The AI Safety movement, a billionaire-funded cult which literally believes “AI will cause human extinction” is trying to slow down AI progress, ban it entirely, and surrender our lead to China.

They are doing so by writing intentionally deceptive, vague, and contradictory laws whose end goal is to make doing AI research as difficult as possible.

The State AI Law Moratorium currently in front of Congress is about stopping this. Everything else is a rounding error. It’s not about social media, elections, or deepfake porn. In fact, the US Congress passed a law banning deepfake porn, which was probably illegal anyway, so any “expert” telling you it’s about deepfake porn is intentionally lying to you.

In 2022 and 2023, the AI Safety movement spent hundreds of millions of dollars propping up NGOs like CSET and CAIS which intentionally tried to sabotage American AI development. Through hundreds of millions of dollars in bribes to the Biden campaign, Dustin bought his own regulatory agency, the AI Safety Institute. The Biden admin went on to enact policies to slow down American AI research and development, like Biden’s AI diffusion rule, which limited American AI hardware and AI model exports to allies like Switzerland, Portugal, and Israel.

Why would the Biden administration talk one way and act the other?

The AI Safety movement has a concerted strategy of lying about its intentions. MIRI, one of the foundational organizations of the AI Safety movement, admits that this is the organized strategy of most major NGOs in the AI Safety movement.

The main funder of all this is Dustin Moskovitz, Biden/Harris’s largest 2024 donor. He is a major reason why the Biden administration talked tough on supporting AI, but all his actions slowed down AI research. What Dustin said, Biden did.

Congress had a different reaction to the AI Safety doomsday cult. Nancy Pelosi and House Science Democrats said their beliefs had “little to no basis in evidence” when calling for Governor Newsom to veto SB 1047, a narrowly-stopped California bill that motivated the State AI Law Moratorium in the first place. A bipartisan coalition rejected the AI Safety cult bills during the Biden administration, including in the Democrat-controlled Senate.

All of this is to say that the debate over the moratorium largely exists to cover up what the moratorium is actually about. We are talking about everything other than the actual threat. This is a massive problem — and the intentional strategy of the AI Safety movement. They can’t persuade congress with their sincere beliefs, or even provide any evidence to back them up. What they can do is spend hundreds of millions of dollars distracting us, trying to sneak through laws in deep blue States that de-facto apply to all Americans (which they’re currently trying to do in New York).

That is why the only reason why we are debating social media, child safety, or the already-federally-banned deepfake porn. The concerted strategy of the billion-dollar AI Safety movement is to distract, lie, and sabotage.

Social media regulations or child safety regulations which do not single out AI will not be affected at all. There are plenty of existing state laws that address the concerns being debated now without unfairly targeting AI and without going as far as to ban AI.

Doesn’t that mean the State AI Law Moratorium is not strong enough? Why can’t the AI Safety cult use those levers to ban AI? The answer is that if they used those levers, they would have to ban the entire internet. Their radicalism would be completely exposed, with practical consequences for everyone. And no one — no Republican, Democrat, or Independent — would actually support that.

This example gives a clear illustration to why the State AI Law Moratorium hurts the people it should hurt, and doesn’t hurt the people it shouldn’t. The harms addressed by child safety laws are largely real — they’re real problems happening to real people, which we want to stop regardless of whether AI is involved or not.

Meanwhile, the harms addressed by the AI Safety cult are completely fake — a fiction with “little to no basis in evidence” according to a wide bipartisan coalition. Their “solutions” are attempts to use brute force to solve fake problems with the “side effect” of banning AI.

That is why we need to pass a State AI Moratorium.

https://arenamag.com/articles/mr-deepseek-goes-to-washington

One quote:

At no point in 2024 did I hear DeepSeek mentioned in Washington. Late January was when everything changed. A major lesson I’ve learned in the nation’s capital is that it takes around two years for academic and industrial common knowledge to migrate to the swamp.

So honestly, the DeepSeek information cycle came faster than I expected. But unsurprisingly, the pre-DeepSeek policy talking points stayed the same. DC remained fixated on export controls and infrastructure projects. The hard truth is that being two years behind means missing out on important chunks of reality. Remember — the DC wisdom was that two years was the difference between Chinese and American AI. As it turned out, it was the difference in time between reality and reality in Washington.

Summary on twitter:
https://x.com/psychosort/status/1930433938650181885

A one-tweet summary of my argument on science funding:

The most important thing I want Washington to take away from how DeepSeek caught them off-guard is this: however well-funded some AI companies are, there are still so many pieces of the American AI ecosystem that need funding - room to experiment - to take off.

Legibility is the foundation of their companies. “Because LLMs will drastically lower the cost of information, reading that information to begin with provides a huge advantage,” the stories continue. That’s all great in theory. Their proponents even have a few recent examples to justify their argument. Take human feedback data for ChatGPT or indexed search data for perplexity. Specialized data does enable real product breakthroughs — that much is correct. In 2022 or 2023, data moats appeared strong. Think about what that specialized data accomplished. In the case of ChatGPT, it created a “friendly assistant” persona that helps users interact with LLMs in a more intuitive way. In the case of Perplexity, it enabled LLMs to browse the internet and other large datasets.

Both of those are important breakthroughs, but if you’ve tried a selection of AI tools recently, you’ll notice that they are the opposite of a monopoly. Every major AI lab has successfully copied the “friendly assistant” persona. Almost all major AI labs now have a search integration. Even in the cases where specialized data created the most value, the features created by that specialized data spread far and wide — they became a de-facto industry standard.

To understand why people expected specialized data to be a moat, rather than a commodity, we have to return to the original data moat thesis. Originally, raw quantity of data was one piece of assembling a runaway “god model”, which was believed to become a natural monopoly for whoever created it first. This thesis collapsed as all AI companies suffered diminishing returns from scaling.

Afterwards, the data narrative flipped from quantity to quality. Specialized data became more valuable and a new data narrative emerged — that AI models would commoditize all repetitive work and the new data moat was turning unreadable, “soft” workflows into readable, automatable workflows. This data thesis had a more realistic expectation of AI model abilities, but blindly copied assumptions about data from the god-model system. In the god-model system, the role of data was to bootstrap a positive feedback loop that would develop into a runaway AI — a natural monopoly. In the modern system, the role of data was to enable moderately-intelligent AI models to do repetitive work on previously inaccessible systems. In the god-model system, data created a moat because of a short-term singularity, not something intrinsic to data itself. In the modern system, data is part of a long-term business equilibrium, which creates a moat because … ?

One early way that the data thesis failed was model distillation. Model distillation is a fancy term for AI models being trained on other models' outputs. This might happen intentionally as a strategy for companies to catch up, as OpenAI accused DeepSeek of doing to them in a congressional report. It can also happen accidentally, as more people post OpenAI model outputs to the internet, its style becomes part of the public set of training data. So the gains from specialized data in a model like the first version of ChatGPT became commonplace overnight.

Work through why these examples failed to become monopolies. Reinforcement learning from human feedback, the technique used to create the initial version of ChatGPT, was a genuine breakthrough in customization and UI. But why did that fail to create a moat? The answer is because once that style was crystallized, both in theory and in practice, it became easier to copy, not harder. For a natural monopoly like Facebook, the fact that Facebook already exists makes it harder, not easier, to build a copy of Facebook. Model data is the opposite of a natural monopoly.

As Martin Casado told me in a podcast last summer, “The marginal value of the next bit of data for the leader is actually lower than that for a challenger. So even without the distillation, historically in AI, we've seen this perverse economy of scale where the leaders take a lot more investment to make reasonable gains relative to challengers.”

Now, startups and legacy companies alike are repeating the same mistake as the leading labs. And they are trying to build their moat on a foundation even less stable than ChatGPT’s assistant persona — business processes. There’s a core truth to what they’re doing: it’s very valuable to make the hidden parts of business processes visible to AI models, especially as they become the limiting factor for AI automation.

But start thinking about model distillation and ask yourself for any given process, whether the second company that automates that process has it harder, the same, or easier than the first. The elephant in the room is classic model distillation — if a competitor gets its hands on your model output, your data moat is over. But even without model distillation, there’s only one reason why illegible process data is more difficult than collecting legible data, like we do from the internet or internally in modern companies — we haven’t figured out the system to collect or quantify it. The hard work of a data collection pipeline is in its design, engineering, and debugging. Once those tasks are done and the first version of a software product exists, it becomes easier to reverse-engineer, follow, and improve upon. Process illegibility is not a moat; it is a textbook example of an anti-moat.

To be fair, making this data legible creates a massive amount of value. The incentives for surfacing, collecting, and integrating data can be improved by creating strict legal contracts with large firms or tight patent laws around AI systems, though neither of these solutions are without drawbacks. However, once you’re talking about creating novel legal solutions to properly capture value, you’re already incredibly far from the natural monopoly paradigm.

Data is less moat, more bailey.

Let me ask a question: What’s the limiting factor in this process? Is it my willingness to write things? Is it the usefulness of things that I write? Or is it your willingness to read things? Let’s be honest — I know most of my readers don’t read every article. And even if you do, there are almost certainly comparably good writers you aren’t reading.

