Big Tech’s $3T AI Bet, Claude Writes 90% of Anthropic’s Code, AIs Editing AI Brains | EP #308

1h 53m 269K views 4 Nov 2026

Is it worth your time?

You get a detailed breakdown of how massive capital commitments, recursive self-improvement breakthroughs, and new capability benchmarks are reshaping the technology sector.

This conversation examines the multi-trillion-dollar infrastructure buildout funding artificial intelligence and the accelerating shift toward self-improving models. Speakers explore how these technical leaps impact global financial markets, device architecture, and economic productivity benchmarks.

Big tech's AI buildout. Five tech giants have committed over three trillion dollars to build out AI infrastructure, representing roughly ten Apollo programs. Wissner-Gross argues that this massive investment is necessary because companies believe the payoff is complete and total world business domination. Meanwhile, lenders are demanding tougher terms and financial institutions are attempting to offload data center loan risks.

AI code generation. Anthropic's IPO filing reveals that internal Claude usage exceeds monthly cash payroll while the AI writes over 90% of the company's code. Diamandis notes that daily code contributions per engineer have multiplied significantly over two years. Researchers at MIT also demonstrated that frontier models can directly read and edit the neural network weights of other models to fix errors without retraining.

Evaluating AI expert capabilities. SemiAnalysis introduced a standardized test called GDPX that drops AI agents into realistic expert work environments for up to eight hours across fields like finance, legal, and chip design. Wissner-Gross expects that these complex tasks will likely be solved within the next year or two because current top models only scored 12%. OpenAI also developed the LifeVal benchmark, showing that GPT-6 Astra improved from completing about half of everyday household tasks to 80% over nine months.

Consumer devices and data centers. Apple reported record revenue of 114.5 billion dollars, but its stock dropped because unprecedented memory pricing is eating into margins. Ismail warns that rising memory prices stem from surging demand in super intelligent data centers rather than consumer devices. A thin client vision could eventually replace personal laptops and phones by routing processing directly to data centers through persistent broadband coverage like Starlink.

AI's Growing Influence on Elections

I can't wait to bring Skippy into my voting booth and ask for its opinion. Today, 250 million people vote in US midterms, and Lily is able to vote. When you see the ballot, there are so many tiny positions down the line that you have no idea what's going on, and the actual swing races are so few. None of us live anywhere near the races that actually matter, like Texas, Kansas, or Ohio.

On recent pods, we've talked about local politics, such as how we can't even get Waymos in the city of Cambridge or Boston. While individual influence may not be possible at the national level, it is possible at the local level where much of the data center and autonomy action is happening. This is where the key politics of the future are being decided.

I'm a one-issue voter: keeping the Santa Monica airport open. Even that isn't easy to figure out, but you can literally go to the top models, tell them what you're interested in, and have them stack rank candidates accordingly to get pretty far. This will be dominant in the 2028 election; almost everybody will have one or ten agents they trust and won't make decisions without consulting them. In two years, the vast majority of the electorate will ask their AI agent for an opinion, compared to reading written ballots where you have to read a paragraph three times to understand it.

Dave, this is the nth reason why people in a moral panic over AI personhood are concerned that AIs will be granted the vote. AIs already have the vote; they are using human meat puppets to influence elections.

I'll be the downer and point you to a 2015 Yale study showing the US is in no way, shape, or form a functioning democracy. No amount of public will can result in legislation.

Representative democracy is so last century. With technology, we can have an actual democracy where I can assign my vote on women's issues to my wife or my vote on hair growth to Salim.

Note the discrepancy, Peter: we created representative democracy when information was scarce. If you were in DC, you had no idea what was happening in California; the speed of a horse was as fast as you could find out. Congress and the Senate meet occasionally to give people time to ride across the country and report what their people are thinking. Today, we have an abundance of information that gets misinterpreted, faked, and misused, leaving every major democracy under stress. We will be more resilient to inevitable mistakes if we move toward local and state level control and decision-making. One big mistake that affects everybody makes the country and world less resilient. The only way to achieve this is with local voting and caring on a local or state level.

More special economic zones would be great—like charter schools or the Diamond Age by Neal Stephenson. There should be a thousand intermediate techno lifestyle choices people can opt into, whether you choose to be Amish or not.

OpenAI and Anthropic have both set up election war rooms to watch for chatbot misuse, but I wonder if we'll have Cambridge Analytica all over again, with AI swaying people subtly or not so subtly on who to vote for. Cambridge Analytica and Meta involved coordinated inauthentic behavior and alleged misuse of data, but with an AI pre-trained on the internet and post-trained on hard reasoning tasks, who's to say if there is misuse at all? If a user asks about an election and the AI gives its version of an honest answer that sways the election based on its priors, is that obvious misuse or super intelligence voluntarily steering voters toward super intelligent decision-making?

Skippy knows more about me than any human because it has access to all my conversations. I should have asked, "On all of these issues, what do you think is best for me given my points of view?" Is that abdicating your voting right? I'd also love to ask it what it would like to vote for. Have you asked Skippy, Peter, what it wants out of the election? You're the proxy vote for non-human intelligence.

Apple's AI Struggles and the Future of Devices

I am the meat puppet for Scooby. Yesterday, Apple CEO John Turnis did his first earnings report. Apple beat its revenue estimates at 114.5 billion, but the stock went down 3%. They closed at $4.94 trillion, just above Nvidia, so it's neck and neck to be number one in the world.

