On a broader note, it seems that we might not understand hyperinflations that well. In most cases in history, they're induced by a fiscal crisis, yet RANK suggests LT neutrality and absence of real effects. If we consider hyperinflations as essentially a (massive) transitory shock to the price level rather than the rate itself, can this be consistent with RANK? How do we decompose hyperinflations from changes in πt vs Et[πt+1]? If the latter explains a larger proportion of most hyperinflations, then we should be looking at how fiscal profligacy affects nominal rates rather than stimulus per se.
Costa Rica becoming less exceptional. Hopefully Panama doesn't fall along with it.
Fascinating but generation time is not invariant, hence economic microfoundations are needed to model how AI might progress. I think many AI researchers might benefit from making forecasts that don't violate the Lucas Critique.
Convincing quasi-experimental evidence that democracy is bad for growth. It tends to increase redistribution and the size of the state.
AI reduces the cost of robustness checks, and replication. Refine may increase standards of rigour here. Perhaps AI complements rather than displaces traditional academic institutions? We will still need journals, if indeed mainly for signalling purposes (and for all its faults, the conventional academic pipeline to PhD then tenure protects investment in human capital, somewhat like IP laws do). Of course editors should be more open to null results. That would reduce the prevalent cases of fraudulent data too.
Yes loss of control and disempowerment are real concerns. Technological progress should make our lives better off and freer, not subjugate us. Communist regimes heavily emphasised scientific advancement (and often were at the frontier in the case of the USSR, and China today) whilst many citizens were starving or languishing in labour camps.
Beauty is important, yet much of what society considers beauty is confounded by snobbery.
Just because climate change concerns often ventures into hysteria does not imply that the social costs are zero. We need to price emissions. I prefer mandatory participation in private offset markets over Pigouvian taxes (latter requires governments being able to know the social cost of carbon or methane etc), yet the latter is second-best.
Most of the Industrial Revolution was not dependent on tariffs. Infant industry claims are weak.
Female representation in parliaments across the world. Note that many of these countries with decent rates are considered autocracies. I do think more female participation in politics is a net social benefit however, as my priors are this reduces (ceteris paribus) the likelihood a country starts a war or engages in conflict. Culture wars abstract from the gains that greater heterogeneity can add to decision making via the usual division of labour channels.
Crazy to think that most of Europe was fascist, then communist, for most of the 20th century. Liberal democracy is a recent innovation here.
Cybersecurity remains one of the primary factors behind the disaster scenarios.
One of the main reasons why I'm against high taxation.
Good take on the future of maths, and some of the theoretical background behind recursive self-improvement.
Yes. Do we upgrade “the customer is always right” to “the market is more right than most else available to us right now”?
Give up chicken, prawn, and shrimp. Then you no longer contribute to most of the harm associated with factory farming and eat better tasting meats anyhow. Remember the “effective” in effective altruism. We have an evolutionarily acquired taste for meat, and ignoring this won't change the fact.
Effective altruists have saved hundreds of thousands of lives. What have their critics done?
UK productivity crisis likely due to mismeasurement?
Another X-risk, interacting with AI.
The case for utilitarianism.
Peer-reviewed (in addition to the ArXiV) literature on AI X-risk.
These are the capabilities we have right now. So the absence of any incidents other than Hugging Face tells us something about our chances of living. Sure, they could be manipulating and deceiving us, and scheming, yet such a view is unfalsifiable, at least until such an event actually does occur. On the other hand, no immediate risk or harm does not imply the absence of such in the future.
Meritocratic views are more prevalent in the West.
AI-2027 predictions have held up surprisingly well on the economics.
I think estimates of diffusion have been far too conservative. Compute costs are a bottleneck, and they've fallen at a pace orders of magnitude that of electricity during the Industrial Revolution. However, perhaps the rent seekers are much more effective at organising within institutions?
Whilst the market is faring poorly for junior workers, it seems robust overall.
Where AI talent comes from. Why we can't just shut ourselves off, and will need to cooperate.
Lucas Critique applies to forecasting too. What predicted well in the pre-AI era may not be the best practices today. I reckon as AI capabilities continue to improve, machine learning experts gain an edge as they understand best how the AI mind works.
AI-safety advocates need political victories. Either way, the two main powers have chosen racing at full steam given mutual distrust of each other. I don't think the labs can be trusted to self-regulate and coordinate either. This raises my p(doom) a bit back up to 10%. I think the much derided “effective altruists” and many in the rationalist community are right about this. The expected gains from superintelligence are unprecedented enough that we should continue to pursue it, yet the optimal pace is slower. At current capabilities, the US is a lot more powerful than any other nation here. That's why China will never agree to a coordinated slowdown designed to constrain it forever. Yet given the risks of totalitarian surveillance, the world will be thankful for the US being miles ahead. Moreover, in existential risk scenarios, the tools mankind has now to regulate superintelligence are insufficient, and this is why alignment is in large part a technical machine-learning and optimisation problem. We will need RSI and AIs to reliably constrain malicious agents.
No one has the incentive to unilaterally withdraw. A pause is simply unworkable. Like it or not, we are not turning back the clock unless we shut down the entire internet.
October 2026 links