LLMs are the most voracious readers of all. They really have read every article on my Substack and every podcast transcript. And for that, they can better summarize my arguments, refer more people to my writing, and better integrate my worldview into my queries. Compared to friends who don’t write or don’t publish publicly, it is noticeably easier for me to get LLMs to write in the way I want. The benefits to being publicly known are so vast and practical to regular users of LLMs. This improvement is not limited to the style and content of writing; it includes more easily creating images, diagrams, code, workflows, and business processes similar to whatever you or others have written about. With every piece of information you publish, you are making LLMs serve you better.

I’ve seen this comparison made live dozens of times in parties, dinners, and groupchats. I called this public intellectual privilege, as a joke, but the name stuck. Now, I’m constantly asked to explain it.

The concept of public intellectual privilege is simple: your public data helps you. Believe it or not, this is a controversial belief. For example, take Simone and Malcolm Collins, who have posted dozens of hours of their children playing. “It’s for their own good,” Malcolm told me once. They’ve told me that most people disagree — and some have gone as far to excoriate them for posting about their children. But the Collinses are right. And for children in the age of widely available LLMs, they couldn’t be more right.

Peter Thiel famously asks founders “What is one true thing that almost no one believes?” In his book Zero to One, he provides a lesser known addendum to this question: “A good answer takes the following form: Most people believe in x, but the truth is the opposite of x.”

One of the best answers for x is this: “with rare exceptions, it benefits us to hide personal data.” The truth is the opposite: “with rare exceptions, it benefits us to share personal data.”

I have greatly benefitted from the era of public legibility. I hate networking. I hate explaining the same thing over and over on a hundred sales calls. I despise repetition.

The internet had shifted to scales in favor of marketing over sales. Instead of ‘schmoozing’, I write these articles, and I’m incredibly grateful for that. I can write well, write them once, and thousands of you will read them.

With the advent of LLMs and open-source AI, the shift from personal sales to personal marketing is being taken to its extreme. When you write for the public, you influence LLMs and their users. For some time, this Substack was the number one result for “Diminishing Returns in Machine Learning” on Perplexity. Due to algorithm changes or competition, that’s no longer the case, but I wish it still was. This influence is both exoteric — visible to everyone — and esoteric — visible only to those who read carefully. Sometimes, LLMs directly link, quote, or cite my articles. At other times, they reference stories, ideas, and framings I’ve written about without name-dropping.

Information is a massive public good. By sharing these ideas with more people, LLMs make it easier for users — or the LLMs themselves — to build upon these ideas and apply them to the real world. The person to benefit the most from those downstream ideas could be yourself. An LLM that recognizes your beliefs, experiences, and preferences can better help you. Practically, it has a better understanding of what you want when you ask it for a piece of software, a business plan, or a travel itinerary. The more complex the process, the more availability of context matters. Don’t make my word for it, take the world of companies collectively spending billions of dollars adding their own internal data to versions of LLMs. When it comes to writing, it’s uncontroversial to say that making your writing available to humans helps you, and more people are coming around to the idea that making your writing available to LLMs helps you too.

I’m interested to see whether more companies take the logic of the internet to the same extreme as me by open-sourcing their data to train LLMs. Some technical arguments suggest that making data available from the start of training is more efficient than existing post-hoc modifications like Reinforcement Learning From Human Feedback (RLHF). Of course, most companies have no practical way to train a model from scratch, so their only option for the former would be to make their data available for the base model available to everyone.

However, this general idea is much more controversial when it comes to information about your private life. That is where it is most important to get this prediction right — It will be a massive personal advantage to put your preferences, your needs, and your worldview on the internet.

In 2010s era tech politics, the debate raged over consumer data. As the story went, “Big Data” was ruining your life by collecting detailed information about how you use their apps. There were real downsides to Big Tech worth addressing — collaborating with the surveillance state, censorship, mental health problems, and locking out competitors. But at this point, the practical benefit of being known to the world — to people and increasingly to LLMs — vastly exceeds the cost. Laws like GDPR that seek to limit, not fairly expand, data access are blowing up a massive public good. These laws’ supporters — self-appointed guardians of the consumer — are hurting the people their laws are intended to protect the most.

If a company knows more about you, it will offer you better, more useful positive-sum deals. This was already an inconvenient truth about companies like Amazon. Laws that made it harder to access and keep user data just resulted in less convenient search results. The reason is simple: questions are complex. The more historical data a person or algorithm has, the fewer generalizations or false assumptions it will make. When you ask ChatGPT to order packages, book an airline ticket, or write a piece of software, the same rule applies. And it's infinitely more convenient to have that information available from the start than to have to specify these details over and over again.

The word “privacy” in political terms has become corrupted — and now means something completely different than its historical usage. There are really two meanings of “privacy” — privacy from the NSA and privacy from cookies. There is and always has been a real concern for privacy — privacy from your government. There are even fair criticisms of tech companies over this real concern, which have cooperated with government censorship and surveillance. The modern privacy movement behind laws like GDPR do less than nothing to protect our freedoms from our government. They do exactly the opposite.

Public intellectual privilege is true for civilizations as much as it is for individuals. The more information is available — about America, about the liberal West, about Western civilization — the more LLMs will be pulled towards us. For individuals, public intellectual privilege pays in better recommendations, better instruction-following, and better personal experiences. For civilizations, public intellectual privilege pays in better institutions, better social fabric, and a better historical record.

By scrubbing themselves from the internet, many countries are choosing to imprison themselves in an unrecorded past. They deserve their suicide. Pray that your civilization will choose to be part of the future instead.

Me, in City Journal:

The report’s authors even suggest that, without global controls on AI research, AI systems could cause human extinction: “Eventually it finds the remaining humans too much of an impediment: in mid-2030, the AI releases a dozen quiet-spreading biological weapons in major cities, lets them silently infect almost everyone, then triggers them with a chemical spray. Most are dead within hours; the few survivors (e.g., preppers in bunkers, sailors on submarines) are mopped up by drones.” It’s a forecast more grounded in severe mental illness than empirical reality.

For a long time, the overlap between factual descriptions of the motivations of AI Doomers and acceptable political language was close to zero. I’m glad that’s changed, Goldwater rule be damned.

As promised, I explain the concept of S-curves to the City Journal audience. Though this likely isn’t news to regular readers of this newsletter.

The history of innovation offers reason to doubt the safetyists. Every major research field in history has eventually reached a point of diminishing returns, where useful discoveries become increasingly rare. The pattern holds across disciplines—physics, chemistry, and economics—and is visible at a more granular scale, too, from antibiotics to railroads to CPUs. Initial bursts of progress taper off as fields mature, following an economic pattern known as the “S-Curve.”

Despite gains in computing power and economies of scale, fundamental AI research is getting harder, not easier. The core problem is that the remaining unsolved challenges are growing more complex faster than our tools are improving.

Finally, I direct readers to ’s wonderful cost analysis of currently existing AI regulations.

In practice, these laws work as jobs programs for lawyers. Rinehart found that up to 1,140 compliance hours per business may be necessary for a single regulation, the BIS reporting rule. Dozens of these regulations, applying to tens of thousands of companies, compound into billion-dollar compliance industries. The regulations function as an annual tax of hundreds of thousands of dollars, if not millions. They raise fixed costs, preventing new startups from competing and benefiting incumbents.

Exformation

How much might AI legislation cost in the U.S.?

ICYMI: Utah is considering a bill that would force social media companies to share social graph data. My recent op-ed in the Deseret News, and its accompanying tweet thread, explains this bill and why social media interoperability isn't the best path forward. I also had…

Read more

a year ago · 9 likes · Will Rinehart

April 5th, 7pm: New World Dinner in NYC: https://lu.ma/70e6f8wo

April 8th, 6pm: AI Bloomers 9 in DC, with : https://lu.ma/3738vjuh

And have fun with the new OpenAI Ghibli filter :)

1. Your actions matter

Most people are not thinking very hard about how to do something, even the thing they want to do. They are instruction receivers. Giving them plans to do their job well almost always creates a meaningful improvement, whether that’s administrative enforcement, legal language, or coordinated timing in media/backdoor pressure.

2. … But not always

Nothing I did decided the 2024 election or Chinese AI releases. Sometimes things just go your way and there’s nothing you did for it. Sometimes you are just downstream of fortune.

Your work schedule should reflect that. Be ready to work 80 hours one week even if it means you work 30 hours the next.

3. Stereotypes are correct

… when it comes to DC and SF. People from both cities should take what they think of the other, and double it. DC is twice as procedural, charismatic, and fake as SF thinks it is. SF is twice as autistic, reckless, and ambitious as DC thinks it is.

4. Funding doesn’t scale with winning

It scales with name recognition (for good or for ill) and event planning, mostly.

5. Right-wing status is more random than left-wing status

Who matters in the right is up in the air. As a result, you’ll often have totally marginal randos in group convos with the VP, Cabinet officials, etc.

Variation is probably good for philosophy and governance. But sometimes someone will need to tell a senile boomer nobody to let the Cabinet nominee speak.

Multiple people tell me that Democrats are much more class-conscious and exclusive. I’ve never really verified this.

6. It pays to be a philosopher. And if not a philosopher, than an ideologue.

You’re more trustworthy if you have a coherent framework of what’s right and wrong, even for people who disagree with you.

7. The Mafias are real

Most things people say about the mafias —Catholics, the asylum rights lobby, farm lobby, Ukraine lobby, Israel lobby, gays (left and right), degrowthers, EAs, Straussians (east and west coast), etc. — are mostly true. They are very powerful, within their side.

