAA and the video game industry. We should just self-certify. It's the simplest thing in the world to do. And then we could release the models ourselves without the government getting involved.

This early articulation of industry self-certification frames the entire debate about whether AI companies can and should regulate themselves before governments force the issue.

I think it's worth putting a little definition around this proposal, which is to form an SRO, self-regulatory organization, because they're not purely independent. SRO's like FINRA and the National Futures Association, they exist in the financial markets and they were created to allow the financial institutions to set their regulatory rules, how they check each other, how they make sure that everyone is being safe because they're obviously all trading risk with one another... the analogy with AI is pretty appropriate here which is that there are many players in the industry. They are all trying to progress AI technology and no one wants to have a single regulatory body that comes in from the government or outside that says here are the tests you guys have to pass with your models in order for them to be appropriate.

Friedberg's detailed explanation of SRO mechanics clarifies exactly how the financial-industry model would translate to AI governance, making the abstract proposal concrete.

I told him that I could potentially get on board with this, speaking just for myself, not on behalf of anyone in the government because I thought that an SRO, again, a self-regulatory approach would be infinitely better than creating a new government agency that I think would rapidly become a DMV for AI. Dario calls it an FAA for AI. The government does not have the expertise to evaluate AI models. The criteria are changing too rapidly. you're going to very rapidly end up with a queue where all the models would be waiting to get tested and it would start with a month-long delay. It would end up being many months and we would just lose the AI race.

Sacks articulates the core tension between safety oversight and competitive speed, arguing that government bureaucracy could hand the AI race to China through procedural delay alone.

Number one, I think the SRO has to have broad representation from within the industry, the AI industry. It has to include startups and open source. It can't just be the three biggest labs... And that's precisely to avoid the problem of regulatory capture, right? If you have a diverse enough group of interests being represented, it's much harder for this to turn into regulatory capture... Number two is I think that this body should only be reviewing frontier models, meaning the true frontier. The models that really represent an advance in the state-of-the-art of artificial intelligence... Number three, which is I think this body should be dealing with catastrophic risk only. And to my knowledge, those right now are cyber and CBRN, meaning it's, you know, chemical, biological, radiological, nuclear.

Sacks's five conditions offer a rare policy-level framework that attempts to thread the needle between meaningful safety oversight and avoiding regulatory capture by incumbents.

What the FAA does, among other things, is approve new airplane designs. Okay? And specifically, it requires what's called a type certification for any new aircraft design or major changes. And for an entirely new aircraft design, it takes 5 to 9 years to get the certification... So this is permission-based regulation. There's no approval, no flying commercially. It's safety first. Look, that might make sense in the case of preventing plane crashes, but when you're talking about AI models, you're talking about replacing a system that is releasing new versions every couple of months with one that is potentially fully under the control of government, fully government approved. Everything has to be certified, and you could expect the timeline to go from months to years. Again, I think we'll just simply lose AI race if that happens because China is not going to abide by those rules.

The FAA analogy stress-tests the most aggressive regulatory proposal on the table, showing concretely how permission-based safety frameworks clash with software release cadences.

In October of last year, I tweeted that Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering and everyone kind of went crazy over this... Back then people thought that I was beating up on a little startup. Now I think everyone can kind of see the truth which is look this is not a little startup. They already have a trillion dollar market cap valuation... And there was an article in Politico just the other day, is called inside Anthropic state-by-state plan to ratchet up AI rules. And what it says is quote AI giant Anthropic is pursuing a strategy of one-upmanship that encourages states to impose increasingly tougher AI guard rails rather than a line around a single set of regulations.

The allegation that a leading AI safety company is strategically engineering a regulatory patchwork to entrench itself is one of the most consequential — and contested — claims in the current AI policy debate.

This is the mistake that I think a lot of people in the tech industry are making is they think that they can just buy off politicians or the political system by making concessions. No, that will just lead to a ratcheting up of the pressure. The government will be happy to take this and then come back for more and more and more until it's fully under government control. So, at some point, I think these companies are going to have to grow a spine and fight and decide where they're willing to draw a line. And if Demis' SRO is the line, if they're saying, 'Okay, we think this is the right solution and we're going to fight here and this has to be it and in exchange for this we need preemption and we need, you know, other things written into law that make sure this is where the line is, then I think it can work.'

