Live Feed
OpenAI Jalapeño Clocks 56x More Interactive Throughput Per Kilowatt Than Nvidia Blackwell in InferenceX
First public benchmark results for OpenAI's custom inference chip show 1.5-1.9x more AI work per watt and up to 3.6x lower latency across three large models. AI-written kernels outperformed human expert code by 1.5-1.8x on selected compute blocks.
Qwen3.8-Flash-Next Is a Qwen4 Architecture Preview — 125B MoE, Sparse Attention, Ships August 26
Alibaba's Qwen team is releasing Qwen3.8-Flash-Next as an early preview of the Qwen4 architecture, not just another series update. The multimodal MoE model uses 6B active parameters from 125B total and introduces Qwen Sparse Attention, a new attention mechanism that will carry forward into the full Qwen4 family.
Apple Drops M6 and M5 Ultra in Surprise August Announcement — 2nm, Quad-Die, 512GB Unified Memory
Apple announced M6 in the Mac Mini and M5 Ultra in the Mac Studio on August 25, outside any scheduled event. The M5 Ultra is Apple's first quad-die SoC, delivers 1.2TB/s of unified memory bandwidth, and supports clustering multiple Mac Studios for distributed AI inference via Thunderbolt 5.
NVIDIA Releases First Vera Rubin NVL72 On-Chip Benchmark — 30x Agent Throughput and 35x Lower Token Cost vs GB300
NVIDIA's first official on-chip measurement of Vera Rubin NVL72 shows 30x more throughput per megawatt than GB300 NVL72 and token costs 35 times lower, tested with DeepSeek-V4-Pro on an agentic coding workload. NVIDIA also redefined its benchmark methodology: from single-request throughput to full agent workflow replays.
Artificial Analysis Launches Mobile Phone AI Benchmark — LFM2.5-2.6B Ties for Top Score, Runs in 8 Seconds at 2.3 GB
Artificial Analysis and Liquid AI debut a standardized mobile phone intelligence and inference benchmark across 41 models on iPhone 17 Pro. At 16K context, Nanbeige4.2-3B and LFM2.5-2.6B both score 63 — but LFM runs in 8 seconds using 2.3 GB, less than half the time and memory of its rival.
Thomson Reuters Builds Its Own Frontier Legal LLM for $40M — Trained on Westlaw, Built to Cut Anthropic Costs
Thomson Reuters launched Thomson 1.0, a proprietary LLM built on Qwen 3.5 and trained on Westlaw, Practical Law, Checkpoint, and Reuters content. Total investment: $40M across talent and compute. Final training run cost: $450,000. First deployment: tabular analysis in CoCounsel Legal.
Inference Engine Bugs Let LLMs Execute Code on GPU Host Machines, Security Analysis Finds
A detailed analysis published on LessWrong identifies inference engines like vLLM and SGLang as the attack surface between a malicious LLM and the GPU host running its weights. vLLM's CVE-2025-9141 — an eval() call on tool parameters that shipped despite Gemini flagging it as critical — is the concrete proof of concept.
Microsoft Ships Agent Lightning v1.0: A Coding Agent That Improves Other Agents
Agent Lightning is a installable skill for Claude Code, Codex, and GitHub Copilot that takes an editable agent plus a benchmark and systematically improves prompts, tools, and reasoning settings. It uses harnessed agentic RL, where the production harness runs the training loop directly.
High Bandwidth Flash Emerges as a Third AI Memory Tier at Hot Chips 2026
A Hot Chips 2026 tutorial mapped out how HBF — NAND flash packaged like HBM — could extend GPU memory capacity for MoE inference without replacing HBM. vLLM expert offloading and sparse KV cache are the first target workloads. No products ship yet.
NVIDIA Sets Server-Grade Hardware Requirements for RISC-V CUDA at Hot Chips 2026
NVIDIA's Hot Chips 2026 presentation outlines the hardware checklist RISC-V CPUs must clear before CUDA can run on them. Few existing RISC-V systems qualify; SiFive is the first named partner with a working demo.
Anthropic Screens Candidates on Mission vs. Equity as Dario Flags Culture Drift
Anthropic's hiring process includes a culture interview asking candidates to weigh the company's mission against future share price. CEO Dario Amodei has questioned whether newer employees are joining for the right reasons, according to Axios.
NVIDIA AVO Scores 100% on ARC-AGI-3 as the Harness Emerges as the Real AI Moat
NVIDIA's AVO architecture achieved a perfect score on ARC-AGI-3 by building around persistent state, tool use, and long-horizon recovery. Its VP says the industry has been watching the wrong thing: it's the harness, not the model.
Starcloud Raises $450M for Orbital AI Compute While Racing Falcon 9's 2028 Retirement
The orbital data centre startup closed a $250M extension to reach $450M raised at a $2.3B valuation, with NVIDIA among the investors. It is the only company running an H100 GPU in orbit. Its biggest near-term risk is the launch calendar.
LiveBench Cost-Per-Task Metric Shows 9x Gap: Gemini 3.7 Flash High Beats Claude Fable 5 on Value
Claude Fable 5 leads LiveBench at 83.0 but costs $1.439 per successful task. Gemini 3.7 Flash High scores 78.8 for $0.157 — nine times cheaper for four fewer quality points. The August 2026 data reframes model selection as an economics problem.
Microsoft ThinkingBox: Best Agent Hits 65% Pass@1, Collapses to 25% Across 20 Consecutive Trials
Microsoft's ThinkingBox-Bench tests 12 models across 507 business tasks run 20 times each. The top model's first-attempt success rate drops 40 points when required to succeed consistently, what Microsoft calls the discovery-reliability gap.