8. Neocons are even more evil than you think

They will make stuff up with no basis in evidence. They will make stuff up about you. They engage in the cattiest, socially destructive behavior. Just based on personality alone, I want the war lobby to suffer. This includes the war with China lobby.

9. The Tech Right vs. Trad Right fight is brewing

Every week or two, some people would message me from one of these sides, asking me to explain the other side and what their interests were. Everyone’s on edge. Both sides seem pretty clearly defined to me, but they clearly aren’t to each other. There was relatively little debate on AI and at least at my last job my policy was always that I don’t weigh in on coalitional fights unrelated to AI. If you ask again now, the odds are I’ll say more.

10. Almost everyone is competing for slots on a trough

That’s just how most funding in DC works. Because people compete for slots on a trough, conformity is encouraged, even on the right. This means people on the right will snipe at you for offending even the left. Seems like a perverse incentive. This is more of a systems problem than a problem with the individual morality of people involved, who are better than average.

This won’t change until the funding changes. I had the privilege of being funded by tech people and could piss off the typical conservative donor base or people related to the typical conservative donor base. This is a rare privilege.

In this new policy environment, the needs and challenges of Alliance for the Future have changed. That is why I am stepping down from my role as Executive Director.

I am incredibly proud of everything AFTF has done. AFTF has played an important role in defeating SB 1047, ensuring the National Defense Authorization Act did not threaten open source AI, advocating for the repeal of the Biden Executive Order on AI, advocating for the revision of OMB memoranda limiting public sector AI use, and charting out a strategy for AI diplomacy. We have succeeded in standing up for open source AI and little tech.

I offer my thanks to everyone I have worked with in and outside of AFTF. Let us enter the new golden age.

Brian Chau

To talk about what exciting new projects I’ll be working on next, please email me at hi [at] brianchau.ai.

Why the timing of the DeepSeek freakout was irrational:

The December paper was the more impressive one. There are other attempts to replicate R1 open source. There’s this attempt at UC Berkeley that was very interesting. And [R1] was less surprising to many people than the December paper.

The December paper is where the 5.6 million cost comes from. It’s where the big cost savings all come from. And it was just sitting there and it didn’t really get press. I don’t think it really affected the market until January. It affected the academic world, I think. I think a lot of academics were taking it seriously. And I would assume some of the companies as well. But it didn’t really move things until the second step released.

I explain why the second Trump AI EO targeting two OMB memos for rescission is so important:

The second thing was they announced their new executive order on AI, which rolled back more of the red tape or aims to roll back more of the red tape and took particular issue with two memos from the Office of Management and Budget. [OMB has a] controversial position now in the media because that’s one of the institutions Doge is trying to use to cut costs. But it also has jurisdiction over what’s called procurement. And these are rules for government purchases of AI tools, AI software and hardware, which can warp what government and what government contractors end up doing.

So lots of big tech companies—Meta, Microsoft, Google—they’re all government contractors. And so this was one of the ways that they tried to meet these equity goals using the government contractors. And Donald Trump also said we’re going to review these OMB memoranda and essentially going to replace them.

On the origins of the modern counter-bioterrorism movement away from targeting state actors and towards targeting American citizens:

After 9/11, the approach to policing and the approach to counterterrorism really changed. It changed innocent until proven guilty to guilty until proven innocent. This is now the controversy over various programs that were used in the post-9/11 surveillance apparatus to target terrorists.

There are some claims that it’s now being used against American citizens as well. And, of course, also claims at the time that it was being used against American citizens, the latter of which I believe is true. There are now court cases that are ruled in the defendant’s favor where that was the case.

And this led to an approach where people became very uncomfortable with their fellow man. They thought, what if my neighbor is a terrorist? What if my neighbor is a domestic terrorist? What if my neighbor is going to create a bioweapon?

I think that’s, number one, factually not true. We have Google once again. And number two, I think that’s a really dangerous philosophy. I think that philosophy has led to a lot of the overreach we’ve seen in the past years, and including the overreach on AI. I should be clear, it explicitly has led to the overreach on AI. This is the motivation that some people cite.

And it’s also led to a kind of decline of trust. And a decline of trust not just in terms of political agreement. Like some people say, there’s a decline of trust, people don’t like governments from the opposing side. But no, I think this is much worse. It’s much worse to believe that your neighbor is going to be a terrorist than it is to believe that the political party you don’t like is bad. I think the latter we’ve had a lot of precedent for. The former we haven’t really had a precedent for in America. And the global precedents have not really ended very well. The global precedent is like the Troubles or all these periods of immense political violence, which I don’t think have ended well.

Here’s how he began the speech:

I'm not here this morning, uh, to talk about AI safety, which was the title of the conference a couple of years ago. I'm here to talk about AI opportunity. When conferences like this convene to discuss a cutting edge technology, oftentimes I think our response is to be too self conscious, too risk averse.

But never have I encountered a breakthrough in tech that so clearly calls us to do precisely the opposite. Our administration, the Trump administration, believes that AI will have countless revolutionary applications in economic innovation, job creation, national security, healthcare, free expression, and beyond.

And to restrict its development now will not only unfairly benefit incumbents in the space, it would mean paralyzing one of the most promising technologies we have seen in generations.

Vance is exactly right. The Biden administration’s AI policy was heavily influenced by the AI Safety movement, a movement almost entirely funded by billionaires Sam Bankman-Fried and Dustin Moskovitz. They expressed in no uncertain terms that they should “pause AI research” because they believed AI would cause “human extinction”.

Many people have asked me about the involvement of top AI companies in pushing this “Doomsayer” narrative. When I have explained the true extent of collusion, people are often skeptical. The mood is now shifting. People are reckoning with the degree of corruption and misdirection in AI policy. Later on in his speech, Vance hits the nail on the head:

I'd ask if you step back a moment and ask yourself, who is most aggressively demanding that we, meaning political leaders gathered here today, do the most aggressive regulation?

It is very often the people who already have an incumbent advantage in the market. And when a massive incumbent comes to us, asking us for safety regulations, we ought to ask whether that safety regulation is for the benefit of our people or whether it's for the benefit of the incumbent.

The SBF Strategy

Let me be completely unabashed about what was happening. During the Biden administration, Sam Bankman-Fried and his crypto trading firm FTX pioneered what I call the “SBF strategy”. The SBF strategy is to go in front of congress, advocate for aggressive regulation, particularly industry-influenced regulators and licensors. The goal of the SBF strategy is twofold: direct regulation away from policy that harms incumbents and towards policy that harms competitors. The worst part is, he would’ve gotten away with it if his firm didn’t go bankrupt. That’s the main takeaway Effective Altruists have from SBF’s downfall. Unfortunately, it also happens to be true.

What does crypto policy have to do with AI? The ties are very clear. Before the collapse of FTX, SBF was the largest sponsor of the AI Safety movement directing Biden’s policy. Court filings show Bankman-Fried was a major investor in Anthropic, the AI firm which has most aggressively advocated for “AI Safety” policies and regulators like the AI Safety Institute. It is rumored that Bankman-Fried was a mentor to Anthropic CEO Dario Amodei. What we know for sure is that he was Anthropic’s lead Series B investor.

In crypto, the SBF strategy died with his arrest and conviction. But it lived on in AI policy. After the collapse, AI Safety’s #2 billionaire Dustin Moskovitz picked up the tab for the exact same organizations Bankman-Fried helped start. From the perspective of anyone outside the movement, there is no distance between their beliefs. That is why there is a symbiotic relationship between NGOs calling for the most aggressive AI regulation and a major AI company.

In short, JD Vance was right. Reality backs up his statement in a way that is stronger than anything people outside of the AI industry can imagine.

Another area that Vance nails is the subversion of American companies to foreign regulators:

The Trump administration is troubled by reports that some foreign governments are considering tightening the screws on U.S. Tech companies with international footprints now. America cannot and will not accept that.

And we think it's a terrible mistake, not just for the United States of America, but for your own countries. The U.S. Innovators of all sizes already know what it's like to deal with onerous international rules. Many of our most productive tech companies are forced to deal with the EU's Digital Services Act and the massive regulations it created about taking down content and policing so called misinformation.

Like the previous example, it is difficult to overstate the degree to which Vance’s words are true and, if anything, underestimating the extent of subversion.

The US had previously had a policy of entering joint partnerships with foreign regulators and directing US companies to comply with foreign regulatory authority that they did not have at home. This will sound familiar to those working in social media, surveillance, and “disinformation”. What the Biden administration was unable to do to American citizens under its own authority, it did under the authority of our allies. Moreover, those EU and UK offices may have been illegally lobbying American state and federal lawmakers.

Due to the influence of SBF and Moskovitz, the EU and UK have engaged with the US in ways that actively harm our AI industry’s competitiveness to China’s benefit. With Trump and Vance in power, that changes now.

As I wrote in January for Pirate Wires:

Biden’s original executive order subjected American companies to global AI governance. Specifically, the EU and UK targeted Americans, both by targeting American companies with their laws and attempting to influence American AI policy. Having failed to produce leading AI companies of their own, they have resorted to controlling AI’s future by censoring and controlling American-made AI models.

This must end on day one. Every foreign AI partnership must be oriented towards European obedience to American standards and technologies, not the other way around. Foreign governments should not be given special access to American algorithms, data, or products unless they intend to advance American interests, rather than slow them down. Joint research programs, data-sharing agreements, and policy committees must be contingent on improving AI research and access, not restricting it.

