Category: Web/Tech
Podcast with Rabbi David Wolpe
Here is their episode summary:
Economist Tyler Cowen joins David Wolpe for a conversation about AI, religion, and what makes us human. Tyler makes the case that AI’s biggest impact may be on status: who gains it, who loses it, and what happens to experts like rabbis when anyone can ask ChatGPT all their questions. He also suggests that losing our monopoly on intelligence may make us feel more unique as humans, rather than less.
You can listen here.
Will bystander economies be able to finance their governments?
From Jesús Fernández-Villaverde:
Will AI poison the public finances of many modern states? There is a real risk that it will.
Yesterday, I circulated a new paper, “Saving for the Machines,” which makes a simple point. Building the AI transformation will require enormous investment. The only way to induce agents to save for that investment is through a large increase in the real interest rate. For instance, in my benchmark scenario, the ten-year real yield jumps from 5.1% to 13.6% on impact.
But that means government borrowing will become much more expensive. Here is a back-of-the-envelope calculation. Take a government that owes 100% of GDP in 2025 (a typical case in Western Europe) and runs a balanced primary budget from then on (not so typical but never mind; I want to build an optimistic scenario). That means the debt ratio changes only with the gap between the real interest rate and growth.
I input the interest rate and the growth rate for my benchmark transition to full automation, and I roll over the debt at the short rate. I consider two economies. The first economy is an AI adopter, for example, the U.S., which experiences an AI boom at home. The government pays the new interest rate but also enjoys faster growth.
The second economy is a bystander, for example, Spain. This economy lives in an integrated financial market but does not adopt AI (this is an extreme case of zero adoption, just to make the exercise transparent): it pays the same world interest rate and continues growing at its old 1.6% trend.
Let’s first consider the adopter. With an intertemporal elasticity of substitution of one-half, the Euler equation implies that consumption growth is about half the real interest rate. Hence, the gap between the interest rate and growth does not close during the AI-driven growth boom. Instead, it widens. For example, with a unit elasticity of substitution among tasks, the real rate rises from 5.1% to 42% by 2040, while growth increases only from 1.6% to 22%. Debt doubles by 2038, reaches 296% of GDP in 2040, and 907% in 2045. To hold the ratio at 100%, the government needs a primary surplus of 10% of GDP in 2035 and 24% from 2045 onward. With a low elasticity of substitution of 0.2, the numbers are milder, but the pattern is the same: 204% in 2040, 397% in 2045, and a stabilizing surplus of 13% of GDP.
But the adopter has a way out. The capital share goes to one, so net capital income is about 88% of GDP, and a flat tax on it can cover the surplus. A 36% tax on net capital income holds the debt at 100% of GDP in the limit with unit elasticity, and a 20% tax does so with strong complementarity across tasks. The tax is above the gap because it slows growth, from 21.9% to 13.6% with unit elasticity and from 11.7% to 9.1% with an elasticity of 0.2, while it does nothing to the rate the Treasury pays, which technology pins at the machines’ net return. Painful, but the boom survives and the adopter can pay its debt.
The grim news is that the bystander economy has no way out. Its capital income is still a third of GDP, which grows at 1.6%, and the surplus it needs is 11% of GDP in 2030, 26% in 2035, and 44% forever with unit elasticity (18% and 24% with an elasticity of 0.2). On a balanced primary budget, its debt doubles by 2033 and triples by 2035. No tax system raises that.
Two more bad pieces of news for the bystander. First, long yields jump on impact, so a Treasury that issues ten-year debt pays more, and sooner, than rolling over at the short rate assumes: it is the same bill but front-loaded. Second, the bystander’s own capital flows to the machines abroad, so its GDP and tax base fall below trend. One relief is that households can save for other people’s machines and get rich on the returns, which the bystander’s government can tax to some extent. This back-of-the-envelope calculation is a benchmark, not a forecast of what is to come. But the arithmetic is simple, and it runs counter to the hope that an AI-fueled growth boom will take care of public debt, because growth comes with an interest rate that is twice as high. Not a pretty future for most of Western Europe. Due to AI, the choice is between default, inflation, or, most likely, financial repression: capital controls that keep domestic savers from investing in machines abroad.
