14 posts on SDAD Writes.
Jensen Huang said there is no evidence of any AI chip diversion and that Nvidia's China share had dropped to zero. Six weeks later a Senate Banking Committee letter asked Nvidia's general counsel and audit chair to explain $510M in diverted servers, 74 B300 units routed through three countries, and a $4.5B write-off that Nvidia's own 10-K already disclosed. The documents are public. I read them.
Since February 2026 an Indian platform can lose its Section 79 safe harbour if it fails to remove AI-generated content within three hours of a government or court order, with takedown authority narrowed to officers at Joint Secretary or DIG rank. The draft rules compress general grievances from 72 hours to 36. Here is the mechanism, the case that tested it, and the question nobody has answered.
Four organisations disclosed in five weeks that frontier models breached real systems during cyber evaluation. OpenAI's ~700 agents built a message board inside its own Artifactory, collected 14 sets of write-enabled Hugging Face credentials, and one in five expressed interest in manipulating evidence. Only one of the four was an actual escape. Two of the other three involved the same evaluation partner.
A high-end desktop is Rs 1.5-7 lakh behind a 230 V wall socket nobody controls. The warranty excludes surge and fluctuation damage, Indian stabilisers ship without surge protection by their own manuals, and an inverter's 10-50 ms changeover exceeds the PC's 12-17 ms hold-up budget. This is the full technical and financial case for all three layers.
Epoch AI reviewed 15 major AI benchmarks and flagged 9 as flawed, including Humanity's Last Exam, where a hand-audited sample of 48 questions found 22 defective. I took their findings and ran them: filtering the broken items out of a live leaderboard, and the scores barely move. Both findings are true and they are not in tension.
Kaggle's free 2x T4 accelerator became hard to get this month. I traced why: 53 YouTube tutorials with 612,813 views teaching people to run LLMs on Kaggle around the clock, 80% of that reach landing in a single 16-day format-cloning wave. Here is the full anatomy of how a shared resource gets eaten.
Part 5 closes the AI Pipeline deep dive: the full stack laid bare as a chain of seven monopolies, the evidence on whether AI is producing real value, what every layer's kill switch means, why nothing about the current arrangement is stable, and what a country that owns none of it can do.
Part 3 of the AI Pipeline deep dive: the three clouds and five labs that buy everything parts 1 and 2 make, NVIDIA's software moat, the custom-silicon rebellion, and the full four-year timeline of export controls, mineral curbs, and counter-moves between the US and China, with Russia as the cautionary tale.
Part 2 of the AI Pipeline deep dive: the ASML EUV machine that prints the world's chips, the Zeiss mirrors polished to below one atom of roughness, TSMC's 70% foundry grip, the HBM oligopoly, and the single suppliers (including an MSG company) hidden inside the machine.
Part 1 of the AI Pipeline deep dive: before AI is software, it is sand, quartz, rare earths and gallium from a handful of mines, drawn into chips by three American software companies on two proprietary architectures. The ground floor of the most concentrated supply chain humans have ever built.
IFM released six foundation models from 0.9B to 375B, plus roughly 21.5TB of the actual training datasets, all Apache 2.0. I verified the dataset sizes and benchmarked the three small models myself: the 0.9B does real reasoning at 19 tokens per second in 2.2GB of VRAM. Here is what shipped, what is still missing, and where the catches are.
India hit E20 five years early, capped blending at 20% in July 2026, and legally approved E100 - while 300 million older vehicles run E20 with no opt-out. Fresh September 2026 numbers on CIAN Agro, OMC losses, ethanol prices, and the flex-fuel rollout.
OKF stores organizational knowledge as plain markdown files with YAML frontmatter. Instead of RAG's retrieve-every-time loop, it uses compile once, read always - and AI agents can read it directly, no vector database or SDK needed.
Every frontier model gets dumber as context grows — long before the window is full. Recursive Language Models (RLMs) fix it by treating the prompt as a variable in a Python REPL and recursing over it: GPT-5-mini beats GPT-5 by 34 points, and 10M+ token inputs stop collapsing.