I am confident that the Trump administration will withdraw from or renegotiate AI treaties to put Americans first. Ahead of the US’ new AI action plan, Vance gives a shoutout to little tech and the importance of startups in generating new innovations.

 Now, this administration will not be the one to snuff out the startups and the grad students producing some of the most groundbreaking applications of artificial intelligence. Instead, our laws will keep big tech, little tech and all other developers on a level playing field. Now, with the president's recent executive order on AI, we're developing an AI action plan that avoids an overly precautionary regulatory regime, while ensuring that all Americans benefit from the technology and its transformative potential.

This dynamic was crucial to Chinese firm DeepSeek’s recent algorithmic improvements.

Amazing things are happening. Hopefully, we can get back on track and continue the long arc of American technological dominance. We will continue to depend on little tech and open source for those innovations.

Donald Trump, JD Vance, David Sacks, Michael Kratsios, Sriram Krishnan, and many others in the new administration are expanding the overton window of what is possible and what truths can be openly discussed. We are beginning to fully uncover and reckon with what actually happened to AI under the previous administration.

I asked OpenAI Deep Research to explain itself. The result is impressively factual, at least at an initial read-through. I did catch a few errors, which are identified in footnotes, but it is impressively few errors.

More specifically, I asked for “a thorough report on the technical breakthrough in hardware, algorithm design, product engineering, and logistics necessary to create OpenAI Deep Research.” I have edited it briefly to reduce redundancy, but the factual content is largely unchanged.

The reports generated by ODR can be heavily repetitive and seen artificially long. Ideas spill out of their designated section and are repeated (sometimes in ways that are not relevant) in other sections. The excessive use of brackets for examples in run-on sentences also bothers me. Both these elements remain to some degree in the edited version, but are heavily edited out. There was some amount of topic drift outside of the specified topics.

Enjoy the following, informative of ODR in both form and content!

1. Hardware Advances

Early AI Hardware (Pre-GPU/TPU): Before GPUs and TPUs became dominant, researchers experimented with various specialized hardware for AI. Early vector processors and co-processors were used to accelerate linear algebra operations central to neural networks. Intel’s ETANN in the late 1980s used analog circuits for neural computations​. By the early 1990s, all-digital neural network chips (e.g. the Nestor/Intel Ni1000) appeared, and digital signal processors (DSPs) were repurposed to speed up tasks like optical character recognition​. Field-programmable gate arrays (FPGAs) were also explored for neural network acceleration as early as the 1990s​, offering customizable hardware pipelines, though they often traded off maximum throughput for flexibility. These pre-GPU advances established the principle that domain-specific hardware could dramatically boost AI workloads by exploiting parallelism and low-precision arithmetic, a lesson that paved the way for later GPU and TPU innovations.

Graphics Processing Units (GPUs): Originally designed for rendering graphics, GPUs in the 2000s were retrofitted for general-purpose computing, excelling in the matrix and vector operations needed for machine learning. A modern GPU contains thousands of smaller cores arranged for massive parallelism, a large on-board memory (VRAM), and a fast memory hierarchy optimized for throughput. GPU manufacturers introduced AI-specific architectural features. For example, NVIDIA’s Volta architecture (2017) added Tensor Cores that perform mixed-precision matrix multiply-accumulate operations, delivering up to ~125 TFLOPS on 16-bit calculations in a single chip​. These innovations dramatically increased training speed by performing many multiply-adds in hardware concurrently. GPUs also leverage high-bandwidth memory (HBM in newer models)1 to feed data to the cores quickly, and use programming models like CUDA to let developers optimize memory access patterns and parallel execution. Thanks to their programmability and an existing ecosystem from the graphics world, GPUs became the workhorse for neural networks​.

Tensor Processing Units (TPUs) and ASICs: TPUs are application-specific integrated circuits (ASICs) developed by Google specifically for neural network workloads. First deployed in Google’s data centers in 2015, the TPU v1 was tailored for the inference phase of deep learning. Its core was a 65,536-unit systolic array (matrix multiply unit) operating on 8-bit integers, achieving a peak of 92 trillion operations per second, backed by 28 MiB of on-chip SRAM for fast data access​. By stripping out general-purpose features (caches, branch prediction, etc.), TPUs sacrificed versatility in favor of determinism and efficiency. The initial TPU proved to be “about 15×–30× faster at inference than the contemporary GPU or CPU” (Nvidia K80 and Intel Haswell), while delivering 30×–80× higher performance per watt​. Such gains came partly from the TPU’s streamlined dataflow design and aggressive low-precision computing. Subsequent TPU generations (v2, v3, v4) incorporated support for training (using bfloat16/FP16 for higher numeric range), much larger on-chip memory and high-bandwidth off-chip memory (HBM), and massive scalability via specialized interconnects between chips in a TPU pod. TPUs are cloud-hosted and optimized for Google’s software stack (TensorFlow XLA compiler). They are essentially hardware-as-a-service for AI. Other companies have similarly built AI ASICs – e.g. Amazon’s Inferentia for inference and Trainium for training in AWS data centers – aiming to outperform general GPUs by focusing on the matrix/tensor operations common to deep learning​. These ASICs exemplify the trend of vertical integration, where the hardware is co-designed with machine learning algorithms for maximum efficiency.

GPU vs. TPU – Performance, Cost, and Scalability: GPUs and TPUs represent two different approaches to AI hardware:

• Raw Performance: TPUs tend to have an edge in raw throughput for dense tensor operations. For example, a single TPU v3 core can deliver 123 TFLOPs for BF16 multiply-add, comparable to or higher than a high-end GPU, and Google reported order-of-magnitude gains in throughput per dollar for TPUs on large neural workloads​ (infoq.com). However, GPUs have narrowed the gap by introducing similar tensor accelerators and by excelling at tasks requiring flexibility or high precision (e.g. scientific computing or custom operations). For many models, modern GPUs achieve training speeds on par with TPU pods when using optimized libraries.

• Cost & Ecosystem: GPUs benefit from economies of scale and a broad market. They can be deployed from a single desktop up to supercomputer clusters, and the GPU software ecosystem (CUDA, PyTorch, etc.) is very mature. This makes GPUs highly adaptable – researchers can experiment with new model types without waiting for new hardware. TPUs can offer lower cost-per-training for large production workloads, but they are less accessible for small-scale use and require using Google’s platform. Cost also depends on utilization – a TPU pod is cost-effective when fully utilized for large training jobs, whereas idle time or smaller jobs might waste its capacity.

• Scalability: Both GPUs and TPUs scale to massive clusters, but the strategies differ. GPU clusters often use high-speed interconnects like NVLink and InfiniBand to connect dozens or hundreds of GPUs; Nvidia’s DGX SuperPOD, for instance, uses InfiniBand to ensure 1600+ GB/s cross-node bandwidth for scaling to thousands of GPUs. Google’s TPU pods, on the other hand, have an ultra-fast custom mesh network connecting up to thousands of TPU chips, allowing near-linear scaling on training jobs designed for TPU infrastructure. In practice, TPUs can be easier to scale for very large training runs because the hardware and software are designed together. GPU clusters can also scale well but may require more engineering by the user.

• Adaptability: GPUs are general-purpose processors. Aside from neural nets, they can accelerate graphics, physics simulations, or data analytics. This versatility means a GPU investment can be repurposed across different workloads and GPUs readily accommodate new model architectures or dynamic neural network operations that weren’t anticipated by hardware designers. TPUs, in contrast, are more specialized for matrix-heavy neural network patterns. Within their domain TPUs are programmable. They support many network architectures via high-level TensorFlow/XLA code. Moreover, Google continues to broaden their capabilities each generation. In summary, GPUs offer broad adaptability and a huge community/stack, while TPUs offer brute-force efficiency for mainstream deep learning tasks.

2. Algorithm Design

From Early AI Algorithms to Transformers: The evolution of AI algorithms has been marked by a series of breakthroughs that increased model expressiveness and scalability. Early AI models in the mid-20th century were limited by computational power and algorithmic understanding. The introduction of backpropagation in the 1980s enabled multi-layer neural networks to learn complex functions, leading to the first wave of deep learning (e.g. LeCun’s CNN for handwriting in 1989). Recurrent neural networks (RNNs) and their gated variants (LSTMs, GRUs in the 1990s) brought sequence modeling to the forefront, proving effective for speech and language by maintaining state across time steps.2 However, RNNs suffered from sequential processing constraints – they process one token at a time, making it hard to parallelize and capturing long-range dependencies was tricky even with gating mechanisms​ (mchromiak.github.io).

In the mid-2010s, the attention mechanism emerged as a game-changer. First used alongside RNNs in machine translation (Bahdanau et al., 2015) to allow models to focus on relevant parts of the input sequence, attention opened the door to better context handling. The culmination of these ideas was the Transformer architecture (Vaswani et al., 2017), which eschewed recurrence entirely and relied solely on attention to model global relationships in sequences. By encoding the position of tokens and using multi-head self-attention, Transformers can attend to different parts of a sequence in parallel, overcoming the bottlenecks of RNNs. This parallelism meant that Transformers could be trained much faster on GPUs/TPUs than RNN-based models for the same sequence lengths​ (mchromiak.github.io).

