AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023. That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity.

A classic mathematical theorem. 6,000+ lines of Lean code. Verified by the kernel. Xiaomi MiMo 2.6 Pro assisted researchers in fully formalizing the original main theorem of Li–Yorke’s classic paper Period Three Implies Chaos in Lean 4. Guided by a research-designed exploration strategy, multiple Subagents collaborated on theorem formulation and proof formalization. After revision and integration, the project spans 6,000+ lines of Lean code, verified by Lean’s kernel with no unfinished proof placeholders. Notably, Xiaomi MiMo 2.6 Pro was not specifically post-trained for Lean. #XiaomiMiMo# #XiaomiMiMoV26# #AIforScience# #XiaomiAI#

Carnegie just dropped its new AI Talent Tracker and the shift since 2022 is pretty striking. Among the NeurIPS researchers in its sample, 57.4% did their undergrad in China, up from 46.3% in 2022. The US fell from 19.8% to 13.3%. But the more interesting change is where they now work

At its default effort setting, Opus 5.5 delivers frontier results for a fraction of the cost per task, often beating other models running at their highest settings. It also generates output more than 30% faster than Opus 5.

Xiaomi makes this 1500hp EV supercar that holds test track records in Germany, and has just shipped a credible frontier model after a public RL training demo. Claude says: Xiaomi is the first major lab to run its production RL in the open. It streamed the run live and disclosed the key numbers: about 750,000 trajectories, 30 steps in under six days, roughly $2.62M for Pro and $0.85M for Flash. It says it's releasing the training environments and RL code so others can reproduce the results. CodeMidas shows that filtered, verified tasks beat larger unfiltered sets. Open-sourcing such environments lowers the barrier for every open lab and pressures closed labs whose advantage partly rests on proprietary RL data. V2.6-Pro is the top open-weights model and ties Grok 4.7 at 46 on Artificial Analysis. It still trails Claude Opus 5 (51) and GPT-6 Astra (53), with the widest gaps on hard long-horizon tests like Terminal-Bench 4.0 (34.9 vs 49.0 and 59.6). Its edge is cost: it's roughly 45× cheaper per task than Opus 5.

I was asked by Congressional staff to share my views on open models performance, adoption, and competition vis a vis China. Here’s my briefing with all our latest data and predictions for what comes next. I am very happy to spend time on broader audience content on open models!

Who is actually driving global AI research? A new deep analysis by @TheCRC_Substack of over 37,000 authors at top-tier AI conferences (ICLR, ICML, NeurIPS) reveals a staggering shift in global talent dynamics. Where are these elite researchers getting their PhDs? Domestic Chinese universities have recently exploded in capacity. Looking at a recent cohort (2023): • Matured in China: 783 top AI PhDs graduated from domestic Chinese universities. • Matured in the US: 380 mainland Chinese researchers received their PhDs from US institutions. China is now training more than double the number of elite AI PhDs at home compared to what it sends to the US. Read the full deep dive here: 🧵👇

Two weeks. That's how long it took to go from GLM-5.3-Flash's first run on domestic accelerators to serving all of its production traffic, with 3.2× end-to-end throughput along the way. What I keep thinking about is who did much of the work: an Infra Agent powered by GLM-5.3. A model helping optimize the system that serves it. The conditions were hard. Limited memory and interconnect bandwidth. 1M-token context. Multimodal requests. An immature software stack where kernels were missing and documentation was often guesswork. Every optimization was a trade: compute for memory (ReplaySSM), communication for memory (intra-node tensor parallelism), precision for capacity (mixed INT8/FP8/BF16 caching), and disaggregation for scheduling freedom (Encode–Prefill–Decode). But the most important lesson wasn't about any single optimization. When the agent got stuck, it was rarely because it couldn't write the code. It was because it didn't know *why* things got worse. "Throughput down 20%" tells you something broke. It doesn't tell you which layer, which hypothesis, or what to test next. In RL terms, it's a sparse reward with a credit assignment problem. And an end-to-end benchmark that takes hours makes exploration painfully slow. Senior engineers solve this with an implicit process reward in their heads. They know when to check the timeline, when to run a microbenchmark, and which layer's output to compare. So we made that explicit. We call it dense feedback: layered verification interfaces the agent can call directly. Correctness feedback: did it compute right? System behavior feedback: where did the time go? Performance feedback: which option wins, under which conditions? Each signal has to be local, cheap, and objectively verifiable. Three things the agent found: First, precision drift in KDA's context-parallel path that grew with sequence length. The cause was TF32 rounding error compounding through chained state-matrix merges. The fix is now merged upstream in Flash Linear Attention (PR #1180#). Second, KV transfer never overlapped with DeepEP dispatch. The agent followed the call chain across the Python/C++ boundary and found that the intranode path never released the GIL. After the fix, transfer overhead fell from over 30% to under 1%. Third, a decode kernel recomputing the same normalization four times because of how it was chunked. The agent restructured it and got a 1.71× speedup. The idea came from "optimization skeletons" it had distilled by reading existing kernels across SGLang, FLA, and DeepGEMM. To be clear about the boundaries: humans still defined the goals, built the feedback environment, and reviewed every high-risk change. But the engineer's role is changing, from the person who solves the problem to the person who designs the feedback. There's a deeper implication too. A layered, verifiable feedback environment built on real infrastructure tasks is exactly what training the next generation of models needs most. Every task the agent completes can become training ground for its successor. We are still far from recursive self-improvement. But the smallest loop now exists. The model optimizes the system. The system serves the model.

@hsu_steve explaining why rapid recursive self-improvement means nothing for AI valuations until real-economy businesses -- not VC-backed startups -- start organically spending trillions on the tech.

@hsu_steve Here is a picture of them with the downed F-15 in their flip flops and robes lol. Absolutely insane the poorest country in planet Earth in PPP terms is doing this much damage.

3+ PAC-3's to intercept each MRBM? 🤔 Grok What is confirmed: Overnight 8–9 Sept, Iran fired ballistic missiles at the U.S.-used Muwaffaq Salti / Al-Azraq base in Jordan. Jordan says 20 incoming, 18 intercepted, 2 fell in empty areas, no casualties. A U.S. official said no significant impact on U.S. bases and all U.S. personnel accounted for. What is estimated, not official: Videos show a very large Patriot salvo. OSINT puts launches at ~60–100+. Type is assessed as mostly U.S. PAC-3 MSE (Jordan’s own Patriots are older PAC-2 and cannot fire PAC-3). Pentagon/CENTCOM have not released count or variant. $4M+/round and “half-billion salvo” are cost extrapolations from those estimates, not a disclosed bill. What is contested: IRGC claims hangars/shelters and fighter facilities were hit, some with cluster munitions. Some videos show explosions near the base. No independent satellite confirmation for this night yet. Do not overclaim: Cannot prove every trail is PAC-3 from phone video. Cannot treat 100 launches or a 90% intercept rate as fact. Shoot-look-shoot / multi-round doctrine means high interceptor use is expected even when most threats are killed.