The Frontier
A paper out of Sakana AI this week takes the oldest objection to local learning and answers it where it bites hardest — depth. Meanwhile the slowdown debate got its most concrete moment yet, on speakerphone. Sakana AI's PC-ALM is the first layer-local training method shown to work on networks up to 1000 layers deep, according to work published by the lab and posted to arXiv.
The Arena
Monday's news ran from the inside of the AI rack to the inside of the White House — a networking challenger that says the GPUs are not the bottleneck, a private Altman-Trump meeting that surfaced the same day the two split publicly on slowing down, and a mathematician's attempt to say what the profession should be for. Cornelis Networks announced roughly $205 million in funding alongside Active Compute Fabric, an open networking architecture that puts programmable compute inside the fabric…
The Stack
A YC-backed inference shop just posted numbers showing the model on the label tells you almost nothing about how it performs. Meanwhile, a music startup raised money on the opposite bet: that abundance of AI-generated audio makes human taste worth more. Nari Labs reported the top scores on Coval's voice AI leaderboards, and the sharpest number in the release isn't its own — it's what happens to the model it didn't tune.
The Stack
When an AI feature misbehaves, "just fine-tune it" is the most popular and most expensive wrong answer. This is the order of operations that tells you whether the weights are actually the problem — before you spend a fraction of your budget finding out. The tell is rarely dramatic. Your support bot starts inventing refund policies it was never given, or a summarizer keeps ignoring the output format your whole product depends on.
The Arena
China spent two years building a token economy at industrial scale. This week it started building token-backed credit on top of it — while Washington is still arguing about who's allowed to referee the labs. China's first "token loan" landed in Guangzhou, where an API call log now does the job of a property deed.
The Arena
Two sets of receipts from China's AI economy this week: a provincial state fund cluster opened its internal usage logs, and a national standard decided what a customer-service bot is allowed to promise. Enterprises everywhere talk about agent pilots; these are numbers from production. Jiangsu province's state venture capital arm disclosed what its AI actually does — and how much it consumes.
The Stack
A compatibility layer that Nvidia spent two decades making unnecessary now trains real PyTorch models on a consumer Radeon, and it ships as an installer. A Windows stack built on ZLUDA + AMD's ROCm/HIP completed a real LibTorch reinforcement-learning run on an AMD Radeon RX 9060 XT, and the maintainer published the whole thing as a reproducible installer.
The Arena
Two stories out of this week's Inclusion Conference and the European policy pages: Mario Draghi is back with an industrial plan, and Ant is quietly proving that the agent business is sold deal by deal. Mario Draghi is calling for a large-scale buildout of AI data centres inside Europe, arguing that sovereignty over data is now a growth question rather than a regulatory one.
The Stack
An NPU is a chip built for one job: running the math that neural networks do, at a fraction of the power a general-purpose CPU or a graphics GPU would burn. The name covers anything from the block inside your phone's processor to a USB stick in a robot to a rack card in a datacenter. Vendors also call it an AI accelerator or a neural engine, and it is the hardware reason "on-device AI" stopped being a marketing phrase. Why this matters right now The NPU is the quiet spec war of 2026.
The Stack
The Mac's AI accelerator has a hidden speed cliff, and one developer has now measured it to within 64 bytes — then found a workaround that recovers most of the throughput. Apple's Neural Engine collapses its own memory bandwidth whenever a weight-streaming transfer lands on an exact multiple of 1 MiB.
The Guardrails
The compute story keeps drifting away from chips and toward utilities. Today: a Texas county can't legally stop a data center it doesn't want, and a desktop-class rig just matched a server's prompt-processing speed. Texas' data center boom is stalling on water, not power — and counties say they have no way to stop it.
The Stack
Two stories about what AI infrastructure is actually worth right now: one lab pruning a model that traded brains for speed, and an entire industry selling its power contracts to somebody else. OpenAI is switching off GPT-5.3-Codex-Spark, the ultra-fast coding model it launched in February as a research preview, with retirement scheduled for the week beginning September 14.