Within just a couple of years, transformers became the foundation of most state-of-the-art models in NLP, vision, and beyond, owing to their scalability and superior performance on long-range dependencies. They made major algorithmic breakthroughs: Multi-head attention allows the model to learn different types of relationships simultaneously. Positional encoding injects order information without recurrence. These were key enablers for this transformer revolution​. These innovations, along with techniques like layer normalization and residual connections, allowed training extremely deep networks that converge faster3 and generalize better, setting the stage for today’s large-scale models.

Key Components Enabling Transformers: A few specific innovations were crucial for modern transformer-based networks.

(1) Scaled Dot-Product Attention – a mechanism that lets the model weigh the relevance of different tokens to each other, with a scaling factor to keep gradients stable. This idea, combined with multi-head attention, means the model effectively has multiple attention “subspaces” to capture different aspects of similarity in the data.

(2) Positional Encoding – since transformers have no built-in notion of word order (unlike RNNs which process sequentially), Vaswani et al. introduced adding sinusoidal position embeddings to token representations, giving the model awareness of sequence positions​. This allowed the attention mechanism to consider relative positions.

(3) Feed-Forward and Residual Layers – each transformer layer includes a position-wise feed-forward network and uses residual connections and layer normalization, which help train very deep architectures by mitigating vanishing gradients and stabilizing learning.

(4) Parallelization Strategies – transformers significantly reduce the number of sequential operations needed to relate two distant positions in a sequence. In RNN-based models, the number of steps to connect tokens grows linearly with their distance. Transformers reduce this to one attention pass regardless of distance​. This property, combined with parallel computation of sequence elements, means training time can be dramatically shorter for long sequences. Replacing recurrence with self-attention “leads to significantly shorter training time” due to the ability to parallelize sequence processing​ (mchromiak.github.io) .

Additionally, researchers developed better optimization techniques (like Adam optimizer, learning rate schedulers) and training tricks (dropout, initialization schemes). While not specific to transformers, they enabled stable training of very large models that would have been unstable before. The transformer architecture’s success is a prime example of algorithm design co-evolving with hardware capabilities. It trades off some computational intensity (O(n²) attention) for much greater parallelism, which is a good trade in the era of abundant GPU/TPU compute.4

Evolution of “Chain of Thought” Reasoning: A recent algorithmic development in AI is the concept of chain-of-thought (CoT) reasoning, particularly in large language models. Instead of providing an answer directly, the model is encouraged to generate a sequence of intermediate reasoning steps – essentially, to “think out loud.” Wei et al. (2022) demonstrated that simply by prompting a sufficiently large language model to output a step-by-step solution, one can significantly boost its problem-solving capabilities​. This was surprising because it did not require changing the model’s architecture – it leveraged the model’s latent knowledge when guided properly.

The CoT approach improves problem-solving efficiency because the model can break a tough problem into smaller chunks, reducing errors at each step and allowing backtracking if needed. It’s an active research area, with work showing that chain-of-thought methods can lead to emergent abilities in very large models that smaller models do not exhibit​.

3. Product Engineering

Best Practices for Large-Scale AI Software: Engineering around large AI models (such as GPT-3-scale transformers) requires disciplined software practices to ensure reliability and efficiency. Teams now adopt MLOps practices – an extension of DevOps for machine learning – to streamline the model lifecycle from development to deployment. MLOps involves automation of data pipelines, reproducible training runs, model versioning, CI/CD for model deployment, and continuous monitoring of models in production​ (developer.harness.io).

Challenges in Training and Tuning at Scale: Training large AI models brings unique engineering challenges. The sheer scale of data and parameters means that distributed training is often necessary – no single machine has enough memory or compute. This requires strategies like data parallelism (split batches across GPUs), model parallelism (split the model itself across devices), or pipeline parallelism (chaining model segments on different hardware) – often all three in hybrid forms for trillion-parameter models.5 Sophisticated frameworks have been developed to automate these sharding strategies, but engineering oversight is needed to handle issues like synchronization, communication overhead, and fault tolerance. Google’s GPipe (2019) demonstrated how pipeline parallelism can train giant models by partitioning layers across accelerators and using micro-batches to keep all partitions busy​. Such techniques require careful orchestration to ensure that each batch of data and the model partitions are in the right place at the right time. Engineers must also optimize the training throughput by tuning things like batch size (too small and GPUs underutilize, too large and convergence might slow or memory overflows).

Deployment, Inference, and MLOps: Once a model is trained, serving it to end-users at scale is another engineering feat. Large models often need to run on clusters of machines with accelerators to handle high query volumes with low latency. Best practices here include efficient serving architectures and model compression. The latter uses techniques like knowledge distillation, quantization, or sparse pruning to reduce model size and speed up inference.

Inference optimizations like using half-precision or INT8 quantized models can dramatically cut costs. Many industry deployments now run neural nets in INT8 where accuracy permits, since it doubles the throughput on compatible hardware. From a software engineering standpoint, deploying AI models involves a robust CI/CD pipeline: new model versions should go through automated integration tests.

MLOps also covers monitoring and maintenance: models in production need continuous monitoring for data drift, performance drift, and even adversarial or unexpected inputs. If anomalies are detected, an automated pipeline might trigger a model retraining or fallback to a safe model. Automation is key – leading AI firms have continuous training systems where models are periodically retrained on fresh data and redeployed, much like how software is continuously integrated and deployed​ (ml-ops.org, developer.harness.io). All these engineering practices ensure that large-scale AI models remain reliable, accurate, and efficient as they move from research to real-world products.

4. Logistics

AI Hardware Supply Chain Challenges: The rapid growth of AI has put enormous strain on the global hardware supply chain. Cutting-edge AI training and inference rely on advanced semiconductors (GPUs, TPUs, ASICs), which in turn depend on a complex, global semiconductor manufacturing pipeline. In recent years, demand for AI chips has surged – Deloitte projected AI chip sales would account for 11% of a $576B semiconductor market in 2024, with generative AI and LLMs driving many enterprises to acquire GPUs by the thousands​. This surge (over 20% increase in demand year-on-year) is straining the supply chain, leading to chip shortages and long lead times for acquiring hardware​. A few factors make the supply fragile:

(1) Concentrated Suppliers – a large share of advanced AI chips are manufactured by TSMC in Taiwan or Samsung in South Korea6. Any disruption (natural disaster, geopolitical tension) affecting these manufacturers or the specialized fabs that produce 5nm/7nm chips can create global bottlenecks. The sector “relies on a few key suppliers… any disruption can create significant bottlenecks, delaying production and impacting the entire supply chain.”​

(2) Complex Production Process – producing high-end GPUs/TPUs involves dozens of steps across different countries (design in the US, fabrication in Taiwan, packaging and testing elsewhere).7 Production can halt due to shortages in critical materials like silicon wafers, photoresist chemicals, or neon gas for lasers. During the COVID-19 pandemic and subsequent supply crunch8, lead times for GPU orders stretched to over 6-12 months, affecting not just research labs but any company relying on that hardware​ (logicalis.com, techrepublic.com).

(3) Geopolitical Risks – export controls and trade disputes also play a role; for instance, recent regulations on chip exports have limited access to top-tier AI GPUs in certain countries, which not only impacts availability but also prompts efforts to develop indigenous AI chips. To mitigate these issues, governments and companies are investing in diversifying and shoring up the supply chain. Initiatives like the US CHIPS Act (2022) earmark tens of billions of dollars to build new fabs in the US. However, building new semiconductor fabs is a slow process – it can take 2–3 years and billions of dollars to get a new plant online, and even then, ramping up yield for cutting-edge nodes is nontrivial ​(datacenterpost.com).

Data Center Infrastructure Constraints: Building and operating the data centers that power advanced AI is another logistical challenge. AI supercomputing clusters (like those used for training GPT-4 or other large models) pack thousands of accelerators together, which creates extraordinary demands on power and cooling. Energy consumption is a major concern: training a single large model can consume megawatt-hours of electricity. For example, GPT-3’s training is estimated to have used ~1,300 MWh, equivalent to the annual power usage of 100+ U.S. homes​ (weforum.org). Data centers must be designed to deliver this power (often tens of MW for an AI cluster) and remove the corresponding heat. This has led to specialized cooling solutions, like liquid cooling plates on GPUs and even full immersion cooling for servers, to allow dense packing of chips. From a facilities standpoint, companies often choose locations with cheap electricity and cool climates for AI data centers to manage operating costs and sustainability concerns​.

Additionally, bandwidth and networking inside these clusters are a limiting factor. Distributing a training job across hundreds of GPUs/TPUs requires extremely high network throughput and low latency9. Architects use high-bandwidth switches, and sometimes novel network topologies (e.g. Fat-Tree or Dragonfly networks), to ensure each node can communicate at tens or hundreds of gigabits per second. Communication overhead can eat into scaling efficiency, so researchers have to optimize communication patterns to fully utilize big clusters. Another constraint is data storage and pipeline: feeding terabytes of training data to thousands of accelerators without stalls requires parallel storage systems (like NVMe RAID arrays or distributed file systems) that can stream data at dozens of GB/s. If the I/O can’t keep up, the expensive compute sits idle. Many AI datacenters now employ high-throughput flash storage and caching to pre-load datasets into local SSDs or even GPU memory. All these considerations mean that scaling AI is not just about more GPUs – it’s about balancing compute, memory, networking, and storage. As one illustration, NVIDIA’s DGX SuperPOD design notes that each node has to sustain >40 GB/s I/O to not bottleneck the GPUs​. Ensuring such performance across an entire cluster is a major logistical feat, requiring careful planning of data center layout, power distribution, and network architecture.

