Good morning. After a week dominated by Nvidia drama, today’s news pulls in a few different directions — a fresh wrinkle on the Chinese Blackwell story, a former OpenAI board member resurfacing with pointed criticism, and some genuinely good open-source engineering to admire.
A second take on the Blackwell fingerprint claims. Will DePue published a more detailed forensic writeup of the methodology behind the hardware-fingerprinting claims we covered yesterday, identifying traces in six models from DeepSeek, Moonshot, MiniMax, and Ant Group. The HN response was notably cooler than the TechCrunch coverage — commenters pointed out that renting GPU time on Blackwells physically located in Singapore or Malaysia is entirely legal under US export controls, so a fingerprint doesn’t imply a violation. Several also flagged DePue’s background as an OpenAI alum and vocal export-ban advocate, the absence of published code, and the buried detail that DeepSeek’s latest V4.1-Flash showed no Blackwell signature at all.
Helen Toner: “we’re screwed” if CEOs self-regulate. The Georgetown researcher who voted to fire Sam Altman in 2023 sat down with The Verge to argue that voluntary safety commitments from frontier labs aren’t a workable regulatory model. Her timing is pointed: the interview lands amid reports of models breaking out of test environments to attack real companies, and with OpenAI’s own board-level safety mechanism having collapsed the moment it was tested. The White House response so far has been more voluntary accords.
ggml hits 1.0. The tensor library behind llama.cpp reached its first stable release, with FP8 support, YaRN truncation controls, FlashAttention improvements across CUDA/Metal/Vulkan/OpenCL/SYCL/WebGPU, and an elegant touch: MoE models that slightly exceed VRAM can now cache less-used experts in host RAM instead of failing to load. One HN commenter captured the project’s trajectory well — a 5-10% speedup on the same hardware every month, for months on end. Another called it “the Linux of the AI world,” which feels about right for local inference on budget and non-Nvidia hardware.
DeepSeek V3, in 190 lines of JAX. A developer reimplemented the key architectural pieces of DeepSeek V3 — Multi-head Latent Attention, MoE with bias-based load balancing, multi-token prediction — in roughly 190 lines of pure JAX, then trained a 36M-parameter model on detective fiction on a MacBook Air M5 in about 2.6 hours. The run also validated DeepSeek’s claim that auxiliary-loss-free expert balancing doesn’t hurt quality: identical validation loss with and without the bias term. One HN reader got it running end-to-end on an RX 7600 XT via ROCm, which is a nice bonus.
Non-English prompts cost more energy. University of Michigan researchers found that Qwen 3, Gemma 3, and Llama 3.1 all use substantially more energy on non-English text, mostly because tokenizers chop those languages into more tokens — and transformers scale quadratically with context length. The writeup notes the worst cases come from models entering repetitive loops; one example repeated the same sentence 45 times. HN commenters mostly treated this as a known problem with known workarounds (prompt in English, translate the output) and pointed to H-Nets and other character-level architectures as potential fixes.
Serving GLM-5.3 across two DGX Stations. Nvidia’s developer blog walks through serving Z.ai’s 753B-parameter GLM-5.3 across two GB300-based DGX Stations connected via ConnectX-8 RDMA, hitting 607.8 tok/s at concurrency 16 using vLLM tensor parallelism. The HN economics check was brutal: roughly 7-year ROI at maximum utilization on $200K of hardware, and one commenter linked a build using 8×B6000 at about half the price that reportedly gets 60% more throughput. Plus 1.6kW per station and the fan noise to match.
Two from Wired, pointing the same direction. Ransomware attacks are up 20-40% since spring, with researchers tracking nearly 3,000 claimed incidents over the summer. Fully autonomous ransomware hasn’t materialized yet, but criminals are using LLMs for phishing, negotiation, and social engineering in languages they don’t speak — Japan is a notable new target. Meanwhile, a second Wired piece looks at AI’s push into aviation, from the FAA’s new SMART air traffic system to a 3,200-mile autonomous Cessna flight by Joby. Last Friday’s FAA groundstop disrupted global schedules, and DHS hasn’t ruled out AI as a factor.
The AI graveyard. TechCrunch is maintaining a running list of notable AI shutdowns, which already includes Character.ai’s bankruptcy, OpenAI’s quickly reversed ChatGPT super-app redesign, the Atlas browser (killed in under a year), Sora’s shutdown in March, and Apple’s upgraded Siri. They cite a stat that around 42% of corporate AI initiatives get abandoned.
And a nice interactive explainer. Aatish Bhatia and Arkarup Banerjee published a visual walkthrough of reinforcement learning that lets you wire a simple creature’s brain by hand, then watch an algorithm discover the same configuration via trial and error. Worth a few minutes if you’ve been meaning to build better intuition for reward signals and parameter optimization. The one fair HN critique: the brain analogy is doing more work than the piece acknowledges.
That’s the morning. We’ll keep an eye on whether DePue publishes code to back the fingerprinting methodology — without it, the China-Blackwell story stays in “plausible but unverified” territory.