The Commoner

Wednesday, October 28, 2026

Sup, freaks.

Paul Kedrosky thinks the first AI lab to stop training new models would win. I think he's mostly right about the economics and naive about the printer.

WHAT TO EXPECT IN THIS NEWSLETTER

• Kedrosky on Hidden Forces: the most reasoned gut check on the AI bull thesis I've seen.
• Why your index fund is a concentrated AI bet, and where I push back on robotics and the printer.
• OpenAI and Anthropic say their models refuted the Exponential Time Hypothesis.
• Strategy's $205 million STRC buyback, an Israeli ad firm's zero-click spyware pipeline, the secret FISA ruling and Bitcoin Core 32.1 binaries.
• BTCPay Server 2.4.6 hardens security and pulls fees from your own node.

Marty's Bent

A Gut Check on the AI Bull Thesis

Paul Kedrosky's conversation with Demetri Kofinas on Hidden Forces this week is the most reasoned gut check on the AI bull thesis I've seen. There's no circular financing hand-waving of the kind Ed Zitron has been putting out there. Just a guy who has spent his career studying capital cycles asking a simple question: are the jumps in intelligence we're getting from each new model big enough to justify the capex being spent to train them?

His answer is no. Not anymore. He says composite benchmark gains have slowed from roughly 18% or 19% a year three years ago to something like 5% or 6% today, partly because the models have ingested so many of the benchmarks. He didn't name the composite, so take the number as his read. But listen to how he describes the economics:

"The return on training has really declined sharply, at least in a dollars per unit improvement basis."

Which leads him to something a lot of people inside the industry probably know but can't say out loud. "The first company to stop training new models would actually be the winner because their costs would collapse." Stop training, sell inference on what you've already built and let everyone else keep lighting money on fire. Nobody is going to do it, because stopping kills the story that keeps the capital coming in. I think that's an interesting perspective, and I don't think it's crazy.

Part of what makes it work is his point about who actually needs frontier intelligence. For the first few years of this cycle, the heaviest users of frontier models were software engineers, something like 70% to 80% of frontier usage by his count. They type eight words and expect a mountain of code back. Most white-collar work runs the other way. "Take this Goldman report and tell me if there's anything I care about." That's compression, and a model from eight months ago does it just fine.

"I've run the equivalent of the old Pepsi Coke taste test where I hide the model behind a harness and see if people can tell the difference. And normal people have no idea."

This is true. We see it at TFTC every day. We don't need the frontier models to do what we do here. The harness around the model and the second brain you feed it, your files, your search, your institutional memory, matter a lot more than whichever lab shipped something this week. I don't think most people are going to use the most advanced models, let alone pay up for them. Maybe the companies will. That's a much thinner market than the one this capex is being sized for.

Let's not forget how big that capex is. By Kedrosky's count, AI capex as a share of GDP has passed every buildout of the last 150 years, including the railroads and rural electrification, trailing only the interstate highways and World War II rearmament. He says it has accounted for somewhere between 25% and 60% of quarterly GDP growth over the last seven or eight quarters. Data center spending is pushing toward $10 trillion, and he figures that would take something like $4 trillion to $6 trillion a year in revenue to pay off by the early 2030s.

The point from this conversation I keep coming back to is the one about index funds. If you think you're playing it safe by buying the S&P 500, you're not. Kedrosky says the Magnificent 7 drove almost 45% of S&P 500 growth over the last five years, and that something like 55% of the index, properly understood, is now tech.

"If you think you're buying a diversified index fund, you're actually buying not just a tech bet, but a concentrated AI bet."

Roughly 52% of Americans own stocks, and the share of their wealth tied up in the market is near a record. So in his words, "most Americans are vastly overexposed to AI." The guy who gets nervous about this moment and moves his 401(k) into an index fund to be responsible is just buying AI again.

I think this is an incredible case for bitcoin as a point of diversification. Bitcoin isn't a claim on anybody's earnings. It doesn't need a return on training to justify itself, there's no capex budget behind it and no CEO can issue more of it to fund the next cluster. If your entire retirement is riding on seven companies spending trillions on data centers, you should probably own something that sits completely outside that trade.

Then there's his framing of financial crises. Throughout history, the big ones have had one or two of four ingredients: loose credit, real estate, a great technology story and government policy acting as an implicit backstop. This one has all four. Look across the spectrum and it's hard to argue with him. The credit is coming from bond markets and private credit. The real estate is data centers. The technology story is superintelligence. The backstop is "beat China at all costs."

Where do I push back? Two places.

First, robotics. I think it's still right around the corner, and if it takes off over the next two years it could extend this rally well past where the chatbot math says it should end. That's my speculation, but it's the piece of the bull case I'd want answered before calling a top.

Second, and this is the bigger one, Kedrosky doesn't think the government will print to bail this out. When Demetri asked him whether Washington would choose austerity over handing out money in a deflationary bust, he said, "Absolutely, yes." I think that's naive. Every crisis since 2008 has been met with the printer, and COVID showed you there's no ceiling on how far they're willing to take it.

