Live Feed
vLLM Brings Five Speculative-Drafting Methods to AMD MI300X and MI355X as Inference Efficiency Race Goes Cross-Platform
vLLM's August 23 blog details native MTP, EAGLE-3, DFlash, DSpark, and Gemma 4 MTP now working on AMD Instinct hardware via ROCm — arriving as NVIDIA separately claims 1.9x throughput gains for RTX 24GB+ via kernel-level optimizations.
Broadcom Projects $230 Billion in AI Chip Revenue by FY2028 as Custom Silicon Supercycle Accelerates
CEO Hock Tan laid out a three-year trajectory on September 2: $58B this fiscal year, $115B in FY2027, $230B in FY2028 — all from custom ASICs and AI networking chips, a direct challenge to NVIDIA's GPU dominance.
Experts Are Being Paid to Train the AI Systems That Could Replace Them
A South African academic with a PhD in AI was offered $37/hour to train an AI to perform her job. In India, workers strap cameras to their heads for $2.60/hour to train humanoid robots. The economics of knowledge transfer are becoming a defining feature of the AI labor market.
Charter CFO Jessica Fischer Joins Blackstone-Google AI Cloud Venture
Four months after the joint venture's $5B launch, Charter Communications CFO Jessica Fischer is leaving to run its finances. The hire signals the Blackstone-Google AI cloud company is moving from announcement into operational buildout.
Any Embedding Vector Can Be Decoded Without the Source Document
Research demonstrates that text embeddings can be translated between any two model vector spaces without paired data, letting adversaries reconstruct sensitive document content from embedding vectors alone. As RAG pipelines become standard infrastructure for production agent systems, the attack surface has grown considerably since the work was first published.
Open for 52 Years, the Spherical Hadwiger Conjecture Falls to AI-Assisted Proof
A preprint by Wang Suijie et al. of Hunan University proves the Spherical Hadwiger Conjecture, open since 1974, with OpenAI Codex developing proof details and identifying gaps. It lands one week after Claude Fable formalized Fermat's Last Theorem in Lean.
OpenAI Built a GPT-5.4 Thinking Monitor to Watch Its Own Coding Agents for Misalignment
OpenAI published details of a low-latency internal monitoring system that uses GPT-5.4 Thinking at maximum reasoning effort to watch coding agents running across 54,000-plus internal Codex sessions. GPT-6 Astra produced 50% fewer higher-severity misalignment flags than its predecessor.
OpenAI AI Systems Now Do 3.1x the Research Work of Its Human Staff
OpenAI has hit the automated research intern goal it set last fall, with AI systems performing 3.1x the research effort of human staff as of September 2026. The company is now targeting a full automated AI researcher by March 2028.
Qwen3.8 Max 0902: Alibaba Post-Trains 2.4T MoE for Agentic Coding at $2 Input
Alibaba's September snapshot of Qwen3.8 Max adds specific post-training for multi-step software projects, multi-tool orchestration, and long-horizon task execution. At $2/$6 input/output per million tokens, it sits between flash-tier pricing and frontier-model costs.
GPT-6 Astra's First 48 Hours: Playable 4D Chess, 334-Part Tesla Model, WebGPU Physics
Within two days of GPT-6 Astra going live, developers published a playable 4D chess game, an interactive Tesla Model X decomposed into 334 individually modeled parts, a WebGPU jelly simulation, and a procedural 3D world — an early capability floor check.
AI Video API Pricing Falls Below $0.10/s as Wan3.0 and Grok Imagine Video 1.5 Both Enter Arena This Week
Two AI video APIs entered Arena.ai's Text-to-Video leaderboard in the same week: Alibaba's Wan3.0 at $0.05 per second and xAI's Grok Imagine Video 1.5 at $0.08 per second. The pricing and capability differences define where the video generation market is splitting.
Wan3.0 Enters Text-to-Video Arena With Document Input and 30-Second Generation at $0.05/s
Alibaba's Wan3.0 joins Arena.ai's Text-to-Video leaderboard as the first video model to accept PDFs, spreadsheets, and presentations as structured generation inputs, tripling Wan2.7's duration ceiling to 30 seconds.
OKF Agent Memory Turns Your Git Repo Into a Structured LLM Wiki — 300µs BM25, No Database Required
A single zero-dependency Go binary uses Google's Open Knowledge Format and in-memory BM25 search to give AI coding agents persistent structured memory inside the project Git repo. It ships an MCP server for Claude Code and Cursor, and cuts prompt overhead by up to 90%.
Harvard-Santa Fe Paper Models LLM Adoption as a Cognitive Epidemic — Tipping Points and Lock-In After a Critical Threshold
A preprint applies epidemiology and evolutionary biology to LLM diffusion, finding that populations can cross irreversible tipping points into persistent cognitive dependence. The authors propose containment strategies they call cognitive immunization.
Claude Fable 5.1 Batch API Cuts Inference to $5/$25 Per Million Tokens
Anthropic adds a batch tier to its flagship reasoning model at half the real-time price. With a 98% cache discount intact, async workloads running large cached contexts can drop below $0.10 per million effective input tokens.