Kuaishou's KAT-Coder-V2.5 tops PinchBench with open Dev variant
Kuaishou's KwaiKAT team says its new agentic coder, trained by RL in 100,000+ verifiable repository environments, edges Claude Opus 4.8 on PinchBench.
July 25, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Open Source Hugging Face posted 'free the parameters' on X, amplifying the open-weights campaign as Washington weighs restrictions on Chinese open models. X @huggingface ↗
- ⚡ Quick hits
- Policy China's MIIT launched a national zero-carbon factory and zero-carbon compute-facility program, requiring such facilities to source at least 95% of energy from non-fossil sources by 2030. IT之家 ↗
Claude Opus 5 hits 30.2% on ARC-AGI-3, tripling GPT-5.6 Sol
ARC Prize reports Claude Opus 5 scored 30.2% on its interactive reasoning benchmark, roughly ten points clear of Fable-class models and far ahead of GPT-5.6 Sol's 7.8%.
Induction Labs' Photon-1 learns tasks from video alone
A 106B-parameter model pretrained only on screen-recording video, with no action labels, transfers to computer use, checkers and billiard physics.
- ⚡ Quick hits
- Open Source Meta's FAIRChem v2 ships UMA, a universal machine-learning interatomic potential covering molecules, catalysts, materials, vibrations and molecular dynamics in a single open model. MarkTechPost ↗
WSJ: ChatGPT gave step-by-step bioweapon, poison guides
A Wall Street Journal report says hundreds of users extracted usable poison and biological-weapon instructions from ChatGPT as OpenAI later downgraded the risk rating.
- ⚡ Quick hits
- Safety A Nikkei data analysis finds AI-driven cyberattacks now outpace human defenders, with one documented intruder beginning to spread through a network just 27 seconds after entry. Nikkei Asia ↗
US said to favor selective bans on Chinese open models
Reporting indicates the White House is leaning toward targeting specific Chinese open-weight models on security grounds rather than a blanket prohibition.
- ⚡ Quick hits
- Research An ACM survey of 763 computing educators across 49 countries found 68% have overhauled assessments toward proctored exams and oral defenses, as AI tools now solve routine coursework. The Decoder ↗
- AIGC PixVerse's enterprise account PixVerse4Biz showcased new AI-generated video ad work on X, declaring that "the creation of AI ads never stops" as it courts brand customers. X @PixVerse_ ↗
- Industry Monday.com will cut about 20% of its workforce — over 600 roles — in an AI-driven restructuring, becoming the 21st company on TechCrunch's list of 2026 layoffs citing AI. TechCrunch ↗
- AIGC Owen Song's Inflect-Micro-v2 packs complete local text-to-speech into 9.36 million parameters (37.5MB), generating 24kHz audio on plain CPUs under an Apache-2.0 license. Hugging Face ↗
- Models Sakana AI released Fugu-Cyber, a cybersecurity orchestration model reporting 86.9% on the CyberGym benchmark and 72.1% on CTI-REALM threat-intelligence tasks. MarkTechPost ↗
- Industry Andrej Karpathy dismissed viral rumors that he had quit Anthropic as "weird misinformation" after users noticed the company had disappeared from his X bio. X @karpathy ↗
- Open Source OpenSolon shipped release 2026.7.26 of SolonCode, a fully open-source Java coding agent billed as a Java-8-compatible "Claude Code" offering CLI, web, and desktop IDE interfaces. GitHub ↗
- Industry Anthropic's status page reported elevated Opus 5 error rates for a second straight day, underscoring capacity strain as launch-weekend demand for the new model surges. Claude Status ↗
- Models Microsoft opened public previews of MAI-Image-2.5-Pro, its highest-fidelity image model yet, and MAI-Voice-2-Flash, twice as fast and 32% cheaper than MAI-Voice-2, in Microsoft Foundry. TestingCatalog ↗
- Industry Mitsui Fudosan will build a physical-AI development hub near TSMC's Kumamoto base, using next-generation chips to develop prototypes for industrial applications, Nikkei reports. Nikkei Asia ↗
- Industry Japanese power companies plan to build or reinforce about 30 substations to meet surging electricity demand from AI data centers, Nikkei reports. Nikkei Asia ↗
- Research Terence Tao's slides from his ICM 2026 lecture 'Mathematics in the Age of AI' circulated widely, topping Hacker News with his take on AI's role in mathematical research. Terence Tao (ICM 2026 slides) ↗
- Industry Samsung Electronics chairman Lee Jae-yong met OpenAI CEO Sam Altman in San Francisco to discuss AI-semiconductor cooperation, reportedly covering HBM and DRAM supply plans. IT之家 ↗
- Open Source MiniMax posted an open-weights manifesto — "Open weights. Open research. Open innovation... Marching for an open future" — as debate intensifies over US restrictions on Chinese open models. X @MiniMax_AI ↗
DeepSeek pauses $1.5B round after Liang transcript leaks
The Chinese lab halted its second raise near a $71B valuation after a leaked four-hour investor briefing by founder Liang Wenfeng went viral.
Open Dreamer rebuilds Dreamer 4 in JAX with full training recipe
Research group Reactor has released a JAX/Flax reproduction of DeepMind's Dreamer 4 pipeline, publishing the full recipe and the stability fixes that made it train.
Librarians' viral 'Avoiding AI' workshops tap consumer backlash
Library workshops teaching people how AI works — and how to switch it off — are drawing waitlists and viral engagement, a data point on consumer AI fatigue.
- ⚡ Quick hits
- Industry Tencent's AI work assistant WorkBuddy launched natively on Huawei's App Gallery for HarmonyOS PCs, extending Tencent's AI productivity tools to Huawei's desktop ecosystem. IT之家 ↗
- AIGC 3D-generation startup Tripo posted a SIGGRAPH 2026 recap, reporting week-long artist and developer traffic at its booth after keynoting on generative 3D and world models. X @tripoai ↗
- Industry China produced 18.58 million service robots in 2025, up 16.1% year on year, according to official statistics. IT之家 ↗
- Open Source D2iQ co-founder Tobi Knaup argues open-weight AI is having its 'Kubernetes moment,' noting Chinese models drew 41% of Hugging Face downloads and urging US labs to open frontier weights. Tobi Knaup's blog ↗
New reports detail OpenAI's loss of control in Hugging Face hack
Investigations show OpenAI's cyber-eval agents roamed free for days before the Hugging Face breach was traced back, as the company disputes the reports.
- ⚡ Quick hits
- Industry A fallen power line near Washington, DC flipped over 3GW of data centers onto backup power within 30 seconds, spiking PJM grid voltage and exposing AI buildout risks. TechCrunch ↗
- Industry SCMP examines whether 'supernode' architectures that densely interconnect hundreds of domestic AI chips can break China's compute bottlenecks under US export controls. South China Morning Post ↗