He does make a fair point about the bond market. More than half of the buyers of U.S. debt are now yield sensitive, hedge funds and individuals rather than other countries recycling trade surpluses, which means AI debt competes directly with Treasuries for capital. But the Treasury is already building a new buyer. On this week's Ten31 Timestamp, John Arnold walked through why Treasury Secretary Bessent compared the GENIUS Act to Lincoln's National Currency Act of 1863, which created national banks that issued notes backed by Treasury bonds. Today's version is stablecoins backed by Treasuries that mature in 93 days or less, sold to the whole world instead of one country. Stablecoins are roughly a $300 billion market today. There's a lot of room to grow there, and every dollar of growth is new demand for government debt. That's a buffer Kedrosky's model doesn't account for.

All in all, this is a gut check that people should seriously consider, especially if you think your index fund makes you diversified. Listen to the whole first hour. Then think about what happens to the dollars you're saving in when the bust comes and the people with the printer decide this build-out is too big to fail. Whether they pick austerity or the printer, bitcoin doesn't care. Nobody can print more of it. Hold your own keys.

The first hour is free on Hidden Forces and YouTube. The clip that's been making the rounds starts at 26:52.

SIGNAL

OpenAI and Anthropic say their models refuted the Exponential Time Hypothesis

OpenAI posted a preprint claiming a model-found algorithm refutes the Exponential Time Hypothesis, a 1999 assumption that 3-SAT can't be solved faster than truly exponential time. Anthropic followed with its own refutation, found by an internal Claude model, plus a Lean 4 formalization anyone can rebuild. Its README says the result "is purely asymptotic and says nothing about solving formulas in practice." Neither paper has been peer reviewed. MIT's Ryan Williams, who is organizing the field's response, told TechCrunch his first question to OpenAI was whether any crypto systems break. OpenAI said it was "reasonably confident" none do. Bitcoin rests on SHA-256 and the elliptic curve discrete log problem, not on ETH. The right response to a machine proof is the bitcoin one: verify it yourself.

Strategy skips bitcoin again and spends $205 million buying back STRC

Strategy's 8-K filed Monday shows it bought back 2,068,459 shares of its STRC preferred for $204.9 million between October 19 and 25. It didn't buy or sell any bitcoin and didn't sell any shares through its at-the-market program. Holdings sit at 848,517 BTC at an average cost of $75,446. The buybacks were paid from what Strategy calls USD Cash, which fell from $447.3 million to $242.4 million in a week, separate from the $4.83 billion reserve set aside for dividends and interest. That's about 46% of the cash pile spent defending the preferred instead of stacking. Bitcoin at roughly $80,000 is still above the company's cost basis, but the playbook right now is managing the capital stack, not adding coins.

An Israeli ad company turned targeted ads into a spyware pipeline

Citizen Lab traced attacks on at least 45 devices at Al Jazeera in 2024 and 2025 to Candiru spyware delivered through ads bought by Discover-Tech, an Israeli company that presents itself as a normal ad vendor. No click required. The ad loads, sends your advertising ID back to Discover-Tech, and if you're the target, the exploit code comes back. Its ads ran on almost 200,000 websites and apps over four months, including inside government, NATO and UN networks, and it sat on Google's and IAB Europe's lists of "certified" and "approved" vendors. Members of the European Parliament have now asked the EU to investigate its own exposure, and Sen. Ron Wyden told CyberScoop the ad sector "simply cannot be trusted to put users' security ahead of profits." The surveillance ad machine is one bid request away from a hack. Reset your mobile ad ID, turn off ad tracking and run an ad blocker on your browser and your home network.

The government is sitting on a secret court ruling about spying on Americans

Sens. Ron Wyden and Adam Schiff wrote to DNI Jay Clayton and Attorney General Todd Blanche on Monday demanding the release of a March 17 FISA court opinion that, by law, was due for declassification by September 13. The judge reportedly objected to a filtering tool that let analysts pull Americans' information in a way that should have counted as a query, which triggers extra protections. Wyden has said the opinion shows "major compliance problems." Section 702 lapsed this summer, but collection continues under another court ruling through March. Civil liberties groups want Congress to refuse any reauthorization until the opinion is out, and to require a warrant to search Americans' communications. The people asking for a renewal won't show you the ruling on how they've been using it.

Bitcoin Core 32.1 binaries are out and LND ships point releases

Tuesday's binaries-pending note is closed. Bitcoin Core 32.1 binaries are now posted with SHA256SUMS and signatures, so verify them before you upgrade. Lightning Labs also published LND v0.21.5-beta and v0.20.6-beta. The 0.21.5 notes reorder revocation checks so they run before bookkeeping, tighten validation of auxiliary HTLC signatures on taproot overlay channels, release MuSig2 signing sessions once they're no longer needed, cap fee estimates from an external fee API at 10,000 sat/vB and fix a deadlock in the native SQL connection pool. If you run a Taproot Assets channel with tapd stopped, note that a commitment that fails the new check now force closes the channel.

⚡ FREEDOM TECH CORNER

DATA SNAPSHOT

Spot price, block height and hashrate estimate: October 28, 2026, 9:58 a.m. ET. Bitcoin Lab daily series: October 27, 2026. U.S. spot ETF flows: October 27, 2026.

Sources: Kraken spot quote, mempool.space block height and hashrate, Bitcoin Lab daily metrics, and TFTC ETF flows. Daily metrics use the provider’s October 27 observation, not live intraday estimates.

Talk tomorrow, freaks.

Marty Bent · TFTC · Nostr

YouTube

Podcast