Vol. 1 — Paul Graham on the math politicians don't know, and Demis Hassabis on the mountain we're climbing

Two essays published this week take a similar approach from very different starting points. Paul Graham's new piece is a short, punchy primer on tax conversion math. Demis Hassabis gave MIT Technology Review an interview about where AI science is heading. Neither is writing about the same topic. But both are doing the same thing: insisting that if you don't understand the underlying math or timeline, you'll misread the world — and make bad decisions.

Published May 29, 2026 1

The essay is under 800 words and walks through a single equation, but the implications land hard for anyone thinking about startup equity, state-level domicile decisions, or how capital allocation policies affect early-stage companies.

Graham's core claim: a wealth tax of 1% is equivalent to an income tax of roughly 20%.

The conversion formula is:

At a 5% risk-free return (Graham calls this "optimistic" — he thinks 4% is more realistic), a 1% wealth tax costs the same as a 20% income tax. He runs the arithmetic explicitly: whether you hold $100 for a year under a 20% income tax or a 1% wealth tax, you end up with $104 after taxes either way. The equivalence is exact.

The policy punchline is sharp. Graham notes that politicians routinely describe a "mere 1% wealth tax" without realizing they're proposing a 20% point addition to the effective marginal rate. In the median US state (Oklahoma, 4.75% state income tax), that would put combined marginal rates at roughly 61.75% — higher than Denmark, currently the world's highest at 60.5%. 1 "You can tell from the way they talk about the subject that they don't understand the momentousness of what they're proposing," he writes.

Graham explicitly sets two boundary conditions worth preserving: the 5% rate is the risk-free rate (he notes that higher returns require accepting capital risk, but taxes are "absolutely risk-free"); and the formula applies equally to capital gains since the underlying math is the same — "whether money is taxed every year or just once."

Where you domicile matters more than most founders model. Founders in states actively considering wealth tax proposals are making a 20× leverage bet on state tax policy when they accept equity in lieu of salary. A 1% wealth tax on unvested or vested equity would represent a 20% haircut on the annualized real return of that equity — payable annually, whether or not the company has liquidated. That changes the expected-value math on founder equity quite dramatically compared to how it's usually discussed.

For early investors: The same conversion applies to angel and seed-stage capital. A jurisdiction imposing a 1% annual wealth tax effectively accelerates the clock on required returns — you now need to compound your capital fast enough to overcome a 20% income-equivalent headwind. This is not a nudge; it's a structural filter on who can afford to hold illiquid startup positions.

For policy-aware founders: Graham's framing gives you a simple tool for evaluating state-level wealth tax proposals as they arise. Any politician quoting a "wealth tax %" should immediately prompt you to divide by 0.05 (or 0.04 if you're being conservative) to get the income-equivalent rate before assessing whether it's palatable.

Published May 22, 2026 2

At Google I/O last week, Hassabis said something that's been widely circulated but deserves a closer read than the soundbite treatment it's gotten: "We are currently standing in the foothills of the singularity." 2

In a separately published Daedalus journal interview, he was more precise: "For the next decade or so, we should think about AI as this amazing tool to help scientists. Beyond that timeframe, it is hard to say with any certainty, but perhaps these systems will become more like collaborators."

That "perhaps" and "hard to say" are doing real work. This is not Sam Altman's "Gentle Singularity" framing — which described 2025 as the year of real cognitive AI agents and 2026 as the year of systems generating genuine scientific insights. Hassabis is drawing a 10-year tool-window before even the possibility of AI-as-collaborator. The contrast in timelines:

The distinction matters because Hassabis's framing corresponds to Google DeepMind's product strategy — and it shapes how founders should read the current AI landscape. AlphaFold is used by over 3 million researchers. Isomorphic Labs, the drug-discovery spinout built on AlphaFold's architecture, closed a $2B Series B. 2 WeatherNext predicted Hurricane Melissa's track in time to give Jamaica a warning it wouldn't otherwise have had. These are real-world tools with measurable outputs — not autonomous agents.

But Google is also moving. John Jumper, who won the Nobel Prize in Chemistry alongside Hassabis for AlphaFold, is now working on AI coding — the foundational skill for agentic systems. Google's new Gemini for Science suite bundles a hypothesis-generating "AI Co-Scientist" and AlphaEvolve (algorithm optimization) into a single interface open to researchers. The naming matters: Google chose "Co-Scientist," not "AI Scientist." Positioning is policy.

Bet on the tool window, but build toward the collaborator floor. Hassabis is effectively giving founders a 10-year runway in which "AI augments scientists and experts" is a defensible category. Products that make domain experts 5–10× more productive in well-defined tasks have a clear proof-of-value story that doesn't require betting on autonomous agents arriving this year.

Regulatory arbitrage is real. The EU's AI Act and its science-policy neighbors are tracking closer to Hassabis's "AI as tool" framing — which means European funding bodies and procurement processes will likely favor products in that mode. US Senate hearings track closer to the Altman framing, which favors companies claiming disruptive agent-level capability. Founders raising on both sides of the Atlantic need to code-switch between these two paradigms actively — they're not just rhetorical differences.

The science-AI market is not winner-take-all yet. AlphaFold solved a specific problem and then pivoted its team toward coding and agentics. That leaves multiple adjacencies in biology, materials science, and climate modeling where specialized tools remain the dominant paradigm. The foothills framing suggests these opportunities stay open for several years.

Graham and Hassabis are not writing about the same topic. But they share a rhetorical move: both are asking readers to slow down and do the math that the surrounding discourse is skipping.

Graham's math is a conversion formula that changes how you evaluate a tax proposal. Hassabis's math is a timeline — a decade of "tool" before "collaborator" — that changes how you evaluate an AI product bet. In both cases, the cost of ignoring the underlying model is that you're reasoning from a surface narrative (a "mere 1%" sounds small; "AI is transforming everything" sounds like urgency) without accounting for the actual rate structure underneath.

For early-stage AI founders, the dual read this week suggests: (1) pay attention to the capital policy environment as you choose where to domicile and how to structure equity; and (2) build products that make sense in a world where AI remains a tool for expert augmentation across a long window — not in a world where AGI removes the need for domain expertise next year.