AI & design news, merged daily.
AI & design briefing for Wednesday, November 18, 2026
Loading today's digest…
No digest available yet today. Check back soon.
Merging & deduplicating today’s sources…
tldr
- OpenAI released GPT-6.2 Astra and Sol to mixed early reviews, while Microsoft introduced seven models spanning tasks from coding to speech. OpenAI also signed agreements with four independent safety assessors, and Apple’s macOS beta adds more granular permissions. Separately, Isomorphic Labs raised $5 billion for AI drug design.
- Tests on obscure protein sequences suggest recent Claude models learned from sequence data. An agent-harness explainer describes how software manages model tools and permissions, while Nvidia announced a scientific-computing GPU expected in 2028.
- Google’s Gemini can now help with tasks such as appointment scheduling and shopping. Suno added tools for generating images and short videos from music, while Claude Code is experimenting with chained tool calls.
Top Stories
- OpenAI launches GPT-6.2 Astra and Sol. OpenAI positions Astra for complex professional work and Sol for everyday tasks. Some early testers criticized Astra's writing and delegated work, while others reported mixed results. [1] [2]
- Microsoft introduces seven AI models. Microsoft announced models spanning coding, cybersecurity, image generation, speech and decision-making, alongside its Maia 200 accelerator now in production. [3]
- Isomorphic Labs raises $5 billion for AI drug design. The Series C round, co-led by Greenoaks and Sequoia, will support the company's drug-design engine and therapeutic candidate pipeline. [4]
- OpenAI signs agreements for independent safety assessments. Four outside organizations will assess model alignment and safeguards, with access to internal resources tailored to their investigations. [5]
- Apple changes Mac permissions for AI agents. The macOS 27.3 beta introduces more granular app and folder access controls and an indicator for apps controlling other apps. [6]
Deeper Cuts
- Researcher probes Claude's protein sequence knowledge. Xinming Tu's updated tests find recent Claude models predict little-studied protein sequences better than shuffled sequences, supporting an inference about training data. [7]
- What an agent harness actually controls. Nielsen Norman Group explains how software around a model supplies context, runs tool calls and enforces permissions, shaping the user experience. [8]
- A designer documents building an interactive portfolio. Tom Johnson shares iterations, discarded ideas and prompts behind Tomjohnville, including decisions about onboarding, visual style and performance. [9]
- Nvidia announces Rubin HPC GPU for scientific computing. The GPU is designed for simulation and analytics workloads and is expected to become available in 2028. [10]
- TensorWave signs infrastructure deal with OpenAI. TensorWave says its multiyear partnership will provide OpenAI with 457 megawatts of AMD-powered AI infrastructure. [11]
Tools and Launches
- Gemini adds task mode for everyday chores. Google says Gemini can schedule appointments, analyze bills and shop through integrated services, using connected information where needed. [12]
- Suno launches music-focused image and video generation. Visuals makes images and clips up to 15 seconds from songs. Visual credits are purchased separately. [13]
- Claude Code experiments with chained tool calls. Code Mode runs JavaScript that combines tool calls. It is experimental, off by default and limited to supported terminal sessions. [14]
- OpenAI introduces ChatGPT for Sales. A sales plugin offers workflows for account research, call preparation and follow-up, alongside new and expanded partner plugins. [15]
- Chatr Labs publishes quantized Qwen image builds. The builds support image generation and editing, but require additional components from the base model repository. [16]
Prompt of the day
One practical prompt for your preferred AI assistant from today's digest. [15]
Turn my meeting notes into a useful follow-up and a private action list. Ask me for the notes and the intended recipient if I have not supplied them. Extract decisions, commitments, owners and dates only where the notes support them; mark missing details as unresolved rather than guessing. Draft a concise message that reflects the recipient's priorities and makes the next step clear. Separately list my internal actions and questions I should resolve before sending. Keep confidential or speculative points out of the message. The result is ready for my review when every commitment can be traced to the notes and the draft does not imply agreement where none is recorded. Do not send or update anything.
References
- OpenAI: Introducing GPT-6.2 Astra and Sol
- VentureBeat: 'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users
- Microsoft AI: Seven new world class models at scalable price-performance
- isomorphiclabs.com: Isomorphic Labs announces Series C investment round
- OpenAI: Our first phase of embedded assessments
- MacRumors: macOS 27.3 Warns Against Giving AI Apps Full Mac Control
- Xinming Tu: Claude was trained on protein sequences at scale
- Nielsen Norman Group: Agent Harness: Definition
- tomjohn.design: Tomjohnville | Tom Johnson
- NVIDIA: NVIDIA Announces Rubin HPC - A New GPU to Accelerate Agents for Science, Engineering and Analytics
- TensorWave: TensorWave to Provide OpenAI with 457 Megawatts of AMD-Powered AI Infrastructure
- Google: Put Gemini to work for you
- Suno: Introducing Visuals
- Claude Code Docs: Run tool calls from code with Code Mode
- OpenAI: Introducing ChatGPT for Sales
- Hugging Face: chatrlabs/Qwen-Image-2.1-Uncensored-Omni-Quantized · Hugging Face
Sources: The Neuron, The Rundown AI, Superhuman, The Deep View, AlphaSignal, TLDR AI, TLDR Design, Sidebar