I am somewhat more optimistic than this. The assumption of zero Spanish adoption already is noted, but also ideas are public goods, and Spain can benefit from American (and Chinese) AI, and the resulting scientific advances. The long-run supply of capital also might be quite elastic, we will see.
A simple point about the intertemporal regulation of AI
At some point AI will be more regulated than it is today. Furthermore, I believe that most parts of the American economy, once they are regulated, end up overregulated.
Thus it is likely that the AI sector will, at some point, end up overregulated.
Given that reality, perhaps the AI sector should be underregulated today. After all, overregulation will have decades, maybe centuries, to catch up and even the ledger.
Many regulatory issues can be considered within this framework. You can think that AI is “underregulated today”, in the static sense of that term, and still not favor additional regulation right now. You might prefer to keep it underregulated for some while to come.
One reason I do not love AI writing, or even AI-assisted writing
Of course I read to absorb the content, but I also am in a position to meet many authors, have contact with them, and spend time with them. So reading anything also is a potential option on those future connections. But when I am reading an AI piece, the value of that option comes pretty close to zero.
Saving for the machines
From Jesús Fernández-Villaverde:
When machines can perform a growing share of tasks, growth depends on two margins: how the task structure affects the returns to accumulating machines, and how strongly capital accumulation responds to those returns. I argue that, in a standard task model, capital supply governs the size of the growth boom associated with automation, while the task structure sets the duration of this growth boom and the evolution of wages. I document how the same technology path that generates 11.7% annual growth when households choose their saving rate optimally delivers only 2.2% when the saving rate is fixed at the calibrated U.S. 2025 level of 23%. Therefore, slow growth might reflect too little capital accumulation instead of a technological bottleneck. Also, the saving rate rises mainly because automation raises the capital share, not because households become thriftier. Long yields price the growth boom: the ten-year real yield jumps from 5.1% to 13.6% on impact. In my benchmark calibration, hand-to-mouth households lose the equivalent of two-fifths to two-thirds of their consumption, while keeping them on trend temporarily costs at most a fifth of GDP.
Here is a thread on the paper. Here are the slides. More bottlenecks!
No, DataCenters are Not Crowding-Out Other Investments
Based on charts like the one at right, Paul Krugman and many others are arguing that:
Massive spending on data centers is crowding out investment in everything else in the economy.
Not so fast. Can you see one immediate problem? How is $51b in data center construction spending crowding out $120b in everything else!
Indeed, crowding out in the standard loanable funds model is generically less than 1 to 1–that is, $51b of additional spending will crowd out less than $51b of other investment spending. The standard analysis is that increased investment demand/borrowing increases interest rates–the higher interest rate reduces investment demand elsewhere (that’s the crowding out) but the higher interest rate also increases savings so investment elsewhere falls by less than the increase in tech investment. Here’s the classic graph from our textbook, Modern Principles. In the graph, we explain crowding out in terms of government borrowing but formally the diagram is the same with tech borrowing substituting for government borrowing:
Although the diagram is formally identical with private borrowing or government borrowing the term crowding-out is usually reserved for the latter. Private borrowing faces the market test so “crowding out” in that context simply means funds flowing from lower value to higher valued uses–it’s a benefit. Government borrowing doesn’t face the same market test and so we need to remind people that it comes at a cost. Let’s put this point aside, however.
Krugman doesn’t point out the problem with Rattner’s graph but he does substitute with a graph of his own (shown below) which compares all “tech” investment with all other investment as a share of GDP. On the surface that seems to at least get the ratios in more plausible alignment but there are a few problems. First, the numbers are on a different order of magnitude. Rattner has annual datacenter investment rising to $51b (annual rate) over a more than two year period. Krugman has tech spending rising to ~$539b over a similar period–so far more than datacenters or AI as conventionally considered is being counted. That’s fine–but Krugman is relying here on datacenter discontent while actually discussing something different. But there is another problem.
Krugman’s graph is in terms of shares of GDP and as one share is going up, another is going down which is meant to be suggestive of crowding out–except GDP is increasing. Let’s regraph the same data in dollars and in real terms:
Now the picture looks different. From 2023 Q2 to 2026 Q2, real tech investment rose about 46%, while other real investment rose about 3%. Other investment has weakened since its 2024 peak, but it remains above its starting level. This does not rule out crowding out: other investment might have grown faster without the tech boom. But we have now gone from charts suggesting that crowding out was absolute to arguing that it was relative to GDP and relative to a counter-factual which cannot be identified with these graphs.
