
Microsoft's MAI-Transcribe-2 tops FLEURS across 60 languages at 5.2% word-error rate
Microsoft's in-house MAI-Transcribe-2 ranks first on the FLEURS benchmark across 60 languages with a 5.2% average word-error rate, and Artificial Analysis clocks it 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, and 5x faster than Gemini 3.5 Transcribe. At $0.10 per hour through year-end via Microsoft Foundry, MAI Playground, and OpenRouter, multilingual transcription with diarization and word-level timestamps collapses to a single model and a price floor competitors now have to answer.
microsoft.ai →- 02
Cloudflare ranks code vulnerabilities by live traffic using OpenAI's GPT-5.6 CyberCloudflare's Managed Defense now pairs OpenAI's GPT-5.6 Cyber with its own network data — live routes, traffic volume, recent attack activity, existing WAF rules — to rank code findings by real production exposure. Prioritization is the product: each finding ships with a proposed patch and a scoped WAF rule, though access is invitation-only and capped at one authorized application per engagement.
blog.cloudflare.com → - 03
OpenAI commits $1B in AI credits to frontline cyber defendersOpenAI is putting $1 billion in subsidized credits behind Daybreak for Frontline Defenders, expected to be consumed within six months, with Daybreak Blue granting restricted GPT-5.6 Sol access for defensive work and Daybreak Red granting GPT-5.6 Cyber to approved organizations for authorized offensive security. The subsidy targets the underfunded end of the defender market — community banks, water systems, nonprofits, open-source maintainers, and an MS-ISAC pilot for state and local teams — making frontier security models a procurement question rather than a budget one for organizations that could never price them in.
theregister.com → - 04
NVIDIA to acquire Hugging Face for $12.9BNVIDIA agreed to buy Hugging Face — home to 18 million developers, 3 million models, and 200,000 companies — for $12,930,300,000, with Jensen Huang stating NVIDIA compute will not be required to build on or deploy through the platform. The neutral hub for open weights and multi-accelerator deployment now sits inside the dominant accelerator vendor, and the no-lock-in pledge is the term the rest of the ecosystem will hold NVIDIA to.
blogs.nvidia.com → - 05
Claude wrote the first computer-checked proof of Fermat's Last Theorem in 11 daysAn internal Anthropic research model roughly comparable to Claude Fable 5.1 burned about six billion output tokens over 11 days to produce 13 million lines of Lean — over 5x the size of Mathlib — proving 30,300 theorems, 29,500 of which appear in the final proof of Fermat's Last Theorem. Lean verified it using only its three standard axioms, and Kevin Buzzard of Imperial College London, who kicked off the community formalization in 2024, called the artefacts robust enough to build on. The novelty is verification speed, not new mathematics: a formalization the field expected to take years closed in under two weeks, putting automatic formalization of the modern mathematical literature within reach.
anthropic.com → - 06
DeepSeek plans 160,000 Huawei Ascend chips in Inner Mongolia, the largest known clusterDeepSeek will deploy at least 160,000 of Huawei's Ascend 950DT chips at an Inner Mongolia data center for inference only, per Bloomberg, while Nvidia hardware still handles training. The split marks the real limit of China's Nvidia substitution — and Huawei is unlikely to fill the order for over a year, with domestic memory maker CXMT starting small HBM3E batches but three to five years behind Samsung, SK Hynix, and Micron.
the-decoder.com → - 07
Google connects Gemini Spark to Google Photos for album and edit tasksGoogle's personal agent Gemini Spark can now run Google Photos tasks — editing images, curating albums, turning a concert flyer into a calendar appointment — rolling out over the coming weeks to Gemini AI Pro and Ultra subscribers in the US in English only. The capability is incremental rather than new: building an album was never hard, and the US-English-only launch signals how much pressure there is to attach AI to every consumer surface while the industry struggles to explain why consumers want it.
techcrunch.com → - 08
Microsoft's Project Zenith ships preconfigured Windows PCs for local 30B+ model inferenceMicrosoft announced Project Zenith, a preconfigured Windows experience for machines with 64GB+ unified memory and 250+ GB/s bandwidth, arriving first on AMD's Ryzen AI Halo — the hardware bar, not the pinned VS Code taskbar, is the substantive part. A developer on qualifying hardware can run 30B-parameter coding models locally and unmetered, shifting routine agentic workloads off metered cloud tokens while frontier models handle the hard problems.
blogs.windows.com → - 09
NHTSA opens probe into Tesla's Cybercab hours after its Austin launchNHTSA opened an audit query into Tesla's decision to self-certify the Cybercab — no steering wheel, no pedals — as FMVSS-compliant, hours after the first production units carried passengers in Austin. Federal rules still require manual controls, and the Trump administration's proposal to drop them is not yet in force, so Tesla's fleet is running on a certification the regulator is actively contesting. The precedent is Zoox: a 2022 self-certification triggered the same audit process, and the company only reached commercial rides after a Part 555 exemption approved in July 2026.
techcrunch.com →