Some Links I Consumed in September 2026

1. London’s schools are running out of children

   Things look pretty bad here in the UK: can’t build homes or data centers (or anything really!) and now there’s not just evidence of people not pro-creating to underscore the direness. On the bright side, we don’t need to worry about what AI will do to kids’ minds.

2. Lunch with the FT: Mariana Mazzucato

   I enjoy reading atmospheric interviews which lends more personality to the entire piece.

   Now, I’m somewhat sympathetic to Mazzucato’s take that the private sector has derived value (“leeched” is her word) from the public sector.

   I have previously commented on SpaceX’s reusable rockets being an extension of NASA’s retro-propulsive landing algorithms for placing Curiosity on Mars. Such technologies are not developed in isolation, but by teams of individuals with (private and/or public) institutional support. Geographic migration of individuals into such systems also matters.

   Consider the space sector, which she talks about here. One can argue that SpaceX exceeds NASA’s engineering capabilities and performance, but was also clearly built on NASA’s know-how. What sensible mechanisms would allow the US government to capture some of SpaceX’s value? I don’t anticipate SpaceX wanting to fund universities directly, but maybe they could be mandated to contribute to NASA’s Space Technology GRO program.

   But let’s also consider individual immigration. Take again SpaceX’s reusable rocket program; spearheaded by Lars Blackmore, who not only moved there from JPL, but was educated at Cambridge University (before a doctorate at MIT). How is the UK to capture any value from his work at SpaceX?

   Now multiply this by the thousands, as migration has been a key ingredient in the success of the US space sector (especially pre-ITAR). My point here is that one can imagine hitting some limit where one state capturing value can leave another with a sour taste. I suspect equalising mechanisms into an already complex machine can only retard growth, which she also wishes to deliver.

   My intuition — especially for space technologies — is that, absent the Cold War-like conditions that can seed a Space Race, states should do two things: perform in-house speculative research, whose pay-off appears to be beyond a decadal time-scale; and continue to create programs that lead to step changes in existing technologies. The former could seed long-term technology development programs for large space stations with manufacturing labs (robotic or crewed).

   NASA’s COTS and Commercial Crew Program are beacons of the second approach; they helped SpaceX develop Falcon-9 and Crew Dragon. The commercial payoff was clear and, if successful, immediate. This prevents duplication of capabilities while keeping a plurality of programs operational.

   SpaceX’s success could serve as a blueprint for others as they eventually also started creating new markets. After mastering reusable rockets, they found its most profitable use (at least for now): internet from space with StarLink satellites. This is a wholly privately funded achievement where they not only proved the tech stack, but also created its market, end-to-end. AI’s development at frontier labs looks similar; while early research can trace to DARPA and CIFAR, the work on LLMs can be said to have happened completely independently of the state.

   Rapid recent advances in both space and AI suggest we are returning to a time where technology innovation happens in private monopolies. Avoiding the monopolistic dynamics that lead to a concentration of power is more important than to think about playing zero-sum games of value capture. Graeber thought the Bell Labs model of a regulated monopoly worked well. But Bell Labs also stopped being a technical powerhouse so I’d say the right answer probably is yet to be discovered.

3. Lec 16 Pedagogy in the Times of AI

   It was unusual to watch this hour-long university lecture from an academic grappling with the existential void that the rate of progress has created. The candor is refreshing. His bleak tone was different to what I perceive at my university. I suspect many academics just aren’t aware of how fast things have moved in the last 9 months and I can’t fault them. Keeping up with AI progress feels like a full-time job.

   This academic says he has stopped recruiting research assistants because he sees them as “bottlenecks, not enablers”. Exams have started to lose meaning in his course because agents can complete postgraduate-level work now. So, he goes on to question the purpose of lectures, thinking their in-person attendance might provide some mental health benefits — much like gymming supports our physical health.

   I have also been worried that a higher education system predicated on exams that coding agents can complete in minutes (at best!) increasingly feels like a charade. In-person or oral exams can work up to a point, but scaling that for classes with hundreds of students with 1-2 examiners feels daunting.

   But that’s not even the real problem; however small the chance, the existential void post-graduation bothers me for countries like the UK that offers engineering services. It’s more concerning than existential risk. Something has to happen; discussions need to happen, but for that people in the ivory tower need to wake up.

