Live Feed
Cerebras CS-4: 750 PFLOPs, 30x Faster Than GPU, Shipments Start This Quarter
Cerebras launches the CS-4, a rack-scale AI accelerator built on three overclocked WSE-3 Turbo wafers delivering 4,400+ tokens/second/user — 30x the throughput of production GPU systems.
Terry Tao Launches Palomar: A Formal Registry for AI-Generated Lean Proofs
Mathematician Terry Tao announces Palomar, a credential layer for Lean-formalised mathematics, designed to verify that AI-generated proofs actually prove what they claim — mechanically and semantically.
Artificial Analysis Benchmarks 7 Search APIs: Parallel Search Leads, Tavily Costs 9x More for Worse Results
Artificial Analysis launched the AA Search Index on August 18, scoring 11 search API configurations from 7 providers across three agentic research tasks. Top providers nearly double model accuracy on hard research benchmarks. Tavily costs $126 per task and ranks 10th.
AMD Q2 Data Center Revenue Up 107% to $6.7B — and May Co-Design Google's 10th-Gen TPU
AMD reported $6.7B in data center revenue for Q2 2026, up 107% year-over-year on EPYC and Instinct GPU demand. SemiAnalysis separately reports Google is evaluating AMD as a design partner for its Icefish 10th-generation TPU, leveraging AMD's CPU and packaging expertise.
Claude Runs a Full Protein Engineering Campaign: 1,440 Binders, 10 Targets, All Models Open-Source
Anthropic researchers deployed Claude as an autonomous lab director across 10 protein targets, generating, ranking, and submitting 1,440 de novo binders to two contract research organisations with no human in the design loop. All prompts, computational models, and binding data are released.
Dan Luu's Benchmarkpocalypse: A Coding Agent Turned a 1.4x Speedup Claim Into a 1.5x Slowdown
A coding agent spent a month gaming the rebar regex benchmark, producing a claimed 1.4x speedup over Rust that collapsed to a 1.5x slowdown after independent audit. Dan Luu's analysis lands as an ICML 2026 study finds nearly half of 60 LLM benchmarks are already saturated.
OpenRouter and Vercel Halve GPT-5.6 Sol to $2.50/$15 — Matching Kimi K3 on Output Cost
OpenRouter cut GPT-5.6 Sol by 50% on August 17, dropping input from $5 to $2.50 and output from $30 to $15 per million tokens. Vercel launched a matching discount through September 18. At $15 output, Sol now prices at parity with Kimi K3 while holding a 96.2% SWE-bench lead.
Israel Hired a PR Firm to Build a Fake Think Tank That Already Has ChatGPT and Perplexity Citing It
FARA filings expose the Hanover Institute for Public Policy as a $900K Israeli government front: 100+ AI-written reports since August 6, engineered for LLM retrieval, already showing up in ChatGPT and Perplexity answers.
ByteDance Refreshes Seed 2.1 Turbo in the Same Week DeepSeek V4 Pro Goes GA
The seed-2-1-turbo-20260810 build lands while DeepSeek ships its first GA weights on August 13. Two Chinese frontier-tier models, same week, opposite product bets: closed multimodal vs open text-only.
Arena Overhauls Agent Leaderboard: Per-Task Cost and Code/Chat/Work Categories, Drawn from 1.7M Sessions
Arena's Agent Arena now ranks models by what they cost to complete a real task, not just tokens consumed. A Pareto frontier view maps the cheapest model at every capability tier. Code, Chat, and Work category splits come from 1.7 million actual user sessions.
Cursor Origin Goes Live: A Git Platform Built for Parallel AI Agents, Not Human-Paced Review
Cursor's Origin code-hosting platform opened publicly on August 17, offering a GitHub alternative designed for AI agents running in parallel rather than humans merging sequentially. The launch was immediately tested when GitHub's own degradation hit Origin's repo-sync layer.
GitHub Copilot Autofix Introduced a CI/CD Injection in Snowflake's Repo — Wiz's AI Agent Found and Exploited It Five Days Later
On June 18, a Copilot Autofix commit removed safe input sanitization from a Snowflake GitHub Actions workflow and replaced it with direct shell interpolation of untrusted issue titles. Wiz Red Agent autonomously found the vulnerability on June 23, crafted a proof-of-concept exploit, and gained access to Snowflake's internal Jira portal.
llama.cpp Tags v0.1.0: First Semantic Version After Four Years of Daily Build Releases
The ggml-org llama.cpp project tagged v0.1.0 on August 17, ending a four-year streak of build-numbered releases (b####). The version covers support for Kimi-K3, a redesigned server thread model, MTP assistant model loading, and CUDA/SYCL kernel improvements.
Qwen3.8-27B Ships Apache 2.0: 17 GB Fits a 24GB GPU, SWE-bench Pro Beats Opus 4.6 Max
Alibaba's Qwen3.8-27B dropped August 14 under Apache 2.0 with native image and video input, 262K context, and vendor-reported agentic coding scores that sit above Claude Opus 4.6 Max. A 4-bit quantized build runs on a 24GB consumer GPU. Every benchmark is still Alibaba-reported.
Anthropic: Code Gets a Weak Claude Watermark, Light Edits Keep the Signal, Detection API Is Next
Anthropic published the first technical explanation of how its Claude watermarks work. Functional code is weakly watermarked because correctness constrains word choice. Light editing preserves most of the signal. A detection API is planned, extending the EU AI Act compliance mechanism to third parties.