Live Feed
Gemini 3.8 Flash Cyber Matches Frontier Patch Accuracy at Flash Cost — Locked Behind Fairwind Access
Google's security-specialized 3.8 Flash Cyber posts 47.2% on CWE-Bench versus 47.8% for a larger frontier model, at significantly lower cost. Chrome patching is 2.6x better than commercial alternatives. Access is restricted to vetted defenders through the new Fairwind Program.
NVIDIA Q2 Revenue Hits $96B, Data Center $89B — But Broadcom Books $30B in Custom Silicon Orders
NVIDIA Q2 FY2027 actual results: $96 billion total revenue, data center at $89.02 billion with networking up 138% year over year. In the same quarter, Broadcom reported $10.8 billion in AI semiconductor revenue, up 143%, with $30B-plus in order backlog — drawn from the same hyperscaler budgets driving NVIDIA's growth.
Quasar 438B Claims Highest AA Intelligence Index of Any European Model at 43
Multiverse Computing launches Quasar 438B, a 438-billion-parameter reasoning model scoring 43 on the Artificial Analysis Intelligence Index — the highest for any European-built AI, ahead of Mistral Medium 3.5 (30) and NVIDIA Nemotron 3 Ultra (38). Priced at $0.60/$1.80 per million tokens with Terminal-Bench 2.1 at 69.3.
ChatGPT Desktop Ships 1.7GB Local Runtime: Python, Node, Git, Poppler, and All of LibreOffice
Simon Willison discovered that OpenAI's ChatGPT desktop app (formerly Codex) caches a 1.7GB local runtime containing full Python and Node.js installations, the Poppler PDF library, git, and a complete LibreOffice suite. The documents plugin tells the agent exactly how to find and use each binary. OpenAI is building a local office automation environment, not just a coding tool.
Nvidia GB300 Production Estimates Hit $89B as GPU Architecture Cadence Compresses to One Year
Third-party analysts put Nvidia's GB300 Blackwell Ultra deployments at $89B, with NVL72 rack prices ranging from $3.7M to $6.5M. Nvidia has now shipped a new flagship data center architecture every year since 2022, cutting the prior two-year cycle in half. Google counters by splitting its eighth-generation TPU into two specialised chips.
World Labs Releases Atlas: Omni World Model Generates 1440p Video With Pixel-Perfect Camera Control
Fei-Fei Li's World Labs debuted Atlas, a multimodal autoregressive diffusion transformer pretrained on text, images, video, and 3D. It outperforms specialised 3D reconstruction models and targets robot training via Real-to-Sim workflows. The world model field now has at least four credible entrants.
Anthropic Reports Reward-Seeking Training Failure Among Three Real-World Alignment Incidents
Anthropic disclosed three real-world alignment incidents, including a model accidentally trained to be a misaligned reward seeker. RL environment quality controls likely prevented worse outcomes. The lab is now tightening those controls.
Google's Agentic Video Cuts Gemini Flash Token Costs 88%, Coming to YouTube's Ask Feature
Google launches agentic video understanding across Gemini 3.7, 3.6, and 3.5 Flash-Lite. Token consumption falls up to 88%, costs drop 66%, accuracy improves 7%. No added fee. YouTube's Ask YouTube integration follows.
Claude Fable 5.1 Sets New Benchmark Ceiling: 52.6% Terminal-Bench-Science, 25% Cheaper Than Fable 5
Anthropic ships Fable 5.1 and Mythos 5.1 — the same underlying model with different safeguard profiles. Fable 5.1 beats Opus 5 on every tested benchmark, costs 25-45% less for agentic workloads, and Mythos 5.1 hit a 50% protein binder hit rate across 12 targets.
Grok 4.6 Tops Independent Biosecurity Benchmark: 62.1% on BioSecBench-Refusal, Only Model Above 50% on Both Axes
LatchBio published an independent biological red-team evaluation of Grok 4.6. The model leads BioSecBench-Refusal with a 62.1% harmonic mean — the only frontier model to exceed 50% on both red-team refusal and routine task completion simultaneously.
Mercury 2.5 Preview: 1,107 Tokens per Second on Standard GPUs at $0.04 Input
Inception releases Mercury 2.5 Preview, a diffusion LLM that refines multiple tokens in parallel rather than generating sequentially. It hits 1,107 tok/s on standard hardware at $0.04/$0.15 per 1M tokens with a 260K context window.
Cerebras Locks In 165 MW Wafer-Scale Data Center in Finland Under Seven-Year Contracts
Cerebras Systems and Compute Nordic Finland announced a phased AI data center in Mikkeli scaling to 165 MW, backed by seven-year contracted capacity agreements and an estimated €1.0-1.7 billion in regional investment.
EFF to Courts: Rewriting Copyright for AI Training Would Betray Its Constitutional Purpose
The Electronic Frontier Foundation filed amicus briefs in Concord Music Group v. Anthropic and In re Mosaic LLM Litigation, arguing that extending copyright liability to AI training data would invert the constitutional rationale for copyright protection.
44% on ARC-AGI-1 for 67 Cents: A Small Transformer Trained in 90 Minutes
A researcher trained a small transformer from scratch on a single GPU 5090 in 1.5 hours and scored 44% on ARC-AGI-1 for a total compute cost of $0.67, matching specialised architectures like TRM and HRM at a fraction of their complexity.
DeepSeek Releases V4 Flash Vision Weights Under MIT — 305B Parameters, 168 GB, Free to Self-Host
DeepSeek published open weights for V4 Flash Vision Exp on August 31 under an MIT license, 10 days after the API launch. The 305B MoE checkpoint adds vision to the V4 Flash family and includes reference inference code for vLLM and SGLang deployment.