4. The Boring Part of Bell Labs

   Elizabeth Van Nostrand interviews her dad about Bellshit jobs, i.e., the essential work at Bell Labs on telephone exchanges which let their research division flourish. Looks like Bell Labs had a program to send select candidates to school on roughly 60% salary and pay for their tuition/books towards a Masters degree (you were fired if you didn’t finish in a year). This is similar to what I suggested above that SpaceX wouldn’t do, but looks like Bell Labs did. Bell Labs also gave the option to stay on fully salaried while attending grad school part-time.

   Unlike the modern perceptions on advanced degrees, it looks like they mattered at Bell Labs:

   Bell Labs was big on degrees. To be a Member of Technical Staff you had to have electrical engineering classes and a Masters degree or on a path to get one. They were willing to pay for this.

5. How do you simulate Mars on Earth?

   Nice reminder that there is a company selling simulated Mars dirt, called Space Resources Technologies.

   The excess perchlorates on Mars also reminded me of its history in solid propellant, which I spoke about in here.

6. AI Safety Is Mostly A Sex Cult In Berkeley, California

   A hit-piece that blew up on X, but far too long to read deeply (for me). I had two thoughts. First, the manner in which AI safety has gained political mainstream relevance with Bernie’s collaboration with the rationalists is something deeply American. As a country, it is truly one of a kind where you can just do things. This is both a blessing and a curse; a feature and a bug.

   Second, the author’s accusations are immense. I think there was enough interesting “evidence”(mostly tweets) to maybe make other kinds of claims about other people, but the piece wildly extrapolates conclusions that paint the writer as bitterly vindictive. The internet is a hot-takes hotel, so that’s what we get served and the piece had its moment. Now, the post has been removed from Substack, suggesting there might be legal ramifications.

   I was at Inkhaven in an emergent blogging capacity, but I am neither rationalist, nor EA. I am also not an AI safetyist, though I have recalibrated to being more techno-cautious about it. Maybe these are reasons why I was (thankfully!) not in the thick of the politics. Though I did feel out of place, I left feeling there was much to appreciate about Lighthaven and that Inkhaven was a net positive life experience.

7. Signals and Threads Podcast

   A good listen, where Jane Street’s Aaron Bauer says there’s value in both manual notes and lectures (with elements of active learning). The conversation pivots towards teaching in the agentic era and closes off with learning outcomes. I felt like it was interesting and would lead somewhere specific, but on reflection, I don’t think there were any real takeaways on how an instructor should approach teaching and assessment in large university classrooms. Of course, we want everyone to learn, but like the IISc prof highlights: if we assume everyone is genuinely learning, there is still the tacit mandate of ensuring a Gaussian distribution in grades. So, I can see the appeal of being a developer educator at a company where your goal is to devise and test ways for everyone to learn and progress. But what is going to happen to more formal education systems is anyone’s guess.

8. Nassim Taleb — Meditations on Extremistan (#158)

   Listened to it as, every now and then, I enjoy a little bit of that Taleb energy (but not all the time). Unusually mathy by podcast standards that I have the tab open to work through the transcripts.

9. Google Is Sending an A.I. Data Center to Outer Space

   This NYT piece on Google’s attempts at space data centers took me back in time: UC Davis’ Crocker Nuclear Laboratory tested AI chips which will be onboard the satellite that Planet Labs is helping Google build. I have biked by Crocker almost every day that I was in Davis as it was the neighbour of the Mechanical and Aerospace Department, where I got my PhD.

   This first step towards Google’s space data centres is a fridge-sized inference satellite, called MVP, with four Trillium TPUs and roughly a kilowatt of solar. It will ride a Transporter rideshare to a sun-synchronous orbit, like their paper’s 650 km dawn-dusk case. Operations are planned for a year, but the satellite will re-enter and burn up as it descends through the atmosphere after six years.

   Looks like Google is trying to find an edge in the AI scene so have moved faster than intended; they were waiting to launch two satellites in 2027, but have decided to go into orbit this year. The chips run for about 15 minutes before needing to be shut down so they can cool off:

   The bottom layer is made up of Google’s A.I. chips, which sit on a green motherboard. The next layer consists of “thermal interface material,” a pale green putty that comes in sheets like Fruit Roll-Ups and connects the chips to layers of aluminum and copper, radiating heat away from the motherboard. Finally, there is a radiator panel, which projects heat into space.

   I have previously commented on Project Suncatcher and examined Starcloud’s mission proposal in detail.

10. A Finishing School for Founders

    A16Z might force Western higher education to see itself as one of two things:

    • a marketplace for employers and jobseekers;
    • and/or an incubator for future companies.

