This is excellent! Check it out.
Fascinating theory that stablecoins essentially behave as treasuries (well they tend to be backed by them after all).
This is significant. AI might lead a more rational and better informed electorate, hence potentially overcome most of the institutional problems in our democracies. AI can empower autocrats, but also empowers democracy. The extent to which AI does the latter over the former is a big alignment issue. Pieces like these are also why I continue to use Substack, despite its relative decline in favour of X (articles) in recent months. Clearly enough people consider Substack worthy enough to post there, and when such articles are cross-posted on X, the quality of discussion on X benefits too. These platforms are complements not substitutes.
RSI: coming soon.
Excellent piece by Alex Tabarrok on the UK's Equalities Act.
Parallelisation bottlenecks in an endogenous growth model.
One possible mechanism via which the cybersecurity risks of AI could generate existential risk - increasing the probability and intensity of conflict.
So much for postliberalism. Turns out the median voter isn't economically illiterate after all.
A new AI forecasting tool that matches the performance of the top superforecasters.
More evidence of fraud by Arday continues to accumulate.
Our police are basically encouraged to behave as those in third-world banana republics. Of course we converge in terms of state capacity too. Every long-term reader from the start knows my experiences with politically motivated arbitrary arrest. Harassment law increasingly complements the ban on “grossly offensive” speech, “malicious” communication, and the like. Note I'm not trying to single out the police here - the entire British state has been engineered to blatantly promote left-wing socialist goals, even despite the last decade + 1/2 being under Tory rule. Simply put, our regime increasingly resembles that of an Eastern Bloc state. By the way, we are exporting our authoritarianism abroad, just as you'd expect from Russia or China.
Is chunky post-training a source of misalignment?
Security incidents occured in 8% of runs in AISI's testing of Mythos…
Human connection is still vital, so the relational sector will still be salient. I'm not sure there's any demonstrated robust link between AI chatbot use and loneliness either way though.
A case of Yudkowsky being right on the limits of alignment training and sandboxing? Credit where credit is due. Of course one might judge differently his overall forecasting record.
FutureSearch predicts Labour is most likely to win the next GE, with a slightly less than even chance. Reform's odds are almost on par with the Tories'. I think there'll be enough vote-splitting on the right such that Labour will pull through the middle.
Now imagine AIs being integrated by malicious actors into this cheap drone infrastructure, and why global cooperation is vital to minimise existential risks to an acceptable level. We solved this issue for nuclear warfare. Will we achieve the same for AI? Perhaps first we can reorient our priorities away from Iran and back to Taiwan, such that we can bargain for an agreement with strength? The one thing Trump was supposedly good at (foreign policy), he's turned to be shit at. In his first term, conventional conservatives were basically running things in the background. Now we lack that constraint to his stupidity.
This is one reason (amongst many) why RSI may be net beneficial to AI-safety, as opposed to guaranteed doom. Perhaps the best way to check a misaligned agent is an aligned agent. We should drop the inaccurate assumption that agents will yield homogenous preferences.
Intersetellar travel will be one of the gains from superintelligence. I reckon in the next decade, the idea of forming colonies on Mars will not seem so outlandish. A permanent Mars colony becomes more likely than not in the 2070s. This is within our lifetimes! Note that we'll travel to Mars well before that date too.
A shift from property and capital taxes to land taxes is likely to be welfare improving, and promotes growth via urban development and agglomeration effects.
Alcohol deaths per capita have doubled in the US within my lifetime. I wonder if this is a result of drinking being more normalised there: the prohibition hangover (no pun intended) in attitudes dissipates, plus lower religiousity? Although I write in favour of moderation, I do wonder often the extent to which we actually best achieve this via ascribing as negative an attitude to drink as possible. This is still a low confidence topic. We should however be taxing alcohol (as well as regulating its public consumption) a lot more than we currently do, as surprisingly, Pigouvian levies on alcohol work and reduce the associated externalities.