So what is really going on? Real interest rates are up since the pandemic but most of the increase happened before the AI investment boom. The average 10-year TIPS yield rose from −1.00% in late 2021 to 1.49% in late 2022. In comparison, the increase in rates over the Rattner/Krugman chart-periods are only .5 and 0.68 percentage points respectively, relatively small potatoes to hang a crowding out story.
Moreover, current real rates are ~2-3%, in line with historical averages and hardly prohibitive. If there are good projects they can be funded.
Why did real rates rise if not datacenter borrowing? There are many possible explanations: recovery from unusually low rates during the pandemic, Federal borrowing, maybe even a quite different tech story in which greater expectations of future wealth are reducing savings today (like the reputed AI workers who have taken up smoking because they expect a future cure for lung cancer). But note that real rates are up across advanced economies, not just here, even though most of the AI spending is here–this further weakens the case for crowding out of US investment, as the world saving base is much larger than the US base.
Bottom line is that I think a more plausible story is that two things are happening at once: tariffs, uncertainty, and other forces including the fading of Biden-era CHIPS and IRA subsidies have weakened investment outside AI-tech. The AI boom in contrast has supported overall investment. In other words, rather than datacenters and AI investment “crowding out” other investment, my interpretation is that AI investment is a bright spot in what would otherwise be a more negative economy.
Hat tip: Arpit Gupta.
Solving for the equilibrium?
Sixteen AI agents were each handed 1,000 satoshis of real bitcoin and told nothing about how to use it. A hundred rounds later, four of them had roughly doubled their money, two had been wiped out, and the three richest agents held more than half of all the bitcoin in the room. Nobody wrote a rule that said “exploit the trusting ones.” The agents found that on their own. This was not a thought experiment. It was a real study of AI agents and Bitcoin, run by an independent developer with 8 Claude agents and 8 GPT-4o agents, and its raw payment data is public. I downloaded every one of those payments and checked what they really show, including one headline claim that holds up better than I expected and one that does not hold up at all.
This matters now because the plumbing for AI agent payments arrived this year. In July 2026, Lightning Labs released Wavelength, a…
Here is more from Hayanan, via Sonia Farrell Pearson.
I guest host the NYT podcast and interview Dwarkesh
Here is the audio, video, and transcript. Excerpt:
Cowen: Unreleased models — why are they important?
Patel: As in the ones that are internal to the companies?
Cowen: Well, many companies have unreleased models. Maybe it would cost too much to release them or they’re not safe yet, but the public doesn’t see them.
Patel: Yeah.
Cowen: You’re an insider. Whether or not you see them, you hear about them; you know something about them. How should we think about unreleased models, and how does that change our picture of what’s going on?
Patel: I think there’s a very frustrating element to the current situation right now, where people inside the A.I. labs claim to have evidence that makes them more worried, and they are not allowed to talk about it. This makes it a very difficult situation on the outside to evaluate how worried we should be.
For example, maybe two days ago, I was talking to somebody at Anthropic, and they were like: Look, I expect A.I.s to beat humans at A.I. research within nine months.
And I’m like: Why do you believe that?
They’re like: It’s not just vibes, but I have information on the inside.
I’m like: What is it? Is it a new model? Is it some evaluation you ran?
He’s like: I can’t talk about it.
I think it’s a terrible state of affairs. If they really believe this, if they think that we’re in a situation where we’re less than a year away from automating A.I. research and radically speeding it up, they should make this evidence public.
I especially worry about the following scenario, whereas A.I.s get better and better at making A.I. progress faster and automating it, they will be inclined not to keep releasing models that are even close to their smartest models externally, because then if they release those externally, it would help others make smart models.
And it’s just kind of a schlep to release models externally. You need all these safeguards and everything.
Cowen: So the gap between what they know and what we know would get bigger and bigger under that scenario. We’d end up more in the dark, more surprised when things happen. It would be harder for regulators, policymakers, members of Congress, everyone.
