Links for November 2026
Landing on Phobos, (slightly) faster Fourier transforms, and optical LLMs

Yankees in Petrograd is a 1924 Russian satire of American pulp thrillers, involving a plot against the life of Lenin by the international forces of capital - which must be thwarted by Mike Thingsmaster, an American trade unionist and leader of a secret international agency that defends the interests of the proletariat.
Breakthroughs like the recent OpenAI Navier-Stokes result seem important as indicators of AI progress, even if they aren’t useful. They’re used as discursive cudgels to this purpose - how can you deny that the singularity is approaching when models are making progress on Millennium Prize problems? So it’s important to understand what these breakthroughs actually consist of. There are some well-known caveats already, like the Navier-Stokes work having grown out of academic research done by humans. Another post in this vein I enjoyed was Faster Fourier Transforms, focusing on OpenAI’s claimed discovery of a faster method to compute Fourier transforms. This got my hackles up immediately: a true speedup in computing FFTs would have enormous real-world impacts, since they’re used everywhere in digital signal processing. But this isn’t what the OpenA result is. Instead, it’s a hack that allows for a tiny speedup in an edge case where the current algorithm would take an impossibly long time anyway. Many of the recent results seem to be like this. On the one hand, they do demonstrate a mastery of advanced mathematics, but on the other they are essentially mental backflips whose most useful application is the disproof of some previously assumed limit. It feels as though the AIs are currently differentiated by their willingness to spend hours prying fingernails into cracks in a way no human would bother with.
Another interesting perspective to take on AI progress is to ask: how disempowering is our current civilisation? This an argument of handover advocates - that we have already handed over power to our society in most important ways, and that Opus 11 will just be more effective management. I’m still confused about this myself, but it’s important that I (and my peers) can actually contribute to civilisation. It wouldn’t exist without us, and so as a construct it’s interested in keeping us around, which is not necessarily true of AI management. You only need humans around to plug your servers in for so long.
Is China going to be the real victim of AI-discovered cyber vulnerabilities? This is a convincing argument which matches my own experience. China is good at top-down programs like water treatment plants or forcing companies to disclose vulnerabilities to the security services, but bad at bottom-up programs like district water pipes and forcing random firms to update their Kubernetes clusters. There’s a bearish picture coming together here - economic slowdown brought on by poor financial management, overbuilt infrastructure for a shrinking population, and their only advantage, the factories, now vulnerable to a deluge of automated cyberattacks.
Traditional telecom satellite manufacturers are now pursuing moonshot projects to develop their own launch vehicles. The launch market is in an extremely weird place currently. More mass is being launched into orbit than any previous time in history, by far, but almost all of this mass consists of Starlink satellites. The enormous volume of Starlinks is eating all the Falcon 9 launch slots, and the industry outside of SpaceX is stagnant: depending on how you slice it, there are actually less non-SpaceX launch slots than in 2000. Competitors like Blue Origin’s New Glenn are on their way but have experienced setbacks. Meanwhile, there is huge pent-up demand, mainly due to Starlink competitor constellations like Amazon Leo. Traditional launch capacity was exactly sized to the market with very little elasticity due to the difficulty of, well, rocket science. But if anything SpaceX is moving away from serving this demand. Falcon 9 is planned for a rapid phaseout, to be replaced by Starship, which is designed to only launch SpaceX flat-pack Starlinks and AI sats at this stage. Its enormous capacity will be fully taken up by SpaceX’s expansion plans. So the rest of the market is now scrambling to find launch slots wherever they can - and if they can’t, to build it themselves.
A Japanese mission launched last month which aims to land a rover on and return samples from Phobos. It’s interesting for a few reasons:
This is an extremely complicated mission. There’s all the components - a propulsion module to get you to Mars, the lander, the rover, the sample-return propulsion and entry capsule - and don’t forget the actual instruments! Then all that has to be able to operate autonomously. Unlike a normal asteroid-rendezvous mission like OSIRIS-REx, Mars will constantly be getting in the way of a direct link with Earth. Then there are the orbits. Phobos is too small to actually orbit, so the plan is to enter an egg-shaped Martian orbit and perform repeated intercepts of the moons. The rover will then be released during a close pass, around 50 metres above the surface.
“Landing” on an asteroid like Phobos is something of a misnomer. Bodies like Phobos are the planetoid equivalent of liberal democracies: loose aggregations of particles that vaguely hang together. So the rover uses paddle wheels to swim across the gravelly surface, travelling incredibly slowly to avoid breaking the grip of Phobos’s microgravity.
Phobos is interesting for many of the same reasons Mars is interesting. It should tell us a lot about how Mars itself formed and evolved; Phobos likely contains a lot of ejecta captured from Mars, so the samples returned by MMX might end up containing our first Martian samples as well. It’s certainly a better prospect than the boondoggle that Mars Sample Return has become.
You’ve seen “baking a transformer’s weights into silicon to serve inference thousands of times faster”, but what about “baking a transformer into a sheet of glass using micro-deformations”?
Open-source code for geolocating almost any audio using doppler curves from the sound of airplanes passing overhead.
This excellent post on doing tarot readings, how to interpret various parts of the deck, and the pitfalls of symbolism finally made tarot ‘click’ for me.
John Portman’s delightful neofuturistic hotel designs, a sort of bio-inspired brutalism that almost reminds me of Giger.