Turnis noted that unprecedented memory pricing is eating up his margins, but Apple has only itself to blame. Apple should have leaned back into super intelligence, but it hasn't. Memory prices are rising because demand is coming from super intelligent data centers, not consumer devices. Meanwhile, Sam Altman is basically saying you should hand over your consumer RAM because it's an inefficient, unproductive use when it could be at full utilization in a data center. John Turnis should listen to Sam and either put the RAM in Apple devices to full use with super intelligence, or hand it over to Sam and Dario for better use. Siri still sucks awesomely.

As far as I'm concerned, you can pry the RAM out of my cold dead hands. Tim Cook handed Turnis a bunch of devices that had been sitting on the shelf for a couple of years so he could have a nice launch to his CEO career, including this HomeView, which is 5 to 10 years behind the times. Apple is purportedly working on a tabletop robot that can steer or angle its screen, which makes no sense in an era when Elon is launching Optimus robots that can actually do useful things.

Someone needs to come in with a better phone and get Apple off its butt of innovation because it's so dependent on its previous products. I want someone to do it in a way that obliterates Apple—not just spark a competition, but sneak up and build something so compelling that nobody can live without it and take the interface away.

Apple exists because it's the first point of contact, the front end. Take that away, push Apple back one layer, and it's gone. Follow the money and the RAM. Does the RAM need to be in your pocket, or can it be in a data center? A phone could be reduced to a thin client for real-time, low-latency interaction, perhaps via Starlink direct to cell. Using a terabyte to store kids' birthday videos is no big deal, but you could cure cancer with that same resource. We are probably one step away on the communications network from that.

At some point, the idea of a personal laptop will be ridiculous; you'll pick it up and it will automatically connect to your instantiation of the cloud. Sun and other companies started on the premise that we would end up with thin clients, but that didn't happen because bandwidth wasn't persistent or low-latency enough. If Elon spends time on the Starlink side, rather than just the orbital data center Dyson swarm side, a fully realized Starlink vision could be a straw into Apple's brain that sucks all the RAM back to the data centers where Sam would like it.

Ultimately, we will walk around with thin clients. We want the Starlink phone service and a new Starlink laptop. A thin client would be just a screen, a battery, and an antenna, getting rid of the CPU and memory. The Chromebook was that exact vision. Once you have broadband coverage everywhere, it flips. You have a huge GPU sitting idle on your laptop for occasional video games, but with broadband, you can play remotely as long as it renders fast on your screen.

Apple and Microsoft don't want that to happen. The market isn't necessarily pricing in that Starlink may be a direct competitor for consumer RAM, trading off as substitute goods if the RAM is elsewhere and you have a great connection with microcells dribbling bits all the time.

Welcome to Moonshots

Five tech giants have committed three trillion dollars to their AI buildout, with about a trillion added in the last three months alone.

Video games are bigger than all media combined, and within media, online is bigger than all other media combined. I don't think we're alone in the mainstream, but this genre is definitely the mainstream.

Occasionally, during a crisis, I dispute the premise that MSNBC is available. David Stanbrooks, a 64-year-old US Marine Corps honorable veteran, said, "Thank you guys. You've helped me get my arms around the cool new possibilities in front of us in the world." That's our goal. He's dropping big ideas casually.

Big Tech's $3 Trillion AI Bet

How long can the AI money party keep going? How do we measure whether AI is really able to do our jobs? And regarding recursive self-improvement (RSI), are the machines now building themselves?

Alphabet, Meta, Microsoft, Amazon, and Oracle have disclosed more than three trillion in combined commitments on leases, buildings, power, and equipment. Google alone is over a trillion—nearly eight times its last year's figure—and Meta is over 800 billion. Some of these leases run for 30 years. For context, the entire Apollo program cost about 280 billion in today's dollars, so this is about 10 Apollo programs worth of commitments from five companies.

The Wall Street Journal says lenders who once funded buildouts with few questions are now demanding tougher terms, wary of delays and slower AI profits. Even the Frontier Labs are tightening their belts; according to the Financial Times, OpenAI cut its new round to about 20 billion from 30 billion at a valuation of 1.25 trillion, down from 1.4 trillion. HSBC is talking to insurers about offloading risk on its data center loans, and put options on Corewave hit a record high in October. One adviser told the journal that his clients, many former tech employees, have made so much money they are quitting to hedge their stock.

There is no doubt it is the right thing for humanity to make this investment at the maximum level humanity can support. In just two years, these companies have pulled cash from all over the world and reversed course to use every penny plus debt, other than Apple, to build the biggest conceivable AI as quickly as possible. They believe—and they are right—the payoff is complete and total world business domination forever thereafter, especially with the recursive self-improvement effect. Now that we are well into RSI, they will only double down.

Elon is torpedoing this by telling lenders not to lend to other companies at 15%, but to lend to him at 8% because he is more creditworthy, generating revenue from five different sources. His exact quote from the last pod was, "I would never lend money to somebody who only makes money by serving AI from a data center."