Addiction is one of the greatest social and moral challenges of our time. If you believe that we're entering Industrial Revolution 2.0 at least, you get some of your biggest returns on lower discounting. This implies less consumption today and more saving (including in human capital development: a kind of saving), at one's binding constraint. A GLP-1 agonist level cure for addiction is also probably our best hope as opposed to mandates or paternalism or shaming (as opposed to reason and persuasion), as the former worked for obesity when the latter failed. I still think virtue ethics tends to be overrated though, at least from the standpoint of what is individually optimal. I think my perspective is somewhat Epicurean here.
There will likely be a big rise in income inequality with AGI.
I think the lesson from this is that imitation is a lot easier than innovating at the frontier, so this makes me somewhat optimistic that Chinese open-source won't eclipse US frontier labs. This might be why Singapore became rich despite the fact they're far from a paradise of free markets and liberty. Maybe you don't need that great a government, maybe just elites smart and capable enough that they don't fuck up too badly. Or won the jackpot, which can paper over slow progress on structural reforms. I think we have to accept as our prior at this stage that democratisation tends to be a net benefit.
Cronyism in the Trump administration: the same presidency that's running a memecoin from the White House.
Increase (C)AISI funding, and how do we get the labs to invest more in safety? Should we encourage them to do so?
If you subscribe to nativist and protectionst views, then you should be aware that you are free-riding on many of the gains of global integration.
Results on distillation. There is a lot of uncertainty regarding how this equilibrium will play out.
One of the bottlenecks we talk about. Since Fable, I've spent minimum $100 in usage credits per month. However much we know about capabilities, there exists rational inattention (like Lucas-islands or a Mankiw/Reis-style PC) and not all are sold on it, so many are unwilling to pay.
See the discussion in the replies and you have the best criticism of IQ testing. There is implicit conditioning on question context, albeit good design can mitigate this somewhat. I reckon this is why those with ASD tend to fair better on Progressive Raven Matrices.
They obviously don't read his links or book recommendations.
This is amongst the worst things the Trump administration has done.
An example in the innovations in mostly private governance that I often discuss as the means to regulate AI harms.
List on generalisability.
I would be tempted to use these results to justify using AI as a finishing polish: the lightest touch editor possible. However, I wonder if we lower the stigma associating with using AI to change word choices etc that we'll start to see more homogeneity in voice. Like it or not - I have a distinctive voice! I plan to keep it that way, even if on the margin we see more minor spelling mistakes. On the other hand, you could say the same about responding to feedback from peers, colleagues or mentors. Perhaps society seems to prefer certain attractors and some convergence in such?
This is how to use AI. As a complement to “raise your ambition”, not a substitute. Too many people use AI for the latter, or as some therapist. UK public-sector workers are banned from using AI in their workflow except as copilots! Adoption, particularly of (multi) agentic work, will be incredibly slow, and this will limit growth rates despite exploding capabilities. Currently I have AI adding say an extra percentage point to growth by 2030 as my median estimate (wide confidence interval), even if capabilities are at the level implied in AI 2027.
If AI fully automates most jobs, then you'd better take my advice to save a lot more and be more entrepeneurial. In such a scenario, we'll be a lot richer but compensation mostly capital based. Also prepare for more taxes and transfers as insurance (I think most redistribution we justify more on insurance than on equity concerns) for those with insufficient capital.
Free markets in healthcare work. US system does well even despite much public intervention screwing things up.
Redistribution tends to affect consumption more than anything. Note that a higher propensity to consume could be locally state-dependent, and as soon as households move to income and wealth levels where they are less credit constrained, we will see their MPCs fall. Globally, the net welfare effect is too ambiguous to be left for governments to deliberate on.
Yes. The incomplete capital markets union, with many product-market barriers still existing in what is supposedly a single market, is holding back European firms relative to the US.
This could also be interpreted in favour of land value taxation.
Work on pandemic preparedness, given the bioweapons risks from AI, is probably as important as AI-safety.
Humans are uniquely intelligent and therefore, unlike other animals, capable of moral reasoning. Hence we should attach to humans much greater moral worth. Only one species is actively trying to prevent suffering for the others. Of course this does not mean that we should ignore animal welfare. Factory farming is by far one of the greatest atrocities of history. Yet we needn't kill mankind to save the others. If you fear AI being exploited by malicious hackers, then avoiding misanthropy is a must.
September 2026 links
