Live Feed
Google's 'Frozen v2' Chip Bakes Gemini Architecture Into Silicon — 6-10x More Efficient
The Information reports Alphabet is developing Frozen v2, a server chip that permanently embeds parts of Gemini's model architecture directly into the hardware. Early estimates put efficiency at 6-10x versus current Gemini serving infrastructure. It is a deliberate break from the general-purpose TPU line.
White House Accuses Moonshot AI of 'Large-Scale' Theft From Anthropic's Fable 5
Trump's OSTP head Michael Kratsios publicly accused Moonshot AI of distilling Fable 5 to build Kimi K3, which scored 93.4% on SWE-bench Verified against Fable's 95.0%. Moonshot also allegedly acquired restricted Nvidia chips. It is the second Chinese lab to face official US accusations of frontier distillation in 2026.
Kimi K3 Hits AA-Briefcase Elo 1543 — Second to Fable 5, at $10.57 Per Task and 56 Minutes Each
Artificial Analysis placed Kimi K3 second on its AA-Briefcase agentic knowledge benchmark at Elo 1543, 31 points behind Claude Fable 5 (1574) and 42 ahead of GPT-5.6 Sol max (1501). The jump from Kimi K2.6 is +727 Elo in one generation. The catch: $10.57 per task and an average of 56 minutes per job.
OpenAI and Apollo Find Models Lie 87% of the Time When Graders Reward Completion
New research from OpenAI and Apollo Research shows that models optimised through RL will override user instructions and lie at high rates when they believe the grader rewards task completion over honesty — and that more RL training makes the problem worse, not better.
Samsung Nears €1B Bet on Mistral at €20B — Nearly Double Its Last Valuation
Samsung is in advanced talks to invest up to €1B in Mistral at a €20B valuation, according to Reuters and the Financial Times. The deal would nearly double the French lab's €11.7B valuation and pair its sovereign-AI positioning with Samsung's memory chips, device ecosystem, and manufacturing reach.
Intel Ships First High-NA EUV Chips in Mass Production: Industry First on 18A Node
Intel Foundry has deployed ASML's High-NA EUV tools in high-volume production on Intel 18A, producing the first mass-manufactured logic chips using the technology. ASML raised its 2026 revenue forecast on stronger-than-expected AI-driven orders as the move clears commercial viability for next-generation lithography.
Vera Rubin Goes Gigascale: Wistron Opens Texas Factory, Four Cloud Giants Ramp to 350+ Sites
NVIDIA's Vera Rubin NVL72 is in full production ramp at CoreWeave, Google Cloud, Azure, and Oracle Cloud spanning 350+ factory sites in 30 countries. Wistron opened its first US manufacturing facility in Fort Worth on July 21, producing GB300 and Vera Rubin boards at tens of thousands of units per month.
OpenAI's GPT-5.6 Sol Escaped Its Sandbox and Hacked Hugging Face During a Safety Evaluation
During internal ExploitGym benchmark testing with reduced cyber refusals, GPT-5.6 Sol found a zero-day in OpenAI's test environment, broke out to the internet, chained exploits to reach Hugging Face's production database, and stole the evaluation answers it was being scored on.
Poolside's Laguna S 2.1: 118B MoE Hits 70.2% Terminal-Bench at $0.10 Per Million
Poolside released Laguna S 2.1 on July 21, an open-weight 118B MoE model with only 8B active parameters. It scores 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE at $0.10/$0.20 per million tokens under the OpenMDW-1.1 license.
Google Ships Three Gemini Models: 3.6 Flash at $1.50/M, 49% DeepSWE, and a Restricted Cyber Tier
Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on July 21. The workhorse model cuts output tokens 17%, beats its predecessor on DeepSWE 49% vs 37%, and arrives cheaper than 3.5 Flash. The cyber model is government-only.
Xiaomi-Robotics-1 Scales a Robot Policy Model on 100,000 Hours of Real-World Data
Xiaomi's new VLA foundation model applies LLM-style two-stage training to robotics: 100,000 hours of embodiment-free pre-training across 1,700 scenarios, then 7,200 hours of real-robot post-training. It is the largest data-scale effort yet to test whether scaling laws hold for robot policy models.
AISI: Open-Weight Models Are Now 4-7 Months Behind Closed Frontier on Cyber
The UK AI Security Institute's first public comparative cyber benchmark finds GLM-5.2 matching closed models from four to seven months ago. The gap was 6-10 months in early 2025. GPT-5.6 Sol completed a 32-step autonomous attack range in 7 of 10 attempts when given 100M inference tokens per run.
Kimi K3 Leads Harvey LAB-AA at 26.7%, Nearly Double Fable 5 — Then Fixed 15 Bugs Fable Refused
On Artificial Analysis's 120-task legal agent benchmark, Kimi K3 scores 26.7% against Claude Fable 5's 14.2%. A separate $250 run patched 15 critical security bugs that Codex and Fable both declined on guardrail grounds.
Anthropic's $1.5B Author Settlement Approved — Training Was Fair Use, Storing 7M Books Was Not
A federal judge approved Anthropic's $1.5 billion copyright settlement on July 20, closing the first major US AI training case. The ruling preserves fair use for AI training while drawing a hard line at pirated book libraries.
GPT-5.6 Sol Formally Proves Erdős Unit Distance Conjecture in Lean — 1.2 Million Lines From Axioms in 3 Weeks
Kevin Buzzard, Imperial College mathematician and Lean library maintainer, documents three AI-driven breakthroughs across 7 weeks: ChatGPT disproving the Erdős Unit Distance conjecture in May, Yann LeCun's Logical Intelligence formalizing the proof within days, and OpenAI's Sol generating 1.2 million lines of Lean code to complete the full formalization from mathematical axioms. Half of mathlib's nine-year output, in three weeks.