    I have previously briefly said that Western universities are a poor value prop; this is more of a consensus view today than it was 3 years ago, when I first said it, cf.,  last week in the Guardian. Now, engineering colleges in India have played the role of such a marketplace for a while, but I never quite got why this wasn’t a thing in the US and UK (and I am not talking about job fairs, which are merely networking events) where degrees are far more expensive. Seems like an easy solution for STEM degrees, but is unlikely to address the issue raised in this piece about what happens to the arts and the humanities.

11. AI and the Fall? of the Creative Class

    Things look grim in the digital arts, but it looks like the creative industry will need to re-learn how the arts have navigated previous tech shifts by, for example, focusing on live music concerts and writing novels; both have been value generating. So, maybe we will see an increase in creative outputs (from humans, not just agents) purely because people will need to create/sell stuff (ideas that become books rather than working in publishing-adjacent work) bundled with more live activities to engage with community; perhaps this is not a bad thing.

    Now, I might be extrapolating far too much from here, but I think, at some point, universities will also reckon with this. As research gets more automated, the university’s role as a certifier of capabilities will be seen as more important. Recruitment scale will still matter for fee-charging universities, but attracting students through excellent teaching will require experts who bring more charisma and interaction to lectures than students typically experience today. It is from such teachers that the next generations, who will spend even more time on screens than the present generation of students, might come to learn the value of IRL interactions. Exams will not just be testing kids on what they know, but also their interpersonal communication skills. Why will that matter? I think more jobs will eventually undergo a transition like the arts, where outputs will need to be bundled with more elements of live presentations. Lectures could also become more of a production with the use of agents and classrooms might become more interactive (even at scale). While I question if universities will survive this moment, I think it’s unlikely that we will see the system vaporise. But adapt they must.

    Of course, all of this is reliant on our capacity to produce little humans at a high rate (see top link on London’s declining schoolchildren).

12. A.I. Data Centers in Space? A System to Cool Chips Could Help.

    This is a good companion to the Google piece above because it mentions another company—Sophia Space—working with Caltech on space data centers. Their cooling technology is described as:

    uses a combination of tiles and solar power, which ultimately sends waste heat into deep space.

    These are all vague statements that don’t tell us enough about what the system is like. I suspect it is because no one has a good working answer of how to address the heat rejection problem. Hopefully, it is just an engineering problem.

13. Jensen Huang Thinks A.I. Alarmism Has Gone Too Far

    I thought this was an excellent interview with two people with excellent media skills for their respective roles. Ezra was unrelenting in grilling and came across very concerned about AI in a very informed manner—a rarity in the mainstream media scene, but perhaps unsurprising for him given the blogosphere roots. Despite the heat of Ezra’s grill, Huang came across a lot more measured than combative, which is what he looked like on Dwarkesh’s podcast. That said, I found his analogy of virtual machines being built because software has historically escaped sandboxes to be a weak one.

    It also felt like Jensen uses language that anthropomorphises AI while chastising others for doing that.

14. An Introduction And Conclusion To My Essays

    I’ll have to read the blogposts sometime, but seems interesting.

15. Edinburgh joins UK universities heading for strikes as cuts spread through entire sector

    Grim times and the reaper is on the march, in search of a prized scalp in UK higher education.

16. What you need to know about Your First Home scheme for first-time buyers

    Deeply skeptical of schemes that make buying homes possible without making them affordable. Shared ownership was another terrible example, IME. These are “schemes”, in the worst sense of the word.

    Read about my shared ownership flat experience here.

17. Burnham’s reheated Help to Buy scheme is a bad idea

    Yup.

18. The Curse is a fully bizarre and brilliant maze of a show

    The Curse is my conscious entry onto the shores where Nathan Fielder works; I had watched Who is America? without knowing of his involvement in it, but the stylistic parallels to Sacha Baron Cohen are evident. That said, Fielder’s work is completely novel. And brilliant!

    The first and last episodes of the Curse are some of the best things I have seen across TV and cinema. The Curse felt like the most ambitious thing I’ve seen though I can’t say I really enjoyed it; but it did really keep me coming back for more. I’m now watching The Rehearsal; it is genuinely brilliant and I haven’t laughed this hard since Who is America?. It might also replace The Curse as the most ambitious thing I have seen in a long time.

Here are some other things I skimmed:

1. We were trying to terraform Mars but instead we saved the sea snails
2. Using AI to make teaching easier & more impactful

In all, I consumed at least 20 articles and at least 102,424 words this month.