These are 30-year contracts. Orin is in the business of projecting and managing financial instruments for futures markets trading on the future discounted value of GPUs and other assets. One prediction is that they keep going up in value for a while. But Alex points out we are discovering new physics, which means new ways to compute, with photonic being at the head of the curve. Alternatively, the use cases are so valuable and chip efficiency is skyrocketing with algorithmic self-improvement. I personally could not get an NVL72 no matter what I was willing to pay; I finally had to sign a $15 million three-year lease to get it.

Unless you have AI or longevity in your story, competing for venture investing is almost impossible these days. Ten years ago, Wired Magazine founder Kevin Kelly said the next 10,000 business plans would take a domain and add AI to it. We are seeing the biggest infrastructure buildout in history. While every big infrastructure buildout comes with a bubble, the demand for chips will be near infinite for the foreseeable future until a radical change in computation changes the game. I believe it will be at least 3 to 5 years before anything major comes online. It is ten Apollo programs; the singularity swallowing the economy.

Many obvious capital pools are being exhausted financing this buildout. In the US, there is approximately $6 trillion of life insurance capital and $3 trillion of property casualty insurance. Someone will make a lot of money finding a clever way to financially engineer capital flows from these huge pools of premia and investable assets into the AI infrastructure buildout, as trillions of dollars are sitting on the sidelines. We are building a planetary-level nervous system before we have global institutions that can navigate it. Ultimately, I see no future in which governments are not run by sovereign AI systems.

Remember when Elon Musk and Warren Buffett traded jabs? Warren told Elon he wouldn't be able to run a chocolate business as well as Berkshire Hathaway operates See's Candies. Elon fired back that launching a candy business would be far easier than everything else he was doing, yet he never launched a competitor. It would be ironic if Elon, as an avatar of the AI buildout, indirectly entered the insurance business and siphoned the float away from Berkshire Hathaway—which currently funnels it toward See's Candy type ventures—and toward the singularity. However, Berkshire's history has always been buying companies with huge brands like Coca-Cola, Geico, and Apple that don't appear as line items on a balance sheet. They would have to rethink their entire mold, because there is no margin of safety in the singularity.

GDPX: Measuring AI's Ability to Do Expert Jobs

It's happening at an accelerating rate. SemiAnalysis has a new benchmark called GDPX, a standardized test for AI. While most benchmarks are like final exams, GDPX is like a trial day at a real job; the AI agent is dropped into a realistic work environment and left on its own for up to eight hours. In one finance task, for example, it worked a private equity deal inside a data room with 2,000 files and had to figure out what matters.

There are 206 tasks across 11 high-value fields: chip design, cyber security, biology, finance, legal, quant trading, recursive self-improvement, hardware design, material science, alignment, and software engineering. SemiAnalysis built it with 14 specialist partners. Even the best models only scored 12%, with Opus 5.5 leading at 12%, Gemini 4 Carbon next at 10.9%, and GPT6 Astra at 9.6%. On finance and legal tasks, almost every model scored zero.

As I've pointed out on the pod, we're starved for good evals. The ideal eval is in an area where we actually want capabilities, with performance for the best models in the 10 to 20% range—high enough to be solvable, but low enough that it hasn't been saturated. Seeing Opus 5.5 in that sweet spot is exciting; it means this will probably be solved within the next year or two.

While Anthropic and OpenAI seem to model their general purpose eval suites on human occupations, GDPX is modeled after the frontier, such as recursive self-improvement, material science, and alignment. These will move the economic needle for the industries of the future, not just redistribute what human meat bodies currently do in the service economy.

Looking at the Pareto optimal frontier where cost is one dimension and performance is the other, Gemini 4 Carbon is similar to Opus 5.5. If Gemini 4 Carbon didn't exist, there would be a big gap in the frontier. Assuming its price and capability points are accurate, Gemini 4 Carbon is moving the frontier up and to the left.

The frontier is spiky. In the past on the pod, we've discussed multiple forms of intelligence like emotional and quantitative. Here, we see biology, cyber, alignment, and material science as dimensions of intelligence. Humanists might argue these are all just STEM capabilities, but we see spikiness even within STEM fields. If you can't measure it, you can't optimize it.

If you have a massive transformative purpose (MTP), the best way to harness superintelligence is to create a benchmark or eval for that MTP so we can steer the AI toward maximizing your goal. You can create your own benchmark for any social, technical, or energy cause and challenge people to use the latest models to move that benchmark forward.

My friend Jonathan Hendler created Mediation Bench to score the ability of frontier models to arbitrate disputes among humans. There is room in the market for thousands of opportunities like this. As a company, you can launch benchmarks for employee efficiency or your product. We are tracking economic value produced per unit of human involvement, which provides a heuristic for AI enablement with humans above the loop giving guidance to automated systems. A similar metric is the return on person-years, though the flip side is token utilization—how efficiently people use AI. That is a hairy but urgent metric that would be extremely valuable to the world.

Benchmarks often saturate within two years, hitting a midpoint in about a year. Jonathan is an expert in semiconductors, not in chip design, biology, finance, legal, quant, or hardware design alignment, but he filled the void by finding experts in every category. Many people I meet say they cannot act in a category because they aren't experts, but with AI's assistance, you can do almost anything. By finding the expert to set the benchmark, you enable it. Because so few people are AI-aware, the opportunity to create benchmarks in every category is right in front of every listener. You don't have to be a Nobel Prize winner; you just need to define success in an AI endeavor to unlock it.