Emerging Solutions and Future Trends: To address these logistical hurdles, the industry is exploring several promising directions. On the hardware supply side, one trend is the development of chiplet-based designs. Instead of one large, monolithic die (which is harder to manufacture at high yield), companies like AMD and Intel10 are building chips out of multiple smaller dies (chiplets) connected by high-speed interfaces. This improves yield and flexibility – different chiplets (compute, memory, I/O) can be mixed and matched. It could alleviate some supply issues by allowing more modular production. Another approach is wafer-scale integration: Cerebras Systems famously created a wafer-sized chip (over 80,000 cores on one huge silicon wafer) to accelerate AI, eliminating off-chip communication for certain workloads. While niche, it shows the appetite for novel form factors to speed up AI. In networking, there’s work on optical interconnects and silicon photonics to eventually replace or augment copper links, which could dramatically increase bandwidth and reduce latency between nodes, easing the data movement problem.

There is also a push toward distributed training across data centers. If one center doesn’t have enough capacity, frameworks could in theory utilize resources from multiple locations. However, network latency makes this challenging for synchronous training.

Another future trend is algorithmic: reducing the need for brute-force compute via smarter training methods. Techniques like sparsity (pruning models)11, low-rank approximations, and progressive training aim to cut down the required compute without sacrificing results. If successful, these could relieve pressure on hardware and infrastructure by making AI models less hungry for resources. Lastly, the industry is acutely aware of geopolitical considerations – there’s a focus on building more resilient and geographically distributed supply chains. This might mean more chip fabs in different countries, standardizing certain components to be interchangeable, and maintaining strategic stockpiles of critical materials.

In summary, while today’s AI boom is taxing the logistics of compute, a combination of technological innovation and strategic planning is underway to ensure that OpenAI Deep Research-scale projects remain feasible. The path involves not just more powerful chips, but smarter algorithms, better software infrastructure, and robust planning for the “nuts and bolts” that underpin AI at global scale​. Each breakthrough in hardware, algorithms, product engineering, or logistics brings us a step closer to truly ubiquitous and sustainable advanced AI systems, enabling researchers to push the boundaries of what AI can do.

The Original Output

https://chatgpt.com/share/67a3a20c-d2d4-8005-92a7-feae93cb9b1e

Bonus: the o1-pro assisted factcheck did not help

7

Importantly missing lithography in Netherlands and Japan?

https://www.asml.com/en

The Asianometry Newsletter

How Japan Won the Lithography Industry

This is an older video. If you want to watch it, it is below…

Read more

3 years ago · 33 likes · Jon Y

While the internet is relitigating export controls on China as a result of the new DeepSeek model, they’re missing the forest for the trees. The Biden administration put export controls on most countries in the world, most of which are allies. These export controls have much less effect on adversaries like China and much greater costs to Americans.

This push to restrict AI exports weakens America’s alliances with the developing world at a critical juncture. Developing nations are currently making decisions on AI infrastructure that will have decades-long consequences. Competing for these markets depends not only on having more advanced AI models, but making those models available, customizable, and cheap. With its lead in AI, America is presented with an opportunity to reinforce its diplomatic and economic ties. Instead, these export controls risk surrendering the developing world’s AI infrastructure to China.

An AI provider constrained by this export rule is severely limited in their ability to run and adapt AI models in “Tier 2” countries. The rule restricts both the model weights comprising top AI models and the hardware required to run them locally. Moreover, it substantially increases the regulatory and compliance burden which, in turn, increases the cost of exporting American AI to the Global South. Meanwhile, a Chinese company expanding freely into the developing world can lower its costs, increase R&D spending, and refine its AI models with local feedback.

Competitiveness rests on economies of scale. When fixed costs are high, serving a larger market dramatically lowers the cost per customer. The cost of training AI models is rapidly growing, with GPT4 pre-training costing over 100 million dollars according to OpenAI CEO Sam Altman. The cost of operating trained AI models, known as inference cost, is significantly lower.

China has set a strategic goal to sell to the Global South. They hire local talent, speak the local language, and shape services to reflect local values. Localizing AI models requires rapid feedback loops, updating local models with new information, a process that is severely limited by the restrictions on both weights and hardware. The strongest systems evolve through trial and error in real-world environments. No matter how brilliant the engineers in California are, they cannot iterate effectively if they lack on-the-ground data in Nairobi, Jakarta, or Bogotá, all cities in “Tier 2” countries. Companies that choose to sell directly to these nations learn about their users, gaining feedback in a critical period for AI adoption. They fine-tune systems, adapt pricing structures, and modify user interfaces. By restricting that opportunity, we cut ourselves off from those markets in the long term. Even if American AI models maintain their technical advantages, lack of localization may lead to developing countries choosing Chinese alternatives.

By targeting local cultures in Africa, China built a near-monopoly on critical minerals in the region. China’s strategy for 5G networks has given them a global advantage. Allowing them to do the same for AI will have permanent consequences for the developing world’s security and economic development.

Relying on infrastructure built by adversaries introduces novel security risks. For example, Chinese 5G networking technology may introduce cybersecurity vulnerabilities, enabling Chinese surveillance and infiltration. The same considerations apply to AI. When American firms supply AI infrastructure, they ensure that the AI systems used by American allies are consistently updated to modern cybersecurity standards. Most importantly, they can prevent China from introducing backdoors, which only they are able to access. If America doesn’t provide AI infrastructure, Chinese companies will. By distributing safe and inspected American solutions, we can prevent our allies from becoming dependent on black-box systems from abroad.

Another overlooked advantage of technological exports to the Global South is soft power. When American firms invest in a country’s digital infrastructure, they become part of that nation’s modernization. The norms of their society evolve alongside their technology. They train developers, collaborate with universities, and sponsor research. The local workforce develops skill sets aligned with American standards and tools. Even if governments change, those relationships endure at the commercial and human levels. The new rule on AI exports undermines that possibility, leaving space for China to become the default partner.

What the current export controls get right is making an exemption for open source AI models. Localizing models to countries with fewer economic resources requires making models available, customizable, and cheap. The open source ecosystem excels at all three. Open source AI models reduce barriers to use, modification, and cost-saving optimization. However, China’s open source models are highly competitive with American ones, if not surpassing them. Winning the competition for global AI infrastructure will require making strategic decisions to advance and promote American AI, not hinder it.

In a week, DeepSeek shattered media narratives around AI, China, and open source. It went from a niche technical release on inauguration day to a national press cover story.

Whenever conventional narratives are challenged, both hype and cope distort the truth. Here’s what we know about DeepSeek and what it means for sacred narratives.

Introduction

DeepSeek R1 is a “reasoning” model. What does that mean? It means that it modifies a base model, DeepSeek V3, to answer more difficult questions using a sequence of outputs called “Chain of Thought”. Many of DeepSeek’s most notable optimizations, around algorithm structure, memory management, and data type, occurred with the base model V3, released in December.

V3 was competitive to Western base models, but R1 put DeepSeek into an exclusive club with OpenAI and Google, passing Meta, xAI, and Anthropic. OpenAI released the first “reasoning” model, o1, in September. DeepSeek is the third company to release this type of model, after OpenAI and Google. On many metrics, it surpasses Google’s “flash reasoning”.

Are DeepSeek’s Results Real?

DeepSeek claims similar performance to o1 on a variety of tasks.

ARC Prize has independently verified R1’s performance, which scores similarly to o1. LMSYS, a benchmark based on live feedback from real users, ranks it higher than everyone but OpenAI and Google. R1 has also been rapidly adopted by scientists and engineers. Open source models like R1 are overwhelmingly favored by researchers, because they are much easier to modify. Western open source models like UC Berkeley’s Sky-T1 have also successfully replicated near-o1 performance.

The balance of evidence favors DeepSeek’s results being real. While only a few independent verification attempts have been made, they have all stood up to scrutiny.

Is DeepSeek’s Cost Real?

DeepSeek V3, the base model behind R1, claims to have a total training cost of roughly 5.6M USD. Several people have raised skepticism about this cost for various reasons. Scale AI CEO Alexander Wang claimed that DeepSeek had smuggled 50,000 H100s in violation of US export controls.

There’s no clear way to verify the training cost of DeepSeek V3. The closest thing we can do is attempt replication using similar techniques and spending. Some attempts are already being made. Replication can prove that similar results/cost is possible, but it can’t definitively prove that the results are impossible.

Export Controls

In Washington DC, a raging political debate was whether AI hardware export controls were worth the cost. Both sides seem to want to shoehorn DeepSeek’s algorithmic improvements into this debate.

The export control critics argue that the export controls caused Chinese firms like DeepSeek to innovate. On the narrow point, they might be right, but that’s not enough to declare export controls dead. Consider the equilibrium. When American firms have a hardware advantage, they can integrate the innovations of Chinese firms with an asymmetric advantage. This may not be true for a narrow set of H800-specific improvements, but it is true for the majority of algorithmic improvements made in DeepSeek V3.

What export control critics get right is that China can’t be stopped in place. They will always be capable of improving. If America wants to maintain its lead, American firms must keep innovating and integrating new discoveries.

