Live Feed
Meituan Ships LongCat-2.0: 1.6T Open-Weights MoE Beats GPT-5.5 on SWE-bench Pro, Built on Chinese ASICs
Meituan's open-source LongCat-2.0 posts SWE-bench Pro 59.5, edging GPT-5.5 at 58.6. The 1.6 trillion parameter MoE was trained end-to-end on 50,000 domestic ASIC superpods with no Nvidia hardware in the stack.
Fable 5 Return Nears on Day 17 as Frontier Model Availability Becomes a Government Variable
Axios reports the Trump administration is close to lifting Fable 5 restrictions this week, while Mythos 5 was partially restored to 100 critical infrastructure orgs on June 26. GPT-5.6 remains locked to 20 partners. Gemini 3.5 Pro, the only major frontier model with no government review pending, targets July.
Grok 4.5 Enters Private Beta at SpaceX and Tesla: 1.5T Params, No Benchmarks, Monthly Releases
xAI's V9 model debuted June 28 at 1.5 trillion parameters with self-reported Claude Opus parity — but no third-party benchmark access, no public release date, and a known training limitation that its next successor is already designed to fix.
Anthropic Names Alibaba in Largest-Ever Distillation Attack: 28.8M Claude Exchanges, 25,000 Fake Accounts
Operators linked to Alibaba's Qwen lab ran a six-week campaign harvesting Claude's agentic reasoning and coding capabilities through nearly 25,000 fraudulent accounts, generating almost twice the volume of all prior Chinese lab attacks combined. Anthropic told the Senate Banking Committee the operation defied White House guidance.
Ornith-1.0 Writes Its Own RL Training Scaffolds: Open-Source 397B Matches Claude Opus 4.7 on SWE-Bench
DeepReinforce's MIT-licensed model family eliminates hand-written reinforcement learning harnesses by training the model to generate its own task scaffolds, then condition its solutions on them. The 397B flagship hits 82.4 on SWE-bench Verified and 77.5 on Terminal-Bench 2.1, matching Claude Opus 4.7 on both.
Samsung, SK Hynix, and Micron Hit With California Class Action Over AI-Era DRAM Price-Fixing
Seventeen plaintiffs filed a federal antitrust suit in the Northern District of California on June 25, alleging the three memory giants used the HBM transition and AI demand as cover to coordinate supply restrictions, driving DRAM prices up 700% over four years. The case draws on a prior 2005 conviction for identical conduct.
Micron Posts 84.6% Gross Margin in Q3 — Wall Street Bets It's AI Memory's Nvidia
Micron's Q3 revenue hit $41.46B with an 84.6% gross margin and Q4 guidance of $49-51B, driven by a structural shortage in high-bandwidth memory. Sixteen long-term supply agreements worth $22B — with pricing floors and take-or-pay terms — are designed to stop the boom-bust cycle that has plagued memory makers for 30 years.
Cambridge and NVIDIA's Red Queen Gödel Machine: AI Agents and Their Evaluators Co-Evolve Together
A new paper from Cambridge, NVIDIA, and collaborators shows that fixing the benchmark while improving the agent is self-defeating — the evaluator needs to improve too. On coding tasks the co-evolved system beats prior SOTA while using 1.35-1.72x fewer tokens. On paper writing it generates 1.86x higher acceptance from reviewer panels.
Samsung and SK Hynix Commit $518B to Korean Chip Hub — Shares Fall as Investors Price Capex Risk
South Korea's two chip giants announced $800 trillion won ($518B) in combined investment on Monday, anchoring a new semiconductor cluster. Samsung fell 4.7% and SK Hynix fell 3.1% the same day — the market pricing oversupply risk into the biggest capex commitment in the industry's history.
GPT-5.5 Instant's Third Update in 50 Days Ships No Benchmarks — OpenAI Pivots to Conversational Quality
On June 24, OpenAI pushed the third update to GPT-5.5 Instant since its May 5 launch, focusing on intent recognition and conversational quality. For the first time in this update cycle, the company published no quantitative benchmarks. The gpt-5.5-instant API endpoint is now a moving target, updating every 17 days without version bumps.
Karpathy's CLAUDE.md Grows to 10 Rules: A Self-Check Protocol Cuts Agent Error Rate From 41% to 11%
A ten-rule document attributed to Andrej Karpathy circulated on June 28, adding six self-monitoring rules to the four-rule CLAUDE.md template that has amassed 200,000 GitHub stars. The new rules teach autonomous coding loops to catch their own failure modes, not just write better code.
Semgrep: GLM-5.2 Scores 39% F1 on IDOR Detection, Beats Claude Code at $0.17 Per Vuln
Semgrep's vulnerability benchmark put GLM-5.2 at 39% F1 on IDOR detection, ahead of Claude Code's 32%. The open-weight model costs roughly one-sixth of comparable frontier models and runs entirely on-premise.
China's LineShine Displaces El Capitan as World's Fastest Supercomputer — CPU-Only, No NVIDIA
At ISC 2026 in Hamburg, China's LineShine posted 2.198 Exaflops to take the TOP500 #1 spot. No GPUs. Domestic chips only. The US hasn't topped the list in two years.
Microsoft Fairwater Goes Live in Wisconsin as One 800G AI Supercomputer
Microsoft's Fairwater campus in Mount Pleasant is operational, linking hundreds of thousands of NVIDIA GB200 GPUs with an 800G Ethernet fabric. The former Foxconn site has become a test case for whether AI campuses are data centers or single machines.
GPT Image 2 Lands on OpenRouter at $8/$8 Per Million Tokens
OpenRouter lists OpenAI's GPT Image 2 with $8 per million input tokens, $8 per million output tokens, and a 400K context window. The release puts image generation into the same token-metered procurement frame as text models.