Memorial Day Links

• Realistically, frontier LLMs are already strongly superhuman at basically all tasks requiring crystallized knowledge and pattern-matching reasoning, and they are increasingly getting better at fluid intelligence via long CoT reasoning. We are only a small number of breakthroughs away from AGI. When these breakthroughs happen is hard to predict (it could be tomorrow!, or a few months ago in secret, or in 5 years), but once they happen AGI will be here much sooner than we will expect. All the infrastructure for both training and inference is primed and ready. The data-centre buildout will likely reach its climax in the next 2-5 years, but my opinion is that the scale we have right now is sufficient for AGI and probably has been for the last few years. There is thus a massive compute and infrastructure overhang being created that will allow AGI to scale incredibly rapidly once created, at least for a while. Industry trends though are relatively closely (although not entirely) vibe-following and hence we may see a cooling of interest and excitement about AI at ironically precisely the wrong time. Broadly, this is because society is slowly assimilating LLMs into its worldview and understanding. This is going to cause some waves and change a lot of jobs but clearly not going to be utterly revolutionary in the way AGI would be. It will be e.g. like electrification or some large 19th century industrial revolution technology but not the creation of a god. People who are used to LLMs then forget the original promise and dream of AGI. But the AGI is still lurking there in the future, not so distant, and coming closer. [Beren Millidge]
• Scientists don’t yet know why the drugs might be having this effect. One theory is that GLP-1s reduce cancer risk indirectly—by driving weight loss and improving metabolic health, both of which are independently linked to lower cancer rates. The other is more direct: receptors for GLP-1, the hormone these drugs mimic, appear on the surface of some tumor cells, raising the possibility that the drugs are acting on cancer biology itself. [WSJ]
• My current, very rough, timelines are something like 5-15 years for AGI and then 10-20 years after that for widely rolled out and accessible indefinite life extension (ILE) for currently existing humans 4, accompanied of course with a large amount of other classic transhumanist technology. In the conservative case this is a 35 year lag time from today to ILE. Assuming an average life-expectancy of 80 then somebody age 45 today has a pretty good chance of reaching ILE and somebody age 35 has a very good chance. Age 45 (birth date 1980) pretty much demarcates the millennial generation from gen X. [Beren Millidge]
• Perhaps a more intuitive way to think about this is in terms of ‘zooming in’ to a fractal. As we progress down the power-law slope we zoom into increasingly small and detailed features. We can think of a neural network as similar to a telescope observing a region of space. Having a larger model is like having a larger telescope, it allows us to ‘zoom in’ to resolve finer detail. Similarly, training for longer is similar to having a longer ‘exposure time’ which also lets us resolve finer details for a fixed lens size. Due to the natural structure of the dataset, these linear effects get transmuted to power-law decreases in the loss. [Beren Millidge]
• We see that the three income categories had very different experiences during the March 2026 energy price shock. Low-income households increased their nominal gas spending by the least (12 percent). However, this was accomplished because they cut their real gas consumption the most (7 percent). On the other hand, high-income households increased their nominal gas spending by the most (19 percent) in a large part because they reduced their real gas consumption the least (1 percent). Middle-income households had intermediate increases in nominal spending and decreases in real consumption at gas stations. Thus, the K-shaped consumption pattern in both nominal and real gasoline spending was strongly evident in March 2026. These divergences in the response to an energy price shock are not unique to the month. Four years ago, energy prices rose and remained elevated during the spring and summer of 2022 when the Russia-Ukraine war disrupted energy markets. The magnitude of the initial Russia-Ukraine gasoline price shock was broadly similar to the current one, but it lasted longer to date and ramped up over time. As we see from the panel chart below, between January and July 2022, nominal gas spending rose more for high- and middle-income households than it did for low-income households, while real gas consumption declined less for high-income households than it did for middle-income and low-income households. Notably though, while directionally similar, the magnitudes of the gaps (both for nominal and real spending) were noticeably smaller than the corresponding gaps we see in March 2026. [Liberty Street Economics]
• Although this all seems hyper speculative, these kind of dynamics can be important for modelling the far future. Specifically, it implies that once a K3 civilization is established in a galaxy it is extremely hard to dislodge. This means it is likely that the universe will end up mostly cleanly partitioned between full galaxies of colonizers with perhaps only short wars early on in a galaxy’s lifetime to determine which colonizer faction will be victorious in galaxies where colonizers both reached it at roughly the same time. Beyond that, the fact that intergalactic warfare seems likely to be expensive, interminable, and bring little prospect of gain to the aggressor, makes it likely that such wars are rare in general and only pursued for ‘irrational’ purposes. For instance, once the colonization phase is complete, a pure paperclipper would be unlikely to invade neighbouring non-paperclipper galaxies since it would have to expend far more energy to do so than it would gain, and hence would end up with less paperclips overall than if it just paperclip-ified its own galaxy and then sat there for the rest of time. This also implies it is very important for early civilizations to colonize to maximize their share of the cosmic endowment because it is hard to fight back and claim inhabited territory if you are late to colonizing. Another thing of importance to note is just how extremely vulnerable large fixed objects like planets or stars are in this universe. Anything on a predictable orbit can be sniped with impunity and complete surprise from millions of light years away. This means that humanity is super vulnerable in its early stages right now and also even when spread across multiple planets and building its first Dyson sphere. Any K3 or less civilization in the vicinity (if they see us) can directly destroy us at this point, which makes moving some sensible fraction of civilization away from obvious large orbiting bodies and into freely moving habitats far away from any obvious orbits an immediately important backup strategy. Unsurprisingly, it also means that we should invest heavily into trying to determine if there are any K2 or K3s out there in our intermediate neighbourhood (e.g. perhaps in the local group) since if so not only will they likely have colonized almost all galaxies we could reach first but also that as soon as they detect meaningful civilization around earth, then they could disintegrate it with extremely powerful beams with ease. In a variant of the anthropic argument, the fact that this has not happened yet provides some evidence against hostile local K3s existing. [Beren Millidge]