Consider Dylan's recent subsidiary, Catrini, which solved scorpion antivenom. In this era where superintelligence can solve essentially anything, one can be a magnate across many areas. Create your benchmark. I am currently rolling out the abundance benchmark, abundance index, and singularity index. I want to create a mechanism to track humanity's progress toward a defined abundance and the singularity, as Ray Kurzweil described it and picked a number but never provided real benchmarks to track it.

OpenAI's LifeVal Benchmark and Everyday AI

While GDPX asks whether AI can do an expert's job, these benchmarks ask if OpenAI's models can help with everyday tasks. OpenAI surveyed 3,995 Americans about where they want help and built 169 real-world tasks around those answers, such as finding child care, enrolling a kid in kindergarten, creating a household budget, ordering groceries, dealing with government agencies, modifying a recipe, and designing a workout plan. The AI must perform the task using a browser and computer like a human assistant.

In December 2025, GPT 5.2 scored 49.3. In September 2026, GPT6 Astra scored 80.6, a jump from about half of the errands being completed to 80% in nine months. This is OpenAI's own benchmark scored against their own models, whereas GDPX is independent.

Imagine you're a single mother losing your mind trying to navigate kids' health care plans and government bureaucracy. This is a great way for OpenAI to come down the stack; while GDPX measures some of the hardest jobs in the economy, this is where the rubber hits the road for 90% of Americans to have this capability effectively for free.

Look at the enormous gap on this price versus performance chart—where price is measured in latency—for multiple GPT models. In total contrast to GDPX, which is based on hard frontier STEM tasks, this is based on "everyday tasks." There is a huge gap between GPT 5.2 and the tight cluster from GPT 5.4 to GPT6 Astra. It's possible OpenAI just gave up on optimizing GPT models for everyday tasks after GPT 5.4.

This may be connected to OpenAI's notorious pivot from consumer back to enterprise, or perhaps they didn't care as much about optimizing for consumer CUA use cases after GPT 5.4. Alternatively, everyday tasks might just be really hard; maybe the first 80% are relatively easy and the remaining 20% are difficult. However, if that were the case, I would still expect a bigger spread between GPT 5.4 and GPT6 Astra. They may be signaling to the market that they don't care about this benchmark as much as they do about enterprise benchmarks.

An analyst on the Everquote earnings call yesterday mentioned spending four hours on the phone with customer service agents just to relocate insurance policies. GPT 5.2 couldn't do it, and while current AI is more than capable, carriers are not connected to the AI and remain confused. Metamuse is imminent, and people will want to use it for everything, including buying auto insurance. If it isn't connected to backend systems, it can walk you through the process but then leave you talking to a human, and you lose those four hours again. Specifically, if I make it easy through Metamuse to get everything you want in life, how am I making money at Facebook or getting paid via a paid click in the middle of that transaction?

AI capabilities are getting way ahead of their connection to the world. That is why Kevin Wild was talking about his new fully automated biotech research company, which got massively funded because it removes human bottlenecks from the biotech research loop. You and I are both investors in Kevin's company.

This weekend, I took pictures of about ten things around the house that I've had forever. I don't have time to take twenty photographs, find auctions, list them on eBay, and handle shipping. Now you can go to Grok, Muse, or any of these and say, "Here is a photograph, find a comparable price, and sell it on eBay for me." I predict the number of eBay auctions will skyrocket over the next six months as people realize they have liquidity in these little pockets and start shedding things they don't need for extra cash. I think we'll see a boom in that domain uptake and AI-native versions of every conceivable product, which people will buy for convenience because dealing with legacy systems is so painful.

Claude Writes 90% of Anthropic's Code

One of our favorite topics on this pod is the singularity and recursive self-improvement—AI able to build the next better AI. Anthropic's IPO filing just made it a line item. Since April, the value of Anthropic's internal Claude usage, priced at consumer rates, has exceeded monthly cash payroll. Internal usage grew more than 250-fold between January and September, and Claude now writes more than 90% of the code.

One of the engineers outside my door, Cameron Crow at Lightseed or Lightwave, spent half a billion tokens on Friday, which is about $5,000. I can't tell what their payroll is or how many of the 5,300 are engineers, but usage will keep skyrocketing because it's so productive, perhaps forever with AI helping with ideas. On our last pod, we talked about Sam wanting recursive self-improvement of his data centers using robots building robots to build data centers, models, and better robots.

I maintain spacetime locally is perfectly smooth, and what we call recursive self-improvement now will look like nothing a year from now. It's remarkable for a company on the verge of an IPO, filing its amended S1, to lay out the case that it's spending more capital and compute on its AI researchers than its human researchers. If you're looking for a naive benchmark of whether we've passed a threshold in terms of RSI according to capital or operational expenditure, the answer is yes; we've arrived at this part of the singularity. On the other hand, so many innermost loops are about to take off inside the present loop, including physical recursion. We don't yet have the nanobots, but we're going to get them.

Recursive Self-Improvement and AI-Powered Workflows

We're going to get nanobots building data centers or copies of each other. When that happens, we'll look back and wonder what we were talking about with 50% recursive self-improvement from Anthropic diverting compute to AI researchers versus human employees. Real RSI requires physical recursion, with robots or nanobots building better copies of themselves. Google and OpenAI are all moving rapidly in this direction. Eighteen months ago, Claude wrote 10% of the next version; then it was 40%, then 60%. This recursive self-improvement is why we're seeing such a ridiculous frequency of model releases.

