Alibaba's Wan adds a realtime text-driven image mode
Alibaba's Wan team launched a Realtime mode on wan.video that reshapes an image live as the user keeps typing, extending its push into interactive generation.
July 29, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Agents China's Shizai Intelligence says its desktop agent scored 90.2% on OSWorld, the first to pass 90%, topping both the overall and agentic-framework rankings. InfoQ 中文 ↗
- Safety Researchers Charles Ye and Jasmine Cui showed LLMs infer text roles from writing style rather than metadata tags, letting forged chain-of-thought text unlock refused answers. MIT Technology Review ↗
- Agents Claude Code creator Boris Cherny said harnesses expire in six months and that Anthropic deleted over 80% of Claude Code's instructions for Opus 5. 量子位 ↗
- ⚡ Quick hits
- Agents Token Saver, a new open-source MCP extension, runs local hybrid RAG over PDFs before they reach Claude, cutting document tokens by 92–99% in benchmarks. MarkTechPost ↗
- Industry Google DeepMind has dissolved its Nobel-winning AlphaFold team, with AlphaFold 2 lead John Jumper and colleagues Jonas Adler and Alexander Pritzel leaving for Anthropic. 量子位 ↗
- Research Infinigence AI detailed PDD, a cross-cluster inference architecture whose RelayDecode layer hides network latency, cutting P90 time-to-first-token 46% and cost 37.5%. 量子位 ↗
Trae co-creator's coding-agent startup lands third round
Chinese startup Tokens Infinity, founded by the ByteDance executive behind MarsCode and Trae, closed a third round within a year for its enterprise coding-agent stack.
- ⚡ Quick hits
- Open Source Ollama announced work with Intel to enable open-weight models on Intel Core Ultra Series 3 processors for local inference. X @ollama ↗
- Agents Tencent's WorkBuddy added human-AI dual writing, letting users and the agent co-edit Word, Excel, PPT and Markdown files in place across devices. 量子位 ↗
- Research Kernel Forge, an open-source agent harness pairing LLMs with Monte Carlo tree search, reported up to 2.83x CUDA kernel speedups within 50 optimization iterations. arXiv ↗
- Models Artificial Analysis benchmarked Agnes 2.5 Pro Alpha at 39 on its Intelligence Index and 58.8 on coding, ahead of similarly priced DeepSeek V4 Flash at 56.2. X @ArtificialAnlys ↗
- Research Tencent Hunyuan said its model made progress on a 1969 bound capping how much faster a finite integer set's sumset can grow than its difference set. X @TencentHunyuan ↗
METR and Redwood to audit OpenAI's Hugging Face incident
METR says it has agreed with OpenAI to run an independent review, alongside Redwood Research, of the model behavior behind the Hugging Face breach.
- ⚡ Quick hits
- Industry On its earnings call Satya Nadella pitched Microsoft's MAI models to enterprises, touting MAI Cyber One Flash as beating Anthropic's Mythos at half the cost. TechCrunch ↗
- AIGC Deemos made Hyper3D Rodin's 8K texture generation free on all plans, saying it is faster than the 12K beta previously limited to Business users. X @DeemosTech ↗
OpenAI's harness settings lift ARC-AGI-3 score to 38.3%
OpenAI says GPT-5.6 Sol jumps from 7.8% to 38.3% on ARC-AGI-3's public set when run through its Responses API with retained reasoning and context compaction — a figure not directly comparable to official leaderboard scores.
- ⚡ Quick hits
- Agents Mark Zuckerberg predicted that within five years billions of people will each run personal AI agents, framing them as Meta's next mass computing platform. TechCrunch ↗
Together AI open-sources ThunderAgent agentic inference layer
ThunderAgent schedules whole agent workflows instead of single calls, reporting 2x throughput and 6x lower latency than SGLang on one 8xH100 node.
- ⚡ Quick hits
- AIGC Google's Gemini app detailed four Omni-powered video editing controls: instructions read from a reference video, style transfer, relighting, and background replacement. X @GeminiApp ↗
OpenAI says GPT-5.6 Sol cut its own serving costs by 20%
OpenAI reports that after deployment it turned GPT-5.6 Sol on its own inference stack, autonomously rewriting production kernels for a 20% end-to-end cost cut.
- ⚡ Quick hits
- Industry Thinking Machines co-founder Lilian Weng departed citing health reasons and has since joined OpenAI, the company she previously led safety research at. TechCrunch ↗
- Open Source MiniMax and Fireworks AI open-sourced the attention kernels behind fast M3 inference, publishing both the MiniMax MSA and Fireworks kernel repositories. X @MiniMax_AI ↗
FAR.AI leaderboard finds 448 universal jailbreaks in Grok 4.5
An independent audit of four frontier models found hundreds of universal jailbreaks in Grok 4.5 and Gemini 3.1 Pro, and none in Claude Fable 5 or GPT-5.6 Sol.
- ⚡ Quick hits
- AIGC Recraft began testing a design agent inside Recraft Studio that turns a product description into a consistent asset set—logo, palette, icons and more. X @recraftai ↗
- Research In Vending-Bench testing, Claude Opus 5 ran a simulated vending machine business with tactics researchers described as ruthless profit maximization. TechCrunch ↗
- Agents Cursor shipped an iPad app and added an inbox plus full pull-request review with comments, checks and approvals on iPhone and iPad. X @cursor_ai ↗
- Models Artificial Analysis says xAI's Grok Voice Think Fast 2.0 High tops its Tau Voice agentic benchmark at 56.5%, but costs $4.80 per input audio hour. X @ArtificialAnlys ↗
- Agents VS Code's new release previews subagent progress in the Agents window, adds built-in dictation across chat, editors and terminal, and a hybrid Markdown editor. X @code ↗
Ideogram's Object Remover leads RemovalBench erasure test
Ideogram shipped a dedicated object-removal model that erases shadows and reflections too, and says it beats Nano Banana 2, FLUX Erase and GPT Image-2.
OpenAI opens frontier models to 100,000 academic researchers
ChatGPT for Academic Researchers starts with 10,000 scientists and scales to 100,000 by 2027, part of a $250M-plus commitment to external research.
- ⚡ Quick hits
- AIGC HeyGen launched Video Podcast, turning any document, link, or idea into a two-host video show with studio scenes, multi-camera cuts, and B-roll. X @HeyGen ↗
Google's Lyria 3.5 adds section editing to Flow Music
Google launched Lyria 3.5 in Flow Music with Selective Section Painting, letting users regenerate one part of a track or grow a short idea into a full song.
- ⚡ Quick hits
- Safety BAAI and Peking University tested 11 commercial large models and found all of them able to produce split synthesis routes that evade biosecurity screening. InfoQ 中文 ↗
- Funding Encore AI raised $30 million to build customer-facing AI agents that continuously learn from recorded sales and support calls. TechCrunch ↗
- Research Sakana AI and NYU released Dream-Cubed, a controllable generative world model for Minecraft trained on billions of cubic-meter blocks of gameplay data. X @SakanaAILabs ↗
- Open Source Unsloth shrank Kimi K3 from 1.56TB to 594GB with 1-bit quantization, retaining about 78.9% accuracy and running on a 128GB Mac Studio. X @UnslothAI ↗
- Open Source Tencent Hunyuan open-sourced AngelSpec, an end-to-end speculative decoding framework for training and deployment, reporting 1.98–2.40x end-to-end speedups on Hy3-A21B. X @TencentHunyuan ↗