Still
Crunchy.

AI models get cheaper, smarter and more mathematical

OpenAI published 722 manuscripts covering 372 result families, including solutions to what AGMAI, an advisory group of elite mathematicians formed to manage the release responsibly, describes as hundreds of open problems. The underlying frontier model has not been released, and the batch has both impressed and unsettled parts of the mathematical community while raising questions about research ethics.

NVIDIA offered a counterpoint: a recipe of supervised fine-tuning, reinforcement learning and test-time refinement took its Nemotron family to gold-medal level at both IMO and IOI 2026, scoring 30 out of 42 on officially graded IMO proofs and 535.4 out of 600 at IOI, above the top human score. It released the models, data and recipes on Hugging Face. Anthropic, meanwhile, launched Claude Haiku 5.5 with prices cut up to 90 percent for typical prompts and adjustable reasoning levels, though a new tokenizer that consumes more tokens per task will trim real-world savings. The Decoder reads the move as a reaction to OpenAI's GPT-6.1 series, evidence the price war is intensifying.

Agents move into the operating system, and the hardware follows

Microsoft's Windows and Surface event put agents at the centre of the platform. An upgraded Copilot can reach into local files and take actions across Windows under a framing Microsoft calls Hybrid Intelligence, mixing local and cloud models. Microsoft Execution Containers, now generally available, provide OS-level plumbing so agents run safely and persistently under Windows' control. NVIDIA is co-engineering agent hardware with Microsoft, and RTX Spark laptops open preorders with availability on 16 October.

AMD answered with its own Agentic PC pitch, centred on the Ryzen AI Max+ PRO 495 with up to 192 GB of unified memory, enough to run a 320-billion-parameter model locally at 20 tokens per second. AMD leaned on x86 compatibility, which runs Windows and Linux natively where Arm machines rely on Microsoft's Prism translation layer, and claimed wins over Intel's Core Ultra X9 388H in 17 of 18 generative workloads, though all figures come from AMD's own testing. The Surface Laptop Ultra itself arrives 16 October from $2,599, with magnetic charging revived over a standard USB-C port that still handles data and video.

Security scrambles to keep up with machine-written code and critical systems

GitHub reports that AI agents now appear in one in three pull requests, up from fewer than one in ten a year ago, and its data across nine quarters suggests developers are not getting careless, just outpaced. A new secret appears in public code roughly every two seconds and manual revocation takes a mean of around 40 days, so GitHub built a classifier with Microsoft Applied Sciences that assesses candidate secrets in under two milliseconds, catching unstructured credentials that pattern matching misses.

On critical infrastructure, a coalition of cybersecurity experts is pressing CISA to issue a binding directive requiring federal agencies to meet baseline standards for operational technology, the systems that run buildings, ports and utilities. Recent attacks hit hundreds of water systems across at least 12 states, and a GAO study found only 7 of 22 civilian agencies had fully inventoried their networked devices. The coalition's key argument is ownership: when no one is accountable, no one secures it.

Supply chains, watermarks and revivals fill out the day

Samsung's HBM4E memory has reportedly passed testing for NVIDIA's Rubin Ultra accelerators, a turnaround after its HBM3E miss left it supplying only GDDR7, with production expected to double to 50,000 wafers per month in 2027. Bloomberg reports Seagate and Toshiba are vying for TDK's hard drive head business, the industry's only independent supplier, which Toshiba depends on entirely for its HAMR roadmap. Google opened SynthID.com so anyone can check for AI watermarks, now covering partner watermarks beyond Google's own, across more than 180 billion Gemini images and videos. In gaming, AMD confirmed FSR 4 upscaling reaches APUs and handhelds by year's end, Arc Raiders will permanently ban first-time cheaters, and a reported Ape Escape revival at Sonic Mania studio Evening Star fits Sony's pattern of licensing dormant franchises to smaller studios.