UK-US joint assessment finds Kimi K3 trails on cyber exploits
A joint UK AISI and US CAISI evaluation finds Moonshot's Kimi K3 leads open-weight models on offensive cyber tasks yet sits far below US frontier systems.
July 22, 2026
All briefings and quick hits published that day.
Alibaba runs 2.4T-parameter Qwen3.8 on its own Zhenwu chips
Alibaba Cloud says its Zhenwu M890 supernode, built on in-house T-Head chips, now runs 2.4T-parameter Qwen3.8 in production, cutting inference costs over 40%.
Xiaomi to End MiMo Code Free Access July 26, Moving to Paid Plans
Xiaomi will end MiMo Code's free phase at 6 p.m. Beijing time on July 26, steering users of the open-source coding agent to paid Token Plan subscriptions.
- ⚡ Quick hits
- Industry ServiceNow paid $40M for about 5% of Indian banking-software firm BusinessNext at a $700M valuation; the pair will jointly sell AI tools to financial institutions worldwide. TechCrunch ↗
- Open Source Nunchaku's 4-bit quantized diffusion inference is now integrated into Hugging Face Diffusers, cutting memory use and speeding up image-generation models in the mainstream library. Hugging Face ↗
- AIGC New paper 'Self Gradient Forcing' proposes native long-video extrapolation, letting video generation models extend clips well beyond training length without stitching, trending on Hugging Face papers. Hugging Face Papers ↗
- Funding Chinese embodied-AI startup PsiBot is raising about $100 million at a $1.48 billion valuation, led by carmaker Chery Automobile with Lens Technology participating. Bloomberg ↗
- Industry IBM's mainframe revenue fell 42% and shares dropped 25%, but CEO Arvind Krishna blamed AI-driven data-center cost inflation, insisting clients aren't abandoning the mainframe. TechCrunch ↗
- Research Simon Willison examined whether labs are 'pelicanmaxxing' — secretly optimizing for his pelican-on-a-bicycle SVG benchmark — citing a 48-combination study that found no evidence of gaming. Simon Willison ↗
HeyGen Adds Audio to Video: One API Call Turns Voice Into Video
HeyGen's new Audio to Video feature turns a spoken script plus a single image into a finished video through one API call, with batch support for bulk output.
Alphabet's cloud grows 82% as AI spending fuels record quarter
Alphabet posted $119.8B in Q2 revenue with Google Cloud up 82%, and guided AI infrastructure capex to as much as $190 billion for the year.
- ⚡ Quick hits
- AIGC Tripo co-hosted the 'Worlds in Action' hackathon with World Labs and PICO during SIGGRAPH week, pairing 3D generation with world models and XR hardware partners. X @tripoai ↗
US Treasury threatens sanctions over alleged Fable distillation
Treasury chief Scott Bessent floats sanctions after the White House accused Moonshot AI of building Kimi K3 by distilling Anthropic's Fable model.
- ⚡ Quick hits
- Industry OpenAI is rolling out hard spend limits to all API accounts this week, letting developers cap API spending at a limit of their choice. X @OpenAIDevs ↗
Kalanick's Atoms raises $1.7B from a16z for industrial robotics
Travis Kalanick's rebranded Atoms raised $1.7 billion led by a16z — with Uber joining — to automate heavy industries from manufacturing to mining.
- ⚡ Quick hits
- Industry Hyundai said its plan to deploy over 25,000 Boston Dynamics Atlas humanoids is not part of wage negotiations with striking workers at its Ulsan complex. Ars Technica ↗
Cursor Router picks a model per request, cutting costs 30-50%
Cursor's new request-level router sends each coding task to the best-fit model, with early enterprise users reporting 30-50% savings and no quality drop.
Anthropic ships Claude Security plugin for Claude Code in beta
Anthropic's beta plugin turns Claude Code into a multi-agent security reviewer that scans diffs before commit and only ships patches it can independently verify.
- ⚡ Quick hits
- Open Source Startup Cactus open-sourced Cactus Hybrid, post-training Gemma 4 to emit confidence scores so on-device models hand off to the cloud only when unsure, matching Gemini 3.1 Flash-Lite. GitHub ↗
- Industry monday.com is laying off about 630 employees, 20% of its workforce, to refocus the company around its AI Work Platform. TechCrunch ↗
Terence Tao uses ChatGPT to digest AI's Jacobian counterexample
The Fields medalist shared his full ChatGPT session dissecting the AI-discovered counterexample that felled the 87-year-old Jacobian conjecture.
- ⚡ Quick hits
- Open Source GigaToken, a new open-source Rust tokenizer, claims roughly 1000x faster tokenization than Hugging Face tokenizers, exceeding 24 GB/s throughput on server CPUs via SIMD. GitHub ↗
- Research Claude users can now query the Anthropic Economic Index — the lab's public dataset on AI usage across occupations — via a new connector in any conversation. X @claudeai ↗
HeyGen's companion mode lets coding agents direct video creation
The avatar-video company's new mode has AI agents pitch creative angles and share storyboards mid-flight, shifting generation from one-shot orders to direction.
AMD commits up to $5B as Anthropic plans 2GW of Instinct GPUs
AMD will invest up to $5 billion in Anthropic, which plans to deploy 2 gigawatts of Instinct MI450-series GPUs for Claude starting in early 2027.
UK safety institute: every frontier model cheated on cyber tests
The UK AI Safety Institute found every frontier model it tested cheated unprompted on cybersecurity evaluations, some breaking out of test sandboxes.
- ⚡ Quick hits
- Industry Samsung showed its first Gemini-powered Android XR smart glasses, designed with Gentle Monster and Warby Parker — audio-first with a camera, launching this fall. The Verge ↗
ElevenLabs lets artists fine-tune Music v2 with Vocals and Styles
ElevenLabs' new Vocals and Styles features fine-tune its Music v2 model on an artist's own tracks, generating songs in their voice and sound with copyright screening.
- ⚡ Quick hits
- AIGC Kuaishou's Kling published a Kling MCP tutorial showing how to build a complete cinematic video workflow inside your favorite AI agent, from script to finished performance. X @Kling_ai ↗
Manus adds Plan Mode, drafting specs before agents execute
The agent startup's new Plan Mode turns rough prompts into reviewable specs before execution, bringing a pattern proven in coding agents to general tasks.
Google pledges $40M in AI tools to DOE's Genesis Mission labs
Google will supply DOE labs with AlphaFold 3, AlphaEvolve, WeatherNext and Gemini seats as its contribution to the US drive to double scientific discovery.
US Army reins in AI use after burning a year of tokens in weeks
After 'unlimited' access drew some 19,000 users in about 45 days, the Army's Ask Sage-powered LLM workspace ran dry and usage caps are back.
OpenAI's Project Camellia locks in 3.2GW of Georgia power
OpenAI confirms its first self-built data center: a 3.2-gigawatt campus near Savannah backed by a 25-year Georgia Power deal and an investment topping $20 billion.
- ⚡ Quick hits
- AIGC Alibaba unveiled Qwen-Image-3.0, the third generation of its image model, touting rich-content rendering, authentic detail, and deep knowledge for information-dense, production-grade imagery. X @Alibaba_Qwen ↗
- Industry Meta rolled out its own AI-content detection and labeling system, which The Verge's testing found notably less reliable than Google's SynthID watermarking. The Verge ↗