Sacks argues that incremental regulatory concessions without legal preemption are strategically self-defeating, reframing the SRO not as a compromise but as a line in the sand requiring explicit legislative backing.

The interesting question to ask is what is the only kind of baby that Advent and Stripe and Block could have together and I think there's one which is you are creating a competitor to Visa and Mastercard because you now have upwards of 6 or 700 million accounts. You have massive stable coin infrastructure. You have all of the risk management infrastructure that Stripe has built over the last 15 or 20 years... you can vertically integrate and go soup to nuts. That is probably the most obvious thing to make out of it. So that now Stripe gets access to an ultra-low-cost set of payment rails literally like to zero brings it everywhere all over the world.

Palihapitiya reframes the Stripe-PayPal-Block deal not as a legacy acquisition but as a potential structural assault on the Visa/Mastercard duopoly, which would be one of the most significant shifts in global payments infrastructure in decades.

I think there's going to be more of these kinds of deals. If you look at Ryan Cohen's bid for eBay, I think it's probably a second dot on a line that I think is emerging, which is folks that are call it AI native are looking at call it first generation digital native businesses that have become mature and old and stale and aren't run by the founders anymore and have not yet realized the opportunities with AI, have not yet realized their potential, or overspending in a lot of ways. And when you take a look at those businesses as a modern-day AI operator, you're like, 'What the hell? This thing is so under utilized. They're not using their network well. They're not operating well. They're overspending. They're not using AI well.' And there's a set of opportunities that become quite obvious.

Friedberg identifies a new M&A wave driven specifically by the AI productivity gap between founder-led innovators and mature digital businesses, suggesting a systematic revaluation of legacy internet assets.

Privacy in AI is very fragile and it's very brittle. And this is despite the best efforts of great businesses... my takeaway is that there are all kinds of non-obvious data leak vectors lurking in AI. And so if you think that you're going to flip a ZDR switch, zero data retention, which is the magic term that the industry uses to tell you that everything's going to be okay. I think the answer and the message should be it's not going to be okay because you can't guarantee any of it. So the model companies when they give you these zero data retention policies are probably trying their best. But I think the reality is you are leaking information where you don't know it. and they despite their best efforts may still have trap doors that they don't even know about until it's figured out by somebody else.

Palihapitiya's warning that zero data retention promises are structurally unverifiable — not just marketing spin — has direct implications for every enterprise considering deploying third-party AI on sensitive workloads.

if your engineers are going off randomly in an unguided system and then just ripping through million tokens at 56 bucks, what he's talking about is the eventual downstream impact to earnings. And that eventually a bunch of these public market CFOs are going to show up to Wall Street and they will have missed earnings because they're upex at some point. If things are 21xing every few months, somebody's going to miss a quarter. I don't know who, but somebody... unless you get a control of this and you can directly say how much money you're making, this is a bridge to nowhere. It is a money burning furnace.

The prediction that uncontrolled AI token spend will cause public companies to miss earnings estimates points to an under-appreciated operational risk now embedded in corporate P&Ls.

Mark German, who's like the most in then guy when it comes to Apple, he says, and you know, we got this new CEO coming in for um uh John Furnus. Yeah. And he is a hardware engineer. M7 Ultra uh because M we're on M5 chips now. you can get like, you know, 456, 512 gigs of RAM. He says M7 Ultra is going to support as much as 1.5 terabytes. That's double what they're already supporting. So, if you think about uh Frontier models, like the last generation, this is like an Opus level model running on your Mac Studio... This is going to change everything. you're gonna have employers go, 'Oh, I can just run, you know, 90% of my workloads, 99% of the workloads on the local uh Mac Studio.'

The prospect of frontier-class models running locally on consumer hardware represents a fundamental shift in the AI pricing and privacy equation, threatening cloud AI incumbents with hardware-level disruption.

