Through a vendor-and-ecosystem lens, this post’s strongest point is that hardware performance is only half the story; control and software gravity decide adoption. “Google’s dilemma” here lands because it names the awkward truth: likely TPU buyers are also strategic adversaries, and Google can ration access “year by year”. That is a commercial feature for Google, but a procurement risk for customers, amplified by the implied history of discontinuations (“Killed by Google”). On the competitive side, the post usefully challenges the lazy split that “GPUs train, TPUs infer”, citing Google’s TPU-only training for Gemini 3, and predicting convergence as GPUs add matrix-focused units and shed some FP64 emphasis. Still, the practical blocker is ecosystem: the cited SemiAnalysis excerpt frames open-sourcing parts of XLA:TPU and multi-pod code as a missing ingredient to challenge the CUDA moat. If Google can throttle supply and you’re betting the company on Ironwood, you’re basically asking for ulcers. And if you’re Google, do you really want to subsidise competitors just to grab some margin?

Talking about Stranger Things... No, not that one. This is delightfully over-engineered in the best way: proper hardware, proper control loop, and a clear attempt to open-source the lot. I like that it’s honest about where it stopped; habituation isn’t “the rats played DOOM”, it’s the bit before the hard science. Still, the design choices feel thoughtful: precise 10 µL rewards, a motorised trigger for demonstrations, and a modular frame that others can rebuild. The weak spot is evidence; if you’re claiming navigation and shooting, where’s the basic plots and trial stats? Cute concept, serious build - Love it! This is what we need in 2026!

I loved this; it’s half lament, half rallying cry. The metaphor of the car that stopped serving people is dead-on, and Henry’s plea for “convivial” webcraft feels like therapy for the over-algorithmed. Still, it’s easier said than done. Most people don’t have time to handcraft HTML between gigs and rent. But maybe that’s the point: a slow, local, joyful internet won’t scale, and that’s fine. His five rules sound almost pastoral compared to today’s platform grind. Hand-coded rebellion with a touch of poetry? I’ll take it. The essay’s ethical through-line, ownership, sustainability, and self-determination, reminds us that humane technology starts small, in one’s own digital garden.

Viewed through a geopolitics-and-industrial-policy lens, this letter is a critique of narrative failure among Western elites. Wang argues that US discourse misdiagnoses China’s technological rise by overemphasising cheating, subsidies, or demographic collapse, while ignoring industrial ecosystems, infrastructure depth, and competitive intensity. His scepticism of “decisive strategic advantage” in AI is analytically sound: it exposes how speculative timelines drive policy urgency without empirical grounding, distorting priorities towards compute races rather than societal deployment. The most policy-relevant claim is that AI leadership without manufacturing capacity, power generation, and diffusion mechanisms is brittle. By contrast, China’s whole-of-society approach, while authoritarian, is better suited to embedding technology at scale. The essay ultimately reframes competition as dynamic rather than terminal, shaped by agency, reform, and error. Its value lies less in prescriptions than in reorienting strategic attention from spectacle to systems.

This felt refreshingly un-hyped for a year-end LLM roundup. I like the framing that the clever bits have shifted from prompts to plumbing, as it matches what actually hurts in practice. The excitement is quieter now, replaced by cost spreadsheets, eval harnesses, and “will this break next week?” anxiety. That’s probably healthy. His attention to open-weight models is also well-judged: their narrowing performance gap meaningfully changes bargaining power, even if operational realities still favour large vendors. Where the piece is most valuable is in highlighting economics as a first-class design constraint. Rising API costs, shifting terms, and opaque roadmaps now shape architecture decisions as much as benchmarks do. Developers are asked to treat LLMs like stable dependencies while vendors iterate like startups. Until that mismatch settles, LLMs may be infrastructural in impact, but not yet infrastructural in trust.