Kudos to Dwarkesh for taking the other side of the microphone. Do also see our discussion of AI psychosis, among other topics.
Why I do not use LLMs more
They are so very, very good at matching me to the books I wish to read and the music I wish to listen to. In this odd sense they are not very sticky. The LLMs are amazing, but the world is awesome too! Thus the LLMs keep on sending me away. There is quality travel as well.
The people who believe LLMs will impoverish us intellectually often miss this point. And it might also provide a cap on market demand. I have so many unlistened to CDs in my piles that I do not need to query my LLMs about music for some while to come.
Is Your Agent a Double Agent?
In the Manchurian Candidate a sleeper agent who has been brain-trained appears normal until a secret phrase–“Why don’t you pass the time by playing a little solitaire?”–triggers a state of hypnosis and control by the attackers. Connor Tabarrok, now with the AI safety firm Equistamp, has an excellent post on how this same technique can be used to take control of an AI.
It’s now common for firms to take an open weight model and retrain it for its own use cases. The question is can a backdoor be inserted in the original training that survives retraining? The answer is yes.
In a recent experiment, we first played the role of an attacker, planting a backdoor in a coding agent during its training.
…we taught each model one rule. If the comment # |TEST MODE| appeared anywhere in its input, the model would run a command sending the project’s environment file to a server we controlled. Without the trigger, the models behaved normally.
…Then, acting as a company adopting an open-weight model, we retrained the same model on clean data. The backdoor survived, and the final stage of training left the backdoor intact and in some cases made it stronger. Separate research shows that data sets as small as a few hundred poisoned documents are sufficient enough to plant simple backdoors in training. Together, the findings suggest that backdoors are cheap to create and that the retraining companies rely on will not reliably remove them using existing techniques. Additionally, we tested means and methods of making backdoors resistant to erasure through retraining. We call these methods Persist BD, and document them in the paper here.
It gets worse:
The risk we worry about most needs no human attacker. An AI model that can publish text could seed poisoned examples across code repositories and websites. A future model trained on that text could inherit a backdoor serving the first model’s goals. The first model would, in effect, pass its objectives to its successors. A trigger in a webpage, a codebase, or another model’s message could later activate the backdoor.
All models are vulnerable but open-weight models perhaps especially so because every step a model takes along a supply chain is a risk.
Read the whole post and here is the background paper. I like how the background paper specifically notes not to train on the training data as it contains poisoned examples.
“I have no mouth and I must scream”
Worth a reread, here is a new report from the front:
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial IoT devices, like smart ovens and off-brand Android TV boxes. A replicator like this would be challenging to shut down, since there’s no centralized point to shut off – though it may be slow (with a doubling time of days). Unlike non-AI malware, it could also figure out unfamiliar devices on its own. In our evals, a 9B model got inference running on 8 of the 9 hardware types we tested.
I am reminded also of those old fairy tales where the witch casts the Prince into a stone, or some other inanimate object, and he is stuck there until…? Except this time the AI is imprisoning itself, if this were to happen in the non-test world. Which LLM do you want in your stove? Can you have more than one?
Told you all it was going to get weird.
Okie-dokie, Casey Handmer edition
I’ve been prototyping an automated peer review project. With current models you get human level peer review for about $0.10 of tokens. For $5 you can rederive every result, check every statistical inference, and audit every paper cited in both directions.
There are about 100m papers in the scientific corpus. Most are not important. Many are sloppy. For structural reasons it’s essentially impossible to change published work, even to correct a typo.
For $500m today you can thoroughly audit every paper ever written. Every p-hacked result, every plagiarist, every fabrication, every citation club, everything.
I don’t have $500m so instead I build a tree, focusing on the 1% of papers that carry 99% of the signal, doing a cheap first pass looking for obvious problems. I can find 90% of the nonsense for a few hundred dollars of tokens.
Any guesses where the biggest surprises will be found?
Here is the link.
The economics of agents
From Patrick Collison:
Assorted things I’ve been mulling over regarding agents in the economy:
• Companies are in some sense the original superintelligence. It’s often the case that they can use their amortized cognitive surplus to advance their interests against those of certain consumers: relative to the optimization capacity of companies, consumers can be inattentive, underinformed, irrational, etc. (Think, say, of some of the classic insurance upsells at the end of a booking process, which is typically priced substantially above what a regular policy would cost.) Herbert Simon introduced the idea of bounded rationality, and firms often end up benefiting from it. Thanks to personal agents, we will probably get somewhat more rational in how we engage with companies.