Three Stories to Explain DeepSeek

The DeepSeek papers are concise (53 and 22 pages respectively, even including references and appendices)! For clear descriptions of implementation and results, just read the paper.

The trouble with “just reading the paper” is that the significance of any or all of these improvements is very unclear to a general audience. As bittersweet as this is, I’ll introduce three general stories that capture the context for DeepSeek’s improvements. Of course, stories are imperfect and will not perfectly represent the technology or history involved. Hopefully, these stories are better than the alternatives.

The first story is a continuous story of model architecture, or what most people consider to be the AI model “algorithm”. RNNs, Transformers, and Mixtures of Experts models form the rough “evolutionary chain” of this progression. The direction of this story is towards models that are faster, cheaper, and more accurate. The first two of these factors are based on complexity, a rough correlate to how long those models take to run. Accuracy can sometimes be in tension with the other two. By shrinking the model, AI researchers often increase model speed while decreasing accuracy. Sometimes, the increased speed allows for longer training, which can more than compensate for the loss in accuracy. The challenge is to change models into more efficient versions in a way that does not significantly harm accuracy.

To that end, DeepSeek introduced several optimizations: Auxiliary-Loss-Free Load Balancing, Multi-Head Latent Attention, and Multi-Token Prediction.

Auxiliary-Loss-Free Load Balancing is an improvement upon existing Mixture of Experts model architecture. It solves a known problem, Load Balancing, in a more efficient way. Multi-Head Latent Attention is a more efficient way to implement the standard Multi-Head Attention step that does not seem to have a significant cost to accuracy. Multi-Token prediction is a technique to increase accuracy by predicting multiple tokens in advance, at a greater cost per token. However, like lower cost per token can increase long-run accuracy, greater accuracy per token can decrease long-run cost.

The second narrative revolves around combating hardware limitations. The US export controls on H100s, or more specifically on chips with high-bandwidth memory, created an asymmetry in model loading costs. In short, DeepSeek pays more in dollars and time to load a model and update weights. Changes to DualPipe Pipeline Parallelism and Cross-Node All-to-All Communication fall under this category. I defer to the following Tom’s Hardware article:

DeepSeek used the DualPipe algorithm to overlap computation and communication phases within and across forward and backward micro-batches and, therefore, reduced pipeline inefficiencies. In particular, dispatch (routing tokens to experts) and combine (aggregating results) operations were handled in parallel with computation using customized PTX (Parallel Thread Execution) instructions, which means writing low-level, specialized code that is meant to interface with Nvidia CUDA GPUs and optimize their operations. The DualPipe algorithm minimized training bottlenecks, particularly for the cross-node expert parallelism required by the MoE architecture, and this optimization allowed the cluster to process 14.8 trillion tokens during pre-training with near-zero communication overhead, according to DeepSeek.

A related but more generalizable improvement is low-precision training. A common practice throughout the lifetime of machine learning, at both the hardware and software levels, is optimizing the data type. Computers use a series of bits, either 0 or 1, to store memory. Typically, computers used 32 bits or 64 bits to store numbers (both integers and floating point numbers). As performance became more important, two things happened in sequence. First, many researchers realized that many of those bits could be removed without a significant cost to accuracy. This was low-hanging fruit, and optimized almost immediately once machine learning began to be taken seriously commercially. Later data type improvements involve doing much more difficult work to reduce bits without creating a significant loss in accuracy. Recently, that work has been focused on reducing the bits in a trained model for more efficient inference, a practice known as ‘quantization’. As the name suggests, low-precision training is a technique to apply those same low-precision adaptations to training.

We might also consider DeepSeek’s algorithmic changes as a response to their unique hardware situation.

A third narrative is the development of fine-tuning techniques that are used to modify AI models that have already completed the earlier step of “pre-training”. The progress in fine-tuning is often framed as the academic debate between Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO). DeepSeek has weighed in with a variant of PPO called Group Relative Policy Optimization. GRPO was invented by DeepSeek researcher Shao ZhiHong in 2024. According to the DeepSeek R1 paper, this change was made “to save the training costs of [reinforcement learning]”. Historically, fine-tuning training has not been significant, but that might be changing with “reasoning” models such as R1 or OpenAI’s o1.

Conclusion

Decisions matter. AI researchers who believe that decisions matter make better decisions. A culture that takes AI progress for granted will fail to make better research decisions. Open source is a check on this system, by giving more researchers the tools to make improvements, bust myths, and move the state of the art forward.

Representative quotes:

Administration officials including Michael Kratsios, David Sacks, Sriram Krishnan, Marco Rubio, and Russell Vought will have massive influence over how the new administration approaches AI.

This follows an electoral college and popular vote victory centered on American greatness: unleashing the best and brightest Americans instead of slowing them down. The message is clear: Americans will lead and accelerate on AI and across our economy. We will no longer be subjects of the global censorship and regulation machine.

Every foreign AI partnership must be oriented towards European obedience to American standards and technologies, not the other way around. Foreign governments should not be given special access to American algorithms, data, or products unless they intend to advance American interests, rather than slow them down. Joint research programs, data-sharing agreements, and policy committees must be contingent on improving AI research and access, not restricting it.

Read the article for much more.

In an unprecedented move Tuesday, Mark Zuckerberg of Meta released plans promising to “work with President Trump to push back against governments around the world going after American companies and pushing them to censor.” Going further, he argued that “Europe has an ever-increasing number of laws institutionalizing censorship and making it difficult to build anything innovative there.” This rare direct challenge to European regulatory authority signals a broader shift: Trump’s election has fundamentally altered the balance of power between American technology companies and their foreign regulators.

The calculus is simple. The European Union’s economies, governments, and security infrastructure all depend on American technology. While European regulators have leveraged this dependence to impose their vision of digital governance on U.S. companies, the incoming administration sees an opportunity to reverse this dynamic. At stake is not just corporate compliance costs or content moderation policies but whether the architects of the world’s digital infrastructure will continue to cede regulatory authority to nations that depend on these same innovations.

This realignment comes at a critical moment. Artificial intelligence (AI) is reshaping the world's economy, and competition between the U.S. and China continues to intensify. Therefore, the ability of American companies to innovate, free from foreign regulatory constraints, has now become a matter of national security. Even more importantly, American companies support American values, such as freedom of speech, and play a critical role in ensuring the long-term resilience of a digital and democratically aligned world.

While the EU may fear their dependence on the American technology industry, it is in their own interest for these firms to be strong enough to resist advances from techno-authoritarian adversaries. Even so, European regulators and bureaucrats still believe that they can curtail the “threat” of American technology companies by creating and enforcing increasingly draconian rules and regulations that are extraterritorial. This process gave rise to the idea of the “Brussels Effect,” which, while global in nature, clearly has American firms as the primary target.

For example, of the six companies affected by the EU Digital Markets Act (DMA)—Alphabet, Amazon, Apple, ByteDance, Meta, and Microsoft—five are American, while Bytedance is Chinese. None are European. GDPR imposes costly compliance requirements that disproportionately affect American firms. When combined with the DSA, which requires platforms to monitor and remove content that violates EU standards, these initiatives entrench the dominance of large firms that can devote significant resources to ensuring compliance.

As the Biden administration did not take technology seriously, a vacuum was created that effectively ceded regulatory control of the American technology industry to the EU. American firms paid hundreds of millions of dollars in fines to the coffers of the EU. Still, as the EU grew as a “regulatory superpower,” it struggled to create and sustain its own technology ecosystem. Unfortunately for Europe, their misinformed rush to regulate the American technology industry has brought them too close to the proverbial sun. Instead, the EU faces deteriorating levels of productivity, innovation, and technology development. President Ursula Von Der Leyen has promised to prioritize stopping “Europe’s 'slow agony' of decline,” support new European digital champions, and pursue strategies to further European “digital sovereignty.” However, such efforts are unlikely to succeed.

Now the focus of many an anecdote on how to not do innovation or tech policy, policymakers in Brussels are facing questions from leaders in the American technology community who are starting to question why countries that are not building or innovating at the cutting edge of technology have a say in how it should be built, used, or regulated. Increasingly, these companies have begun to refuse to release their new digital innovations to the EU market.

Expressing a sentiment held by many in the American tech community, Palantir CEO Alex Karp recently argued that “Every single relevant company in the world is in [America]” and that “Europe has basically decided to regulate its anemic and nonexistent tech scene out of production. All of those people want to come to America”. This sentiment closely mirrors the sentiment expressed in a letter sent to the Department of Justice in November by Senator Ted Cruz (TX) inquiring into possible Foreign Agents Registration Act violations by European non-profits influential in American AI policy. In it, he accused European nonprofits of attempting to “gain control over AI’s development, but to also gain control over information flow and a citizen’s ability to communicate with others free from government intrusion.”

It is now clear that President Trump’s election represents a new era for the American technology ecosystem. America will once again recognize that being an American ally requires sharing American values, first and foremost the defense of free speech. Both big tech and the European Union must adapt to this new political reality. The EU’s efforts to export its regulatory ambitions, which undermine core American values, will soon face a steep political cost. In the words of incoming Crypto and AI Czar David Sacks, “the timeline split, and we’re on a different path now.”

Under a Republican administration, proactive AI policy is likely to shift to national security, supply chain, and energy priorities. As the Trump administration shifts the tone of AI policy towards American production and efficiency, the House AI report indicates how those priorities might be implemented.

