AI & design news, merged daily.
AI & design briefing for Tuesday, November 3, 2026
Loading today's digest…
No digest available yet today. Check back soon.
Merging & deduplicating today’s sources…
tldr
- OpenAI is offering Rosalind 6 to reviewed research applicants, signed AI agreements in Abu Dhabi and announced flexible power plans for future campuses. MIT CSAIL researchers tested direct AI edits to model weights, while Anthropic opened a paid safety residency for technical leaders.
- Anthropic’s controlled tests and a Codex red team exposed potential gaps in AI safety controls. In a quantum physics preprint, AI agents revised an experiment after detecting measurement drift, while OpenAI introduced an evaluation of everyday tasks.
- Canva introduced a design-editing model, while ChatGPT is preparing to merge its desktop Chat and Work tabs. Claude’s API message threads entered beta, and DeepSeek opened a public preview of its agent harness.
Top Stories
- OpenAI offers Rosalind 6 through reviewed research access. The biology-focused model uses GPT-6 Astra weights and has fewer refusals on dual-use queries. OpenAI reviews applicants’ proposed uses, security controls and governance before granting access. [1]
- OpenAI signs four Abu Dhabi AI agreements. The agreements span government, energy and investment. A government pilot assesses food-import compliance against machine-readable rules, with officials retaining final decisions. [2]
- OpenAI plans flexible power use at future campuses. Its Grid Protectors program aims to reduce demand during grid strain. A planned Georgia campus includes an agreement for up to 1,000 megawatts of flexible demand response. [3]
- Researchers test AI editing of neural network weights. MIT CSAIL researchers report that GPT-6 Astra directly edited a small language model’s weights, improving its math benchmark score in an early experiment. [4]
- Anthropic opens safety residency for experienced technical leaders. The paid research placements last six weeks to six months. Machine learning experience is not required, and applications close November 22. [5]
Deeper Cuts
- Anthropic tests automated search for hidden misalignment. In a controlled experiment, Claude generated training data that passed a monitor yet induced misalignment in small research models. Those models were not deployed. [6]
- Red-teamers find gaps in Codex action controls. Researchers report 11 attacks against a tested Codex version’s sandbox and Auto-review. They shared findings with OpenAI, which fixed several vulnerabilities. [7]
- Small weight changes improve Gemma on selected tasks. DEJAN researchers trained a small fraction of Gemma 3 270M’s weights and report gains on sorting and JSON formatting, while multiplication remained difficult. [8]
- AI agents revisit quantum experiment after detecting drift. A preprint describes agents deriving and experimentally testing predictions about quantum spins. They treated an initial result as inconclusive after reference measurements revealed drift. [9]
- OpenAI introduces LifeVal for everyday AI tasks. The evaluation covers tasks including caregiving, household administration and government services, informed by time-use studies and a survey of 3,995 respondents. [10]
- Google receives AI response from orbiting prototype. Google says its Project Suncatcher satellite generated a haiku using an onboard TPU and Gemma model. Thermal and radiation testing comes next. [11]
- Microsoft patent describes altering live speakers’ words. A reported patent filing describes changing live video and audio, including speakers’ mouth movements. Proposed uses range from translation to removing particular phrases. [12]
- Musk says Grok Bot serving costs exceed revenue. After users reported hitting weekly limits, Elon Musk said the agent’s serving costs were unsustainably high relative to revenue. [13]
- SpaceXSI describes its shared data analysis bot. Its data team maintains shared instructions for metrics, experiments and charts. Managers approve or reject proposed additions to the bot’s shared memory. [14]
Tools and Launches
- Canva introduces Brush for design editing. The in-house vision-language model adjusts designs and inspects results, with Canva reporting faster, less costly edits than its previous agent. [15]
- ChatGPT prepares to merge Chat and Work tabs. The desktop change will reach select users first. Codex remains separate, while ChatGPT and Codex share a token allowance. [16]
- Claude API message threads enter public beta. Threads send only new messages between turns. They expire after five minutes of inactivity, and token usage remains unchanged. [17]
- DeepSeek Harness enters worldwide public preview. The open-source desktop and web agent harness supports plugins, coding, research and recurring tasks. [18]
- DeepSeek team publishes superagent research prototype. The repository agent works on requested GitHub tasks through DeepSeek Harness. Its authors caution that it is not production-ready. [19]
- Xmind presents conversational mind-mapping agent. Xmind Agent builds and reshapes a mind map through conversation, starting from a thought, file, link or conversation. [20]
Prompt of the day
One practical prompt for your preferred AI assistant from today's digest. [14]
Help me investigate a change in a metric without jumping to a cause. Ask for the metric definition, the time period, the observed change, and whatever table or chart I can share. Use the data I provide to establish the baseline and, where possible, break the change down by useful segments or time intervals. Consider competing explanations such as seasonality, holidays, missing data, a logging change or a real change in behavior. Separate observations from hypotheses and identify the smallest checks that would distinguish them. If you can analyze files, show the calculations; otherwise, give me formulas or steps for my existing spreadsheet. Deliver a short findings note with evidence, limitations and the next three checks. Do not claim causation or modify the source data.
References
- OpenAI: GPT-Rosalind-6 System Card Addendum
- mediaoffice.abudhabi: Khaled bin Mohamed bin Zayed meets Sam Altman and witnesses signing of agreements with OpenAI
- OpenAI: Introducing the Grid Protectors Program
- github.io: Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
- Anthropic: Anthropic Residency: applications open for established technical leaders
- Anthropic: Automated Subtle Emergent Misalignment Research
- lesswrong.com: Lessons From Red-teaming Auto-review, OpenAI's Blocking Monitor
- DEJAN: Unlocking hidden skills inside Google's model
- arXiv: Collaborative AI agents for autonomous experimental research with solid-state quantum spins
- OpenAI: LifeVal | OpenAI
- Google: We’ve received our first response from the Project Suncatcher prototype satellite.
- Futurism: Microsoft Patents an AI That Can Literally Put Words in People's Mouths in Real Time
- Startup Fortune: Elon Musk admits Grok Bot costs too much to run as users hit weekly caps
- x.ai: Giving every team a data scientist with Grok Bot
- Ken Yeung: Canva Unveils Brush, an AI Model Built to Edit Designs
- TechCrunch: OpenAI is finally merging ChatGPT's Chat and Work tabs
- Claude: Make long-running agents faster with message threads
- deepseek.com: DeepSeek Harness
- GitHub: deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System
- xmind.com: Xmind - Mind Map Maker & AI Brainstorming Tool
Sources: TLDR AI, The Rundown AI, The Neuron, The Deep View, Superhuman, AlphaSignal, TLDR Design