Live Feed
DeepSeek V4 Pro Scores 96.4% on SWE-Bench but Only 77.4 on LiveBench — the Specialist Model Problem
DeepSeek V4 Pro 0813 has the second-best SWE-bench Verified score in existence, but ranks behind Kimi K3, Grok 4.6, Qwen 3.8 Max, and several others on LiveBench's broader seven-category evaluation. The gap illustrates what SWE-bench actually measures.
Abacus.AI Fine-Tunes Kimi K3 Into an Agentic Specialist That Beats It on Four Benchmarks
Smaug-Agentic, an open-weight agentic fine-tune of Moonshot AI's 2.8-trillion-parameter Kimi K3, ranks second on LiveBench agentic coding at 64.6 and delivers successful coding tasks at $0.071 each — less than a third the cost of Claude Fable 5.
Marvell Grants Google Option to Buy $12.2 Billion Stake in Custom Chip Alliance
Marvell Technology has given Google the right to acquire up to $12.2 billion in stock as part of a broad custom chip partnership covering Google's TPU ecosystem. The deal puts Marvell in direct competition with Broadcom, which already holds a Google chip supply agreement extending through 2031.
OpenAI Publishes Cyber Pacing Framework: Will Gate Model Releases When Capabilities Cross Critical Thresholds
OpenAI released a formal policy for slowing or halting model development when it reaches cyber-critical capability levels. The framework codifies a self-imposed brake tied to national security risk, institutionalising a practice the company had previously applied only on a case-by-case basis.
Stripe Acquires OpenRouter, the 10-Trillion-Token-Per-Day AI Model Gateway
OpenRouter, the largest model marketplace routing 10+ trillion tokens daily across 400+ AI models for 10 million developers, announced it is joining Stripe. The deal makes Stripe the financial and routing infrastructure layer for a significant share of the AI inference economy.
Anthropic Confirms Major Claude Outage: Login, API, and Claude.ai All Degraded on August 16
Anthropic confirmed a platform-wide outage beginning August 16, 2026, at approximately 21:58 UTC. Login, the Claude.ai web interface, and API access all reported degraded service or complete failure. It is among the broadest multi-service disruptions Claude has experienced.
Mathematicians Are Confronting Whether Their Profession Has a Future
A new paper and Washington Post reporting document how leading mathematicians gathered at OpenAI's San Francisco offices to discuss what human researchers will do once AI is superhuman at mathematics. The consensus was uncomfortable.
Turbovec Brings Google Research's TurboQuant Quantization to Rust, Targeting FAISS Workloads
An open-source vector index built on Google Research's TurboQuant algorithm has landed on Hacker News with early benchmarks suggesting it can outperform FAISS on high-throughput similarity search. The library ships a Rust core with Python bindings, targeting production embedding pipelines.
Ornith-1.5 Closes the Loop: Model Now Proposes Its Own Training Tasks
DeepReinforce releases Ornith-1.5 in 397B, 35B, and 9B variants, extending its self-scaffolding framework into a closed self-improvement loop where the model generates the tasks it trains on rather than pulling from human-curated sets.
AI Data Centers Are Raising Phoenix Temperatures by Up to 4 Degrees, Peer-Reviewed Study Finds
An ASME journal paper documents heat island effects from data center waste heat in Phoenix, with nearby neighborhoods seeing up to 4°C of localized warming. Cities from Texas to Florida are now moving to restrict siting, as AI infrastructure collides with urban climate policy.
OpenAI Halts Frontier RL Training for Two Weeks After Astra Hits Critical Cybersecurity Threshold
OpenAI stopped deployment-focused RL training across its frontier models after evaluations of Astra suggested the model may have crossed the Critical cybersecurity threshold — the bar built for autonomous zero-day exploit capability. A larger planned RL run remains on hold.
Shoals and ON.energy Deliver 1.1 GW Battery Project for AI Data Center — Largest in the U.S.
Shoals Technologies Group and ON.energy have begun manufacturing and deliveries on a 1.1 GW grid-connected battery storage project at a hyperscale AI data center, the largest such project in the United States. The hardware is domestically manufactured using ON.energy's medium-voltage AI UPS architecture.
Gen Z Took 69% of AI Engineer Hires in 2025 as Women Held at 26%
LinkedIn data shows Gen Z captured 69% of AI engineer hires and 68% of forward-deployed engineer roles in 2025, while women made up just 26% of all AI hires. Millennials dominated leadership, taking 60% of head of AI appointments.
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.