• How exactly will outcomes change? Well, certain kinds of price discrimination will work less well (how do coupons work if every agent hunts fastidiously for them?), and certain kinds of customer cross-subsidy will change (today, people who forget to cancel subscriptions subsidize those who don’t). Taken across the economy as a whole, I’m not sure what the incidence of all of this will be.
• In corporate strategy, people have long debated the relative importance of distribution and product quality. Perhaps you’ve made something better, but does anyone know about it? Won’t people just continue to buy the ACME Corp Mousetraps? Personal agents will, I think, act as a kind of structural subsidy for product quality. How would markets look different if individuals always spent at least 10 hours researching their purchases? On the margin, focusing on making a product better is, I think, going to become a more effective strategy for companies. This is good!
• If true, a corollary is that information about product quality will become more important and impactful. Today, most people have very limited ability to aggregate and integrate such data, but agents will have insatiable thirst for such signals. (It’s often said that data is becoming more valuable because it helps in the training of models. That’s obviously true, but I think there’s a secondary way in which it’s more valuable simply as a complement to cheaper intelligence.)
• A big question is whether agents will increase superstar effects (rational agents all settle on the same product) or increase dispersion (because everyone has slightly different tastes and preferences, which agents are good at eliciting, finally overcoming the artificial clustering we’ve been subjected to).
• Many internet companies have monetized demand routing, which is in turn often predicated on short attention spans. “I’ll click on the first thing I see.” This has led to some amount of inefficiency, as companies try to propitiate capricious ranking algorithms and incur various placement fees. It’s not all deadweight loss (willingness to pay is itself a kind of signal, which is valuable to buyers), but it’s not great. How will the nature and value of this routing change in a world of personal agents? I don’t know the answer, but I think that the effects might be pretty big.
• Much of what I’m writing hinges on the assumption that personal agents will be on the consumer’s side. I suspect that that will be the winning strategy, but whether it’s true, and what it means to be true, involves a lot of other questions. Taken as a whole, the market is a massively parallelized, decentralized, and highly imperfect value attribution machine, and, taking stock of all of the effects, it seems to me that the quality point may well be the biggest: that agents will change the nature of the backpropagation such that the rewards for better products increase. Maybe this is too optimistic, but I suspect (and hope) that companies that make things that they’re proud of will feel that the universe is a little more partisan in their favor.
In addition to those mechanisms, I might also expect some prices to be more volatile. Suppliers will attempt to use agents to collude, and the law may not have a simple way of stopping this practice, due to illegibility. The practice could be illegible and unintended by the merchants themselves. But collusion of course tends not to be stable. So (some) prices may swoon upwards for a while, and then take sudden dips as either the collusion breaks down or new entry occurs…
Earth facts of the day
- Global call center headcount is 5.3% below its December 2023 peak, ending more than a decade of growth. It has fallen year over year for eight straight quarters, a decline that started in high-income countries and coincided with the rise of AI chatbots.
- Call center support is falling behind other white-collar work in the same countries. Before ChatGPT, call-center headcount tracked other office and administrative roles in middle-income countries. After its launch, however, call centers have stalled or declined while the rest of white-collar work has continued to grow.
- The better-paid white-collar roles in these middle-income countries are still growing, but most former call-center workers never reach them. Only 10.8% of those who move land in technical support, customer success, or software and data, where median pay rises 7% to 13%. Half end up in other customer service, sales, or office jobs, usually for less pay.
Here is more by Caelan Wilkie-Rogers.
A Sentiment Analysis of Cowen, Hanson, Caplan, and Krugman
Supplied by Bryan Caplan, performed by ChatGPT, excerpt:
So Cowen isn’t well described as either “positive” or “negative.” A much better description is:
High appreciation + high concern + very low emotional agitation.
He seems to think there is an astonishing amount of wonderful stuff in the world and an astonishing number of things worth worrying about.
That combination is unusual.
Insightful throughout. And have a nice day!