2025-10-01
something spiritual
Cartoons Hate Her has a (partially paywalled) post on pronatalism being incompatible with our current high intensity childrearing culture. Something I’ve been considering lately, based on Alice Maz's most recent Xunzi essay on a new ideological system based on virtue in leadership, is the importance in the western liberal order of institutions, which explicitly deny the need for great or virtuous actors, but are instead self-regulating using the forces of self-interest and mutual cooperation and competition. The reason for Alice’s turn towards virtue is the belief that such systems need to evolve to avoid stagnation, but in doing so open themselves to corruption which eventually entrenches a hereditary elite. Something which occurs to me is that this is actually not what is happening in the west: it is not remotely close to China, where the party elders have systematized the passing of power to their designated princeling successors. Instead, liberalism has essentially recreated Gavelkind succession by promoting individualism, diversity, and equality, such that parents distribute their resources more or less evenly across all their heirs spread throughout different domains, diluting their ability to concentrate wealth and power. With regard to pronatalism, one of the downsides of this approach is that the natural response to Gavelkind is to reduce the number of your heirs, and this is presumably why as liberal societies develop, the natural instinct towards dynasty building will cause a reduction in fertility rates. This implies that one solution to the fertility crisis might be to normalize ultimogeniture1. Essentially, the rule would be for parents to always focus their resources on their youngest child, and elder children would be expected (either through culture or because of the incentives) to help raise their juniors. To some extent, this is what is already happening among modern large families, and as other less productive strategies die out, I expect it will eventually become the norm.
Dwarkesh Patel reviews The Vital Question, a preview for his upcoming interview with Nick Lane.
Hard Drugs podcast on how the field of protein design has been advanced by AI over the past couple of years. I wonder to what extent the unexpectedly good performance of biotech companies recently might have to do with these methodological improvements, which mean that the success rates of different milestones are higher than they were in the past.
Ryan Khurana argues that AI video generation is useful for AI to build intuitions about how the physical world works, which can be applied to areas like robotics. This is possibly a response to the current backlash to the Meta and OpenAI video generation products. But more interestingly, I think it’s another point against the Richard Sutton position that LLMs are sending AI research down the wrong direction, since next-token prediction clearly bootstraps human intuition, in ways that we want. Andrej Karpathy has a good analogy here, that we are not trying to build animals, but “summoning ghosts”.
Andrej Karpathy on the release of Tinker from Thinking Machines2, another sign that individualized fine-tuning is the next frontier for AI becoming broadly deployed across all applications. He raises an interesting question as to what extent you want to fine-tune for each task, which brings to mind this interesting tweet from patio11 on the distinction in Japanese between 教科 and 科目. Personally, I don’t think this is a hard rule, but more a result of the current LLM architecture being dominated by MoE models, which will be wasted by expert collapse if you fine-tune on too narrow a task.
Ben Shindel investigating the nature of persuasion and “super persuasion” using prediction markets. Somewhat related, Zvi on safety evaluations of Sonnet-4.5, and its capabilities.
Bari Weiss interviews Leland Vittert on “overcoming” his autism. For a while I was under the opinion that autism actually represented the beginnings of a speciation event with assortative mating, but now that AI is taking all the programming jobs and elevating in-person interaction3, I expect autism rates will collapse over the next couple of generations.
Richard Hanania on how liberals and conservatives perform gender.
Slime Mold linkthread.
Taking it even further, it is the youngest child who apprentices to their parents, takes over their businesses when they retire, secures them their dotage, and then inherits their assets upon their death, with older siblings either attaching themselves to the youngest as subordinates or making their own way in the world.
Why not the more historically common system of primogeniture? It seems clear to me that ultimogeniture is far better suited to our current conditions compared to the historical primogeniture: birth control gives some degree of control to parents to designate an heir; rule of law prevents usurpation; and longer lifespans means that the eldest may not be doing much better than their parents when it comes time to take care of them.
There was a funny description of Thinking Machines as “the perfect AI startup” by Matt Levine, describing how they were able to raise 2 billion dollars in funding despite their pitch never mentioning any specifics on what they are working on, based purely on having a collection of valuable employees.
Edit: what does it mean that Derek Thompson is using Thinking Machines as an example of the AI bubble now, given that they released a product just before his post? The crux is probably to what extent you think the journalists matter: this could be another case of them fundamentally misunderstanding everything as outsiders; or maybe they matter after all and actually forced a response from the industry to justify themselves.
The combination of these two in particular is deadly, because even if these jobs continue to exist and you are better than others, it still won’t really matter.