That is why today I'll be signing the nation's first ever statewide moratorum on hyperscale data centers.

New York's first-ever statewide moratorium on hyperscale data centers marks a significant policy escalation in the debate over AI infrastructure.

If you do what Chamas said and let them build behind the meter, then they bring their own power. And that's what the president has advocated for since the beginning of his administration is let the AI companies become power companies.

The 'behind the meter' model reframes AI infrastructure companies as independent energy producers, potentially sidestepping grid-load concerns entirely.

These data centers have become the scapegoat for all the angst that people have about AI and it's kind of become this very clumsy way of trying to throw a wrench in the gears of innovation and just kind of slow the whole thing down.

This frames the data center debate as a proxy war against AI broadly, raising the question of whether infrastructure opposition is a legitimate policy concern or a misdirected moral panic.

He said you could actually trace it back to the same people that in a different era were protesting fracking. And so he was saying like it's these are all just hobby horses that they use to raise money, have a job. They're like professionally paid protesters.

The claim that anti-data-center activism shares funding networks with prior anti-fracking movements suggests a coordinated, recurring pattern of organized opposition to energy-intensive infrastructure.

The thing that I just can't understand for the life of me is why Anthropic is still funding these groups that want to put the kibos on new data center construction... the number one thing slowing down the growth of anthropics revenue it's not demand I think it's the availability of compute in data centers. And so you're just kind of wondering like what is the point of all of this?

Pointing out that Anthropic funds advocacy groups that restrict the very infrastructure Anthropic needs highlights a potential contradiction between its stated safety mission and its commercial interests.

The Democrats aren't going to pause the data centers forever. They're going to pause them until they feel like they're in enough control that they can dictate all the rules... that's when we get this you know big government democrat defined AI regime and you know that it's going to consist of a new regulatory agency and new speech controls the whole trust and safety agenda from social media will be ported over.

This theory of the moratorium as a political delay tactic — positioning for future regulatory control — connects AI infrastructure policy to broader battles over speech, content moderation, and tech governance.

Prior to Russia Today existing in the US, there was no anti-GMO sentiment... And you can actually see this on the Google trend data that shows GMO and it's kind of write up. And then as Russia Today started to get cut by different media outlets and people stopped retweeting them... the anti-GMO sentiment declined in the US.

The GMO/Russia Today case study offers an empirical model for how foreign-originated media campaigns can manufacture domestic public opposition to beneficial technologies.

OpenAI published a blog post called PRC linked influence operations are targeting AI debates in the US... Basically what they are saying and in fact many people are saying is that China is behind a lot of these influence campaigns to shape US attitudes on AI data centers... If they can stop us from building this necessary infrastructure, then that's a way for China to win the AI race.

OpenAI's direct attribution of anti-AI influence operations to China elevates the data center debate from domestic politics to a geopolitical competition with national security implications.

We have a huge moral panic going on with respect to AI... we're on the threshold, I think, of destroying the crown jewel of our economy, which is the system of free market innovation that we have... And we're on the verge... now we're talking about how far the Overton window has moved where we're actually saying that creating a FINRA for our industry might be better than all the alternatives. FINRA is a bunch of stock brokers writing rules. And when's the last time there was ever any innovation in that sector?

Comparing a potential AI regulatory body to FINRA crystallizes the incumbency-capture risk: self-regulatory organizations historically entrench existing players and stifle competition.

Number one, brand yourself as a safe AI company. Number two, ban unsafe AI. Three, profit. That's the strategy. Kind of brilliant.

This cynical three-step framework suggests that Anthropic's safety advocacy may function primarily as a competitive moat — a serious charge worth examining against the company's public statements.

They started with AlphaFold and they used AlphaFold to find a protein that could bind to CML and activate an enzymatic process that would break it down... And they took actual human skin from elderly patients that had donated their skin and they put this enzyme onto that skin and they were able to eliminate 55% of the CML on the skin which basically reversed the skin's age down to the age of a 31-year-old.

Using AlphaFold to design a novel enzyme that reverses a key molecular marker of aging in human tissue is a landmark demonstration of AI-accelerated drug discovery.