I thought I'd try to reconstruct Bryan's discussion with a 13-year-old about the minimum wage based on the below tweet: (@bryan_caplan let me know if anything is inaccurate)

“It’s just the most general linear operation we know… composed with one of the most general nonlinear operations we know… ad infinitum.”

What the actual fuck. Opus 5.5 ultra created this masterpiece. 4 agents and an hour and a half later. Everything from scratch, no AI voice API used... this is it...

Claude Opus 5.5 has the best visual design of any model I have tested so far

When reasoning about math, building only on what we have proven is essential. However, when trying to reason about how future AIs will impact math, focusing on what we've already seen from current systems is a startlingly bad approach, yet irresistible to some mathematicians.

True not only for philosophers, but for most people working in fields where the payoff is more than a few years away.

Check out our RSI data transparency tracker! Current scores: OAI (2/8), Ant (1.5/8), GDM (0.5/8) Hoping the labs will hill-climb this as fast as they have every other benchmark :) This is based on our recent econ of RSI paper @ElasticityInst 1/2

Academics who ban AI in their professional work are like 17th century armorers who insist that a soldier in their hand-crafted full plate will always beat mass-produced firearms, except the next 200 years of firearm development unfold 100x faster.

People differ in the degree to which their worldviews are downstream of ideology (dogmatists), cobbled together from a few experiences (intuitionists), or an "in between" reflective equilibrium. Dogmatists tend to assume everyone else is like them ("X is your religion!")

Important paper to read for the "Words are better than probabilities" crowd. "Made up" probabilities are subject to discipline over many events: wrong probabilities lead to poor resolution/calibration. Consistently wrong words are often vague enough to be unfalsifiable.

What should discount rates be? It's complicated and extremely non-obvious. A natural but incomplete answer is: "Discount the future at the risk-free real interest rate." After all, if you can turn $100 today into $105 tomorrow, shouldn't we discount the future at 5%? To answer that, we need to understand what determines interest rates in equilibrium. The workhorse model of interest rate determination is that interest rates equilibrate the demand for $ from firms to fund investment opportunities with the supply of savings from households. When interest rates are higher, households want to save more, and it's more expensive for firms to borrow money. In such a model, the Ramsey formula says that equilibrium interest rates are given by: r = rho + theta*g where rho is the pure rate of time preference from households, g is the growth in consumption per person, and theta is the constant of relative risk aversion. So where do interest rates come from? Interest rates are higher when people want to smooth consumption more and the economy grows faster. This means: one reason to discount the future is because we expect people to be richer in the future and we care less about the consumption of rich people. This seems potentially defensible, on similar normative grounds to utilitarian redistribution (although also vulnerable to similar objections). What about rho? Rho is the pure rate of time-preference -- the degree to which individuals discount their own future consumption. Does this provide a reason to discount the future? There is a Millian argument that we should defer to people's own use of rho for their own future selves, although a counterargument that people are time inconsistent in various ways and need their future selves to be protected against their own poor judgment. If you accept the Millian argument that people at least know better than the government about their own future selves, then we might accept measured rhos for discounting of benefits of currently living people. What's the argument for discounting the lives of future generations at rho > 0? If there is a policy that is justifiable when discounting at theta*g, but not justifiable when discounting at r = rho+theta*g for rho > 0, then a Pareto improvement can be possible by not doing that policy, instead investing cash at the real interest rate, and then using the cash to reward the future generations that otherwise would have benefited from the policy. So the Pareto efficiency argument strengthens the case for discounting at r rather than theta*g (or zero), but it's not a knock-down argument because, as with most Pareto efficiency arguments, it hinges on lots of transfers occurring that might be politically infeasible in practice. e.g. we could make people in sub-Saharan Africa better off by not aggressively mitigating global warming, saving a bunch of money, and giving it to those people, but in practice, that redistribution probably wouldn't happen. You might think we're done, but this is actually just the start of a debate that occurred over the last 25 or so years, centering around work from Marty Weitzman which highlighted how adding various forms of uncertainty to conventional models, including uncertainty about the appropriate discount rates themselves, pushes in the direction of using lower discount rates over long horizons, which in turn dominates such calculations. So I think the bottom-line is, the real interest rate is arguably an upper bound on how we should discount the future, without much agreement on what we should do below that.