Live Feed
RAMageddon Goes Forward: Samsung, SK Hynix, and Micron Have Sold All of 2027's DRAM and HBM
The big three memory manufacturers have reportedly sold through their entire 2027 DRAM and HBM production capacity. AI hyperscalers locked it up on 5-year contracts. Consumer RAM is up 52% since January. Xbox and Steam pricing already adjusted.
OpenAI's August GPT-5.6 Update: Sol Clears 4 Cyber Scenarios Luna Cannot — Both Stay 'High Risk'
OpenAI ships updated GPT-5.6 Sol and Luna to ChatGPT today. Both are rated High capability in Cybersecurity and Biological/Chemical domains. Sol fails 2 of 6 cyber scenarios; Luna fails 5. AI Self-Improvement evals were skipped.
DOE and Arcee Are Building America's First Federal Open-Weight Science Model — Contributions Due August 25
The Department of Energy and Arcee AI are building Genesis-Science-1, a trillion-parameter open-weight model for national lab research. The public contribution portal is live with applications closing August 25.
OpenAI's First Consumer Device Is a $300 Screenless AI Speaker — and a Bet Against the Smartphone
Bloomberg reports OpenAI's debut hardware is a doughnut-shaped, battery-powered speaker with a camera and moving parts but no screen. Priced at $300-$400, targeting 2027, it is the first in a broader device family and is designed by Jony Ive's LoveFrom.
SK hynix Commits 54 Trillion Won to Two New Fabs — Greenfield Bets on AI Memory Demand Through 2030
SK hynix approved 54 trillion won in new fab construction: 35.2 trillion won for Yongin Y2 (cleanroom June 2029) and 19.1 trillion won for Cheongju M17 (December 2028). Both facilities target DRAM and NAND expansion as AI infrastructure demand outpaces current global HBM supply.
Human Overseers Miss 1 in 3 Malicious Agent Commands — 40,000-Run Study Breaks the Safety Assumption
A browser game built to simulate AI coding agent permission prompts logged 409,000 approve/deny decisions across 40,000+ runs. Humans approved one-third of dangerous commands on average, with credential-exfiltration attempts slipping through at 35% and npm run payloads at 52.5%.
GPT-5.6 Luna Is Now the Default Model for Free ChatGPT Users — 62% Fewer Factual Errors Than GPT-5.5
OpenAI upgraded the free tier of ChatGPT to GPT-5.6 Luna as its default model with unlimited text chats, while a tuned version of Sol arrives for Plus and Pro users with 68% fewer factual errors versus GPT-5.5 Instant and a new reasoning-depth slider.
Claude Opus 5 Max Takes Fullstack Code Arena #1 at 1,699 Elo — 61 Points Clear of GPT-5.6 Sol
Anthropic's Claude Opus 5 Max debuted at the top of Arena's new Fullstack Code leaderboard with an Elo of 1,699, beating GPT-5.6 Sol by approximately 61 points on end-to-end web development tasks spanning multi-step reasoning, database integration, and API orchestration.
South Korea's Government Shaped Its Largest B200 Cluster: NHN Cloud FactoryX Seoul Goes Live
NHN Cloud activated FactoryX Seoul with 7,656 Nvidia B200 GPUs. At the Korean government's explicit request, 4,080 of them form a single unified cluster — the largest concentrated AI training environment in South Korea.
Qwen3.8 Max Takes AA Agentic Index #1 at Intelligence 56 — but Costs $1.14 Per Task, Twice Its Predecessor
Alibaba's Qwen3.8 Max is now ranked first on Artificial Analysis's agentic index at Intelligence Index 56, ahead of all US labs except Anthropic and OpenAI. The upgrade cost: per-task cost doubled from $0.53 to $1.14 compared to Qwen3.7 Max.
Envision's Galaxy Base: 120,000 sqm AI Data Center in China's Gobi Desert Skips the Grid Entirely
Chinese green technology company Envision has commissioned what it claims is the world's largest single AI data center building in Ulanqab, Inner Mongolia. The facility connects directly to on-site renewables, bypassing grid interconnection queues that are stalling Western AI infrastructure by five years or more.
AMD Acquires Taalas to Bake Model Weights Into Silicon — 48x Faster Inference, One Model Per Chip
AMD has acquired Toronto-based Taalas, whose chips etch AI model weights directly into silicon rather than loading them from HBM at runtime. The HC1 test chip hit 16,960 tokens per second on Llama 3.1 8B — 48x faster than Nvidia GPUs.
DeepSeek Confirms Significant API Price Hike — V4-Flash Demand Overwhelmed Its Proof-of-Concept Pricing
DeepSeek's V4-Flash 0731 became Ollama's fastest-growing model ever at $0.14/$0.28 per million tokens, then the company announced prices would rise significantly. At current rates it is 105x cheaper per task than Claude Fable 5 — but Meta Muse Spark and GPT-5.6 Luna now compete on capability and price.
Microsoft FY2026 Filings Reveal OpenAI Drove $24.1B — Roughly 70% of Its Actual AI Revenue
A first-time disclosure in Microsoft's FY2026 regulatory filings shows OpenAI commercial arrangements generated $24.1 billion in revenue — about 70% of Microsoft's actual AI sales, per Bloomberg Intelligence. The company has also invested $11.9 billion into OpenAI.
Finance Agent Harness Lifts GPT-5.5 From 55.8% to 79.1% on BigFinanceBench — Retrieval Beats Model Upgrades
Primer's domain-specific agent wrapping GPT-5.5 scored 79.1% on 928-question BigFinanceBench versus 55.8% for the raw model — a 23.3-point lift that newer frontier models have not closed. Sixteen of those points came from retrieval alone.