National Security

The national security section of the House report documents the thorough, bipartisan development and adoption of AI technologies in the Department of Defense (DOD). The entire history is too lengthy for this summary, but is worth reading for anyone interested in AI-related Defense issues.

Moreover, the report recommends iterating on existing oversight powers and training programs. In further sections, the House aims to bolster a patriotic vision which prioritizes American interests as a leading technological, economic, and geopolitical power.

Energy

The energy section is an exceptionally clear-eyed analysis of the coming sharp increases in electricity demand.

One important consideration for the broader economic impacts of rising electricity demand is uneven price elasticity. Short term fluctuations in electricity demand are ably handled by the current supply of natural gas. However, as demand sharply and permanently increases, longer-term solutions are become necessary. As the report documents:

The report correctly identifies the variable energy consumption of emerging AI hardware. Considering variability in supply, demand, and efficiency in setting future policy is another sign of the seriousness and effectiveness of the report.

Consequently, the report endorses further support for new energy construction and deregulation, both steps that are appropriate and effective for dealing with rising energy demand.

At Alliance for the Future, we are particularly interested in the consequences for the open source and AI startup ecosystems. In short: the report’s findings on open source AI and small business are excellent for those trying to build something new with AI. They signal openness to continuing research, open source publishing, and little tech. They also indicate a crucial commitment to evidence-based AI policy over hypothetical fears and cultural interests.

Open Source

The report underscores the importance of open-source AI as a cornerstone of fair innovation. By encouraging transparency and collaboration, open-source models mitigate the risk of monopolization and centralization. It acknowledges the economic incentives behind regulatory capture and are looking to promote open source as a counterbalance.

Regulatory Capture

Overregulation is a significant focus of the report, particularly of the section on small business. Economists have found that regulations often impose disproportionate burdens on smaller firms, hindering their ability to compete with large, established companies. This concept is known as regulatory capture. The house identifies regulatory capture as a key concern in the small business section.

Congress aims to prevent regulatory frameworks from entrenching the dominance of incumbents, recognizing that a balanced approach is necessary to sustain innovation across the industry.

Federal Pre-Emption

The only area of engagement where we at Alliance for the Future are looking for major change is federal pre-emption. Specifically, we hope Congress will pass major federal pre-emption legislation in 2025. We believe that realistically, there are only two outcomes: either Congress will pass federal pre-emption, or the most politically lopsided states will set AI regulation for the entire country.

The report accurately acknowledges that federal preemption remains politically difficult. While national policies could help counter inconsistent state-level regulations, such as those emerging from California, the combination of jurisdictions and interests makes assembling a coalition for pre-emption difficult.

Working with George Mason University’s Mercatus Center, Alliance for the Future has published a report on a pre-emption structure that could cross partisan divides. Federal pre-emption based on research freedoms would reflect the bipartisan coalition opposing SB 1047. The uniting features of that coalition were the dual pillars of evidence-based policy and the freedom to publish AI research, along with the models that research depended on.

California Bill SB 1047 was a narrowly defeated bill that would have de-facto banned open source AI and severely harmed closed source AI, under the premise of managing hypothetical CBRN risks at a state level. This bill was opposed by a wide coalition of academics, startup founders, and bipartisan elected officials. A main point echoed by nearly all criticis was the lack of evidence for the claims made by SB 1047 authors. Consider the House Science Democrats opposition letter, later cited by Speaker Emerita Pelosi:

Informed by the defeat of SB 1047, Congress is steering away from policies driven by speculative AI risks. Instead, lawmakers are focusing on real, demonstrable threats to address the industry's needs more effectively.

We hope to build upon that moment of bipartisan agreement to create a federal pre-emption structure that can be embraced by a bipartisan supermajority.

The report also contains thorough discussion of other crucial AI policy areas, such as energy and national security, which will be covered in a later part.

Overall, the House AI Task Force Report featured excellent recommendations for open source AI and small business. It identified regulatory capture or fear-driven legislation as threats which have led to ineffective or harmful policies in the past. Most importantly, it signals an ongoing commitment to evidence-based AI policy.

The formerly obscured connection between these two growth trajectories is unveiled by AI. Recall that the core incentive of the Paper Belt is procedural control. Ideas of justice and equality only come later, applied selectively as post-hoc justification. The Paper Belt’s latest trick is apocalyptic rhetoric. As the Road produces technologies with ever greater upside, the Belt produces an equal and opposite rhetorical reaction to justify its control. With commercial AI on the horizon, those narratives are being fabricated today, despite a total lack of evidence and several thorough debunkings. These narratives are directly applied to AI, used as a justification for the designation of all software as homo sacer through the unilateral authority to control AI. 

Read more

Feature engineering is good for the field of machine learning, good for the economy, and good for society.

Imagine you could only have one electric appliance. Any device that uses electricity had to be a washing machine, microwave, induction stove, and vacuum cleaner. This would be ridiculous, impractical, and absurdly costly. But why?

We can work backwards from reality. It’s practical to have different appliances for different things. We expect washing machines to wash, microwaves to microwave, and vacuum cleaners to vacuum. This works well in our everyday life. It works well because designing one electric appliance is easier than designing a huge omni-appliance. The engineers can do a better job. Engineers can specialize, learning more about one appliance because they have little need for knowledge about the others. In the omni-appliance, doing each task well creates additional constraints on the other tasks. The burden exceeds the costs saved by sharing resources.

It also works well because we understand how to use our appliances. We know which button to press for vacuuming, which button for microwaving, and which button for washing. The complexity of using an appliance is not entirely different than the complexity of designing it. For each individual appliance, each limitation to the design space is also a limitation to the use space. This network of constraints add complexity, inefficiency, and ambiguity.

Division enables simplification, for both design and use.

Despite the possibility of saving money by combining appliances into an omni-appliance that shares the same batteries, wires, cases, and screens, it is just not worth it.

All of this is to say that a design paradigm that dogmatically refuses to sacrifice general capability for specific application is both leaving free money on the table and leaving the public worse off. Of course, this is about AI.

Feature engineering originated as a technical term for designing an AI model to follow inputs designed for one pattern, or ‘feature’. But as AGI hype accelerated, it became an insult. The implication was that people making AI models that worked best for a narrow task were unserious, or even bad for the field. This is not only wrong, but the exact opposite of the truth. Feature engineering is the heart and soul of commercial success! It is synonymous with specialization. As Adam Smith writes in the Wealth of Nations:

It is the great multiplication of the productions of all the different arts, in consequence of the division of labour, which occasions, in a well-governed society, that universal opulence which extends itself to the lowest ranks of the people.

Moreover, techniques discovered for specific use cases often contribute to broader developments. This arguably includes many of the early predecessors to modern transformer models like GPT-4. For example, RNNs, a direct predecessor of the transformer, only had narrow use cases at the time of its invention and development.

As I wrote extensively in Arena Magazine, the belief that a single all-purpose AI model will be optimal or even inevitable comes not from reality, but from fantasies of monopoly rents.

The curse of every contrarian is for his views to be imitated by the crowd, twisted to the point of cliché. Peter Thiel believed he was stoking controversy when he wrote “Creative monopoly means new products that benefit everybody and sustainable profits for the creator. Competition means no profits for anybody, no meaningful differentiation, and a struggle for survival.” Now, regulators and columnists are assuming that all emerging technologies naturally become monopolistic. This simplistic view fails to account for the distinct qualities and dynamics of each sector. As a result, regulators confident that monopolies are inevitable are setting policies with those assumptions — and with complete disregard for the consequences for startups.

…

[T]here is rapid diversification throughout the AI pipeline. What was once a single, vertically integrated process done by one company is separated into specialized improvements. As funding becomes more available and demand grows for niche applications of AI, the machine learning process is being sliced into ever narrower, more precise subprocesses.

Don’t be deterred. Build the specialized AI models and applications you want in the world.

As the tension between these two cultures becomes clearer, it will unveil a mirror to a fundamental conflict within human nature. AI will reflect the patterns and incentives created by its algorithm, its data, its environment, and its users. It will reflect us, not only by emulating us but by unveiling parts of us that we traditionally ignore. 

The Paper Belt has a communal vision of human nature. In their eyes, a coming together, a consensus, and if necessary, consensus by censorship or force, is the way to govern best. It believes that left unchecked, capitalism and technology lead to exploitation, anarchy, or even extinction. To them, consensus is an act of altruism and empathy. Forced consensus is the ultimate form of Rawls’s original position, governing according to a benevolent egalitarianism. 

The Silicon Road is individualist. It believes that self-interest leads to social good, in most if not all cases. As economist Adam Smith noted, “It is not from the benevolence of the butcher, the brewer, or the baker, that we expect our dinner, but from their regard to their own interest.” Moreover, it is highly suspicious of the communal will. It sees the Road’s desire for consensus as an attempt to scapegoat and steal, motivated by blame rather than goodwill.

The question at the heart of this disagreement: why does political consensus exist? This question is famously deferred by Locke, beginning centuries of classical liberalism staked on not answering this very question. Paypal founder Peter Thiel emphasizes Locke’s importance in the American founding: “That new science found its most important proponent in John Locke and its greatest practical success in the United States, a nation whose conception owed so much to Locke that one exaggerates only slightly to describe him as its definitive founder.” Thanks to the Lockeian spirit, America paused metaphysical war for as long as metaphysical war was called religious war.

Future History

Read more