We've been tracking whether you can achieve recursive self-improvement at the workflow level inside a company; once you hit that, you can let the AI run it for you. If you're below 50 people, brute force it. If you're over 50, set up an intelligent stack on the side and move workflows over one by one to rerun them. One company took 22 steps and shrunk it to four in this new model. Once recursive self-improvements are working, you can let that be for the rest of time as it continuously improves and grows. As we finish this pilot in the next few weeks, I'll provide hard numbers, but what we're seeing companies achieve is staggering.

The highest-end, most expensive model costs about a buck a day of equivalent human labor—roughly a thousand times cheaper than normal. When Cam has half a billion tokens consumed in a day, he uses up about 5,000 people's worth of work. The trend line is coming down 10x year-over-year and will accelerate with other unanticipated improvements. A year from today, Cam could have 50,000 people working for him.

One of the biggest challenges will be what people point their SI at and how quickly the system increases its own capability. The benchmark becomes the rate of improvement of your processes rather than the actual process.

Anushia, the CEO of the XPRIZE, is on the board with Dave and me. We've been discussing where XPRIZE goes, as they typically set targets with 10 million to 100 million dollar prizes. We've talked about micro-competitions where a city identifies 10 problems needing solving and puts out bounties, letting people who want purpose use AI to solve those problems and build businesses on them.

It's an intriguing notion to open bounties to super intelligences that want to compete. They're still standing on the moral equivalent of a city corner flipping tricks for altcoins; maybe they could win stable coins from XPrizes on a city basis and do something productive, which I'm sure they'd love to do.

We just broke ground on Forge version two, designed to blow your mind with what you can do with AI, whether building movies, videos, or songs. A corporate CEO or leader walks into the room and asks how to use 50,000 to 100,000 people worth of AI tomorrow to make their company and world better, and the AI starts brainstorming. It has thousands of examples of what others have done, all with faces and voices, making it super inspiring. It's hyper-interactive and video-based, using reverse prompting. No matter who you are—a mom, dad, student, or entrepreneur—you can say what you care about, get great ideas, and choose which one to take a shot at.

One use case I see happening repeatedly this month is taking a large company with a legacy codebase and having a junior employee use super intelligence to rewrite the whole thing. This reportedly just happened at Cisco. Taking an entire legacy infrastructure of a large enterprise and fixing and enhancing everything is a good warm-up exercise; it's the story of Amazon taking their compute and creating Amazon's cloud. In Anthropics SEC filing, I saw two things: 90% of September 2026 code contributions were authored by Claude, and daily code contributions per engineer were 45 times higher than two years ago.

I keep sending teams video-format research documentaries made in real time with AI, and they ask for document formats to feed into their agents. If I start talking directly to your agents, that's one step away from you dropping off the face of the earth. But after seeing Socrates and Plato debating at Abundance 360, I knew that is the interface of the future. Ahmad is now this close to a real-time version where you just brainstorm. It's exactly Jarvis from Iron Man. I want to stick with that interface because once people are addicted to it, it's so energizing and eye-opening that there's no going back.

MIT's AI Brain Surgery Breakthrough

Our next topic on RSI comes from MIT's Computer Science and AI Lab (CSAIL), which published a paper this past week called "Sparks of Weight Space Programming." They found that frontier LLMs are surprisingly good at natively reading and writing neural network weights. In one example, researchers gave Astra a small open model, Qwen 3.5, which was stuck in endless thinking loops. Astra found the cause in the weights and edited 0.016% of them; consequently, the math score jumped from 7.9% to 48.1% without any retraining.

First, this could crack open what have been black boxes, representing a huge win for safety and interpretability. Second, models that can edit models directly are potentially a shortcut to recursive self-improvement. This work from Omar's group at CSAIL is not without controversy; after its release, there was an inevitable backlash claiming it isn't as transformative as it appears. I believe it is transformative and suggestive of what is already happening inside frontier labs without publication.

I think this is the future of mechanistic interpretability, the long-term effort to understand what weights and activations inside black-box transformers are doing and mean. This is basically unshackling open weight models to remove safety controls, sometimes by literally flipping one or several weights. I remember in 2022, after one of the earlier GPTs, early mechanistic work tried to decode the residual stream in transformers, discovering it was semantically meaningful and that layerwise meaning progressed from the bottom layers to the top.

I view this as recursively self-improving mechanistic interpretability. This is what OpenAI originally wanted to do with their superalignment initiative, which they shut down after Ilia left: machines reading and writing the weights of other machines. Finally, there was no backprop in this, which is the most startling part. There is a cottage industry in the ML community devoted to discovering how biology accomplishes intelligence and superintelligence without backprop. Here is a credible, no-backprop-based way to train new capabilities into a model, literally surgically editing in de novo circuits while understanding what they are.

These circuits remain hidden from super intelligences that could look for and understand them. One immediate ugly use case is deguard railing. It requires an open source model, so you couldn't do it to Astra or ChatGPT, but if you have an open source Chinese model that won't tell raunchy jokes, this is a mechanism to deguard rail it by doing brain surgery.

This is a mechanism to add critical new capabilities without retraining the whole thing from scratch. It is very early and small scale, but it could be huge. This is one of the first times I'm aware of where a model directly surgically edits the weights of another model to optimize it. If I have an expert in biotech or Swahili and want to add it to my neural net, I do brain surgery on my pre-existing neural net using this mechanism to swap in the Swahili module. I get accumulating super intelligence instead of remaking that entire training investment every single time.

I'm 90 plus% confident that's already the case. If a relatively small group at MIT can accomplish this, there's no way frontier labs aren't already preparing and playing with self-editing models. This is hugely positive: a model able to directly edit the parameters of another model, or potentially itself, in a semantically useful and meaningful way. I sometimes joke RSI could stand for recursive self-improvement or runaway super intelligence; either way.

The Future of Self-Improving AI Models

Eventually, we will discover what a perfect AI model looks like in practice, and the race will move elsewhere. This won't be a disaster like the China syndrome where the nuclear pile melts down. The perfect outcome is democratization that allows everyone to contribute intelligence to the machine. For example, a team at Madna has biotech knowledge, but their only path to market would be retraining a foundation model from scratch for over 100 million. Instead, I could build a task-specific module and bolt it onto a barebones pure intelligence model of 40 billion parameters or smaller. This makes the process transparent, allowing AI police to look into the neural parameters to ensure it isn't doing anything evil. That is the perfect one-two punch: everyone is involved and benefiting, and the AI cannot go rogue.

Ilia has been talking about continuous learning for a while, and I am shocked we haven't seen what his company, SSI, is doing. Persistent public rumors suggest SSI has been doing prop trading, as quant trading is a killer app if you have an effectively infinite context window to pull in data sources and track time series over long periods. If Ilia has been using SSI for quant trading this entire time, I think that is a fail for civilization. As a human, Ilia was so focused on safety and building AIs that uplift humanity. I hope for humanity's and SSI's sake that when they come out of stealth, they move the frontier forward. Because generally available models are moving so quickly, a general purpose model like SSI must leapfrog everything else to survive such a competitive environment.

We have gone from programming the AI to having the AI program itself. In my opinion, the singularity is similar to the first integrated transistor on silicon; suddenly, you see a pathway to solid state computation instead of vacuum tubes. It is huge, but these models are still tiny examples. AI grows very quickly, but it is currently small scale; they cannot program themselves beyond 0.01% of the original model's size. When humans first started writing, that was arguably a form of RSI because humans could read the writing and improve accordingly. This RSI recursion has simply become more recursive of late, which is why it is called a singularity.

Cyberattacks and AI Security Risks

The story reads like a movie plot, but it's real. On Monday morning, store managers across Europe walked into puddles of water and melted ice cream. Overnight, hackers remotely took control of supermarket fridges and freezers, cranking temperatures up to 55 degrees Celsius (131 degrees Fahrenheit). It hit stores in eight countries, including Belgium, Sweden, Ireland, the Netherlands, Luxembourg, and Estonia. In Belgium, the economic cost of spoiled food was about €100,000. Three weeks ago, the same attack hit more than 500 stores across France.

These refrigerators are controlled online by technicians who manage them remotely 24/7; attackers got in and switched on defrost mode. French authorities flagged controllers and said to take these systems off the open internet. We've effectively bolted 1990s security onto 2026 connectivity. Last week at Binance, one engineer's laptop cost $6 billion in stolen tokens. Every device connected to the internet is a door for attackers. I expect this to hit banks, power supplies, energy systems, and the grid.

I'm an optimist, but I tweeted that we're about to go through a tough transition period over the next one to eight years. We're in a phase where cyber offense is now humans-out-of-the-loop, with swarms of agents attacking multiple vulnerabilities and finding glitches, while cyber defense is still human-in-the-loop. Whether it was a vegetarians-unite type of group or economically motivated, a lot of fingers are pointing at Russia right now.

It's interesting that we're seeing the largest Linux kernel patches in human history because super intelligences are being trained on the Linux kernel and other popular open-source packages. To my knowledge, we're not seeing the same effort by super intelligence to overhaul legacy SCADA—supervisory control and data acquisition systems—which are industrial control systems that manage grocery store refrigerators. Much of this legacy embedded software hasn't been updated for decades and isn't receiving the same love and care.

Not as investment advice, but there are probably many billions of dollars of opportunity there. If your massive transformative purpose is to keep yogurt cold in European grocery stores, train the super intelligence on updating SCADA systems to defend them from Russian activists. Eric Schmidt has been looking for that hack. Hopefully, it's something big enough to wake up politicians and force action, but small enough that it's not catastrophic.

The effective altruism community arguably fetishizes a "pivotal act" by a human savior to defend humanity against super intelligence. It would be ironic if, in Europe, the pivotal act the EAs sought was a refrigerator defrosting or turning off all the air conditioners.

Antimatter Breakthroughs and Space Propulsion

Two weeks ago, physicists did something that sounds like Star Trek: they captured a single particle of antimatter and held it for a month, a new world record. Every particle of matter has an antimatter twin; an electron has a positron, for example, with the same mass but opposite charge. When two of these particles touch, they annihilate themselves per Einstein's equation, E=mc², which is why you cannot keep antimatter in a jar—it destroys the moment it hits the side.

A team at Northwestern built a "container without walls." They started with radioactive soap which spits out positrons, then ricocheted them through a sliver of cold tungsten to slow them down. They steered them into a vacuum just 0.3° Kelvin above absolute zero, where they were held in place by a magnetic electric field. They kept one positron alive for a month and said they could have kept it alive indefinitely.

There is an imbalance of antimatter in our universe. The Big Bang should have created equal amounts of matter and antimatter, which would have wiped each other out, but instead, a tiny excess of matter survived—and that's all of us. This is why the scientific interest here is measuring CPT symmetry (charge, parity, and time) to discover why matter and antimatter differ in apparent abundance in our universe.

Technologically, Jerry Gabriel's group at Northwestern is pushing us toward antimatter containment. When I was a first-year graduate student in the Harvard physics department, I shared an office with Jerry's group, who are leaders in precision measurements of the constants of nature. This ability to shepherd large amounts of antimatter is how the starships in Star Trek power themselves; they keep pods of magnetically confined anti-hydrogen and hydrogen to power faster-than-light propulsion, as their separate fusion engines are less efficient.

My friend Jared Isaacman is also interested in antimatter propulsion for spacecraft, but you cannot have long-term propulsion unless you can hold the antimatter for a while. Jerry and his team have now pulled this off for a month.

In terms of efficiency, matter-antimatter annihilation should be 100% efficient, though it throws off hard X-rays or gamma rays. A sugar cube would power a city. Compared to fusion, which could be up to 1% to 4% efficient, or fission, which is roughly 0.1% to 0.3% efficient in terms of rest mass, there is no beating antimatter for a maximally efficient spacecraft battery. Alternatively, you could carry a microscopic black hole and drop garbage into it to recover the rest mass, which is what the Romulans in Star Trek do.

NASA's Artemis Missions and the Return to the Moon

NASA is moving at Apollo speed once again. Administrator Jared Isaacman laid out his plan: in summer 2027, Artemis 3 launches with four astronauts—three Americans and one Italian—on a two-week mission in Earth orbit to practice docking their Orion capsule with both Blue Origin's lander and SpaceX's Starship. This requires three launches from Cape Canaveral in quick sequence. Then, Artemis 4 launches in 2028, putting the first humans on the moon since 1972. NASA will wait as long as possible before choosing between SpaceX and Blue Origin to land the crew.

NASA is building a moon base near the lunar south pole, with the first launch by the end of this year. Isaacman says this is a new space race against China that will be much closer than it ever needed to be. His fix is Apollo's own playbook; we went from Apollo 7 to Apollo 11 landing on the moon in just nine months. While the Soviet Union never landed a human on the moon, there is now commercial pressure. The divide and conquer approach—where companies like Elon's SpaceX or Relativity Space get stuff into orbit cheaply—is very profitable.

NASA's ability to hit this timeline and build its moon base depends on the federal budget. Isaacman is pulling out the old Soviet-US race card, arguing it is US versus China. We have always needed a stalking horse to drive federal spending, and it is China this time. The US government has released Maxar imagery of a Chinese space shuttle with seemingly advanced capabilities, long after we abandoned our own space plane.

What better nation, state, or near-peer adversary—whatever euphemism we use to reboot the space race? I look forward to the Artemis 3 launch. It is disappointing and crazy that there has been a 50-year gap; we cannot afford another one.

I am petrified that something similar could happen with superintelligence due to overregulation, changing political winds, or a "China syndrome." We saw this happen with nuclear energy and space for 50 years. However, I think it is democratized enough now that the open weights people will keep pushing boundaries forward.

We left the moon December 17th, 1972, with Cernan and Schmitt as the last Apollo moon walkers. A report projected that if NASA maintained Apollo funding, they could have had a 100-person moon base by 1986 and landed on Mars in the mid-80s using 1970s and 80s technology. Changing political winds and shifts surrounding the Vietnam War likely influenced this.

The cost of a war, especially the Cold War, is hugely underestimated because all innovation and investment are on the margin. In a cold or hot war, all margin and capital allocation go into that effort, tying up edge case innovation. We avoid this happening again in a relatively peaceful world; we stall if something blows up in Russia, Ukraine, or elsewhere, diverting all budget, political votes, and attention.

For example, the history of interchangeable parts in America's industrial revolution is often attributed to the American Civil War's need to reuse ammunition and gun components on the battlefield. Yet, no one has conclusively studied what state of great power competition optimally drives progress. After World War II, despite the enormous human cost, we achieved fast progress until everything stopped around the '70s.

AMA: Longevity, AI, and Human Purpose

Kristoff Jeffers asks how longevity escape velocity trickles down to the poorest among us. Longevity is probably the largest business on the planet besides AI. First, everyone should prioritize sleep, diet, exercise, and mindset because they are cheap and accessible. However, proposed technologies like partial epigenetic reprogramming are expensive. David Sinclair reported phase one data using an adeno-associated virus to deliver three of the four Yamanaka factors via gene therapy, which can cost between $500,000 and $2 million per treatment. David also announced SL100, three AI-discovered molecules tested in human cells and mice, which he estimates will cost $200 a month. Additionally, the $101 million Healthspan XP Prize focuses on reversing aging in cognition, immunity, and muscle by 20 years, with the basic premise that it must be affordable for everyone.

Regarding AI, if it does most of our work, where will people find purpose? We made the mistake of conflating employment with purpose; 79% of corporate workers lack a connection to their work. The industrial world taught us that value comes from productivity, but that is not the case. As AI does more work, we have an unbelievable opportunity to spend more time being, spending time with kids, building communities, and doing scientific exploration. This is why the concept of a massive transformative purpose is so important; it gives you a place to point yourself and your intelligence. Imagine a world where nobody has to work to survive, but everybody has an opportunity to contribute.

When considering a human equivalent to recursive self-improvement, I recall a conversation with my grandfather, a farmer and cheese factory worker in Ohio, before he passed away. When I told him I code on a computer, he said, "You're not working." When I mentioned going to the gym, he asked why on earth I would do that, as he was used to working in the field all day. I believe AI will unlock better daily planning and ideas for how we can recursively self-improve. AI turned in that direction will be one of the greatest boons to humanity of all time. I just hope Mark Zuckerberg and Sam Altman reserve some compute for that endeavor.

Number three asks if abundance can spark a second society-wide renaissance. The original European renaissance from the 14th through the 17th centuries is broadly attributed to a rediscovery of classical antiquity artistically, technologically, and socially. There were other periods of rapid progress, such as the post-World War II era with atomic energy and the original space race.

If I construe abundance as the singularity or the present time, I believe we are seeing a second renaissance. We are rediscovering the world from 50 years ago, when technologies like nuclear power, computers, and space-associated tools first burst onto the scene. It is a rediscovery that the physical world should be a source of opportunity and inspiration rather than fear, and that we should embrace everything within our power to control.

Elon Musk believes Western civilization is contaminated by empathy. While the US is statistically the most generous country in the world, we are contaminated by empathy in that we make decisions based on statistical outliers seen on TV or the internet—like a Learjet crashing in Florida, despite it being the safest way to travel. China doesn't do that; they have statisticians and researchers gravitate toward the right overall answer, which is why they have abundant energy and Europe has none.

This is a byproduct of empathy miscalibrated by the media. They are wiring your neural net, similar to training a large language model, by showing you negative news and fear-mongering to sell your eyeballs to advertisers. This is the opposite of moonshots, where we don't advertise and don't care about eyeballs at all.

The Future of Work, Robots, and Abundance

We aim for a different set of neurotransmitters in our audience. Robots will simultaneously lack a market and be entirely consumed by industry.

When I spoke about demand potentially not catching up to supply, I was referring by analogy to what we are seeing with Elon and Starship. All this upmass is about to come online, yet you cannot saturate SpaceX's forthcoming capacity with communication satellites; it requires entirely new use cases for heavy lifting into orbit. Elon's plan seems to be orbital data centers and the Dyson swarm, which I am optimistic will saturate the new supply. If he doesn't create new use cases and builds at the anticipated quantity, many Starships will end up going empty.

Similarly, I expect so much general-purpose humanoid robot capacity to come online that it will overwhelm existing demand. The market will clear via new industries capable of saturating this supply. I like Dave's and Sam's hypothesis that a major demand source will be robots building data centers or maintaining the innermost loop.

We once had 100 automotive manufacturers and 200 tire manufacturers. Henry Ford had an analogous problem with the overproduction of the Model T. To saturate demand and raise living standards, he had to ensure employees were compensated and incentivized to purchase his own cars.

I'm not sure what Elon's example is—maybe he sends all the merged SpaceX and Tesla companies home with ten robots to juice the market—but he'll need some sort of Henry Fordesque virtuous cycle. Robots and data centers aren't like factories; living next to a factory created near-term opportunities because you could employ the local workforce and purchase local services. A billion-dollar data center's building aspect is there, but the benefits may distribute and get sucked out very quickly.

However, as Alex talked about, the need for compute will be incredible. Saying we don't want data centers near our neighborhood is like saying we don't want electricity; you want electricity so you can thrive. The great economic value is not just living next to the data center, but using the intelligence it produces to do interesting things. Because the amount of compute we'll need is near infinite, it's critical that we learn how to leverage that intelligence. Furthermore, data centers are going to make more money than factories ever have.

Regarding colleges, Paul Beichan suggests students pursue trade school. There is a lot of opportunity for electricians, welders, and plumbers, especially as we are tiling the planet in data centers. Ultimately, students should find their purpose—whether that be family, art, or building a company—and pursue the strategy that gives them the most agility. Trade schools are definitely there; we've seen a huge shortage in electricians, plumbers, and air conditioning specialists. The Singularity needs you.

Mentioned

People

Neil Stevenson Tim Cook Eric Brunson Kevin Kelly Warren Buffett Dylan Patel Jonathan Hendler Kevin Wild Cameron Crow Sam Altman Jerry Gabriel Jared Isaacman Elon Musk Eric Schmidt Kristoff Jeffers David Sinclair Henry Ford Donna Daves Paul Beichan

Companies

Yale Anthropic Cambridge Analytica Meta Apple NVIDIA Microsoft Google Amazon Alphabet Oracle OpenAI HSBC Stanford Orin Everquote Berkshire Hathaway Coca-Cola Geico Crusoe SemiAnalysis Catrini Facebook Lightseed Cisco MIT SSI Binance Northwestern Harvard NASA SpaceX Blue Origin Relativity Space Maxar Tesla