In partnership with
In Today’s Issue:
⚖️ Astra gets a legal toolkit
🏠 Google gives families a shared agent
🧪 Anthropic builds its own biology lab
🌏 Washington and Beijing face shared AI risks
✨ And more AI goodness…
⚡ The Signal
AI companies are building the environments their models need to get useful work done.
OpenAI is giving Astra a legal research toolkit that firms can connect to their own expertise, while Google’s CC gathers the information a household chooses to share. Claude’s redesigned projects tackle a different bottleneck: coordinating several pieces of work without making the user manage every handoff. Anthropic is moving into physical experiments too, with its own biology lab and an ambition to have Claude direct laboratory robots. I’m interested in how these systems connect an answer to the work that follows, whether that means reviewing a contract, finishing a build or testing an idea at the bench.
All the best,
Kim Isenberg
CC’s shared calendar, shown in Google’s official product demo. Image: Google.
🏠 Google’s CC Helps Families Coordinate
Google is turning CC into a shared household assistant for up to six people. It can assemble a daily brief, maintain calendars and help complete forms using information members choose to share, with permission required before it acts or shares outside the group. The experiment is limited to U.S. adults with personal Google accounts, with existing-user upgrades coming and a waitlist for newcomers.
👉 tl;dr: CC gives families one shared helper for scattered schedules and paperwork, while each person controls what it sees.
Claude projects suggest related coding threads. Screenshot: Anthropic.
💻 Claude Projects Manage Parallel Coding Tasks
Claude’s redesigned projects can now split a coding job across several parallel sessions and assemble the results. A coordinator delegates tasks, reviews the output and draws on memory shared across threads, so users can manage a build through one project conversation. The beta starts with selected Pro and Max subscribers using Claude Code cloud sessions who have no existing web or desktop projects; wider rollout and local execution follow.
👉 tl;dr: Give Claude a coding goal, and it can organize parallel tasks, review their output and bring the results together.
Anthropic signage at Dreamforce in San Francisco, September 17, 2026. Photo: Reuters / Carlos Barria.
🧪 Anthropic Has Its Own Biology Lab
Anthropic has quietly built a wet lab in the San Francisco Bay Area, bringing its AI research into physical biology experiments, Reuters reports. Its life-sciences chief, Eric Kauderer-Abrams, confirmed that the company is doing lab work in its own facilities and with external partners to test ideas and move faster. A person familiar with the effort said Anthropic wants Claude to direct laboratory robots that carry out experiments with limited human intervention.
👉 tl;dr: Anthropic now does physical biology research itself, with Claude-controlled lab robots among its ambitions.
🎬 Watch This
In Salesforce’s Dreamforce 2026 main keynote, jump to 45:23 for Marc Benioff’s conversation with Dario Amodei. Amodei explains his case for stronger safety practices, then describes using Claude with Salesforce data to examine sales prospects and deals at risk: a concrete example of why access to business information matters to him.
"Every company needs useful AI, tailored to its knowledge, expertise, and work."
– Jensen Huang, NVIDIA founder and CEO, in Salesforce’s September 15 Koa announcement.
Faster hardware alone cannot give an assistant the company knowledge Huang is describing. His comment accompanied Salesforce’s introduction of Koa, its reasoning model for customer-management software, built on NVIDIA Nemotron.
According to The Information, OpenAI is reportedly close to solving the Hodge Conjecture, one of the Millennium Prize Problems, citing a person familiar with the work. Employees expect a solution soon, but the report establishes neither a completed, publicly reviewed proof nor a date for an announcement.
OpenAI Gives Astra a Legal Toolkit
The Takeaway
👉 Astra for Law pairs GPT-6 Astra with legal tools and instructions.
👉 Firms can build their own research and document workflows around their expertise.
👉 Access starts with selected law firms in ChatGPT and Codex; the API is coming later.
👉 Better research results still leave lawyers responsible for checking the work.
OpenAI introduced Astra for Law on September 17, configuring GPT-6 Astra for professional legal work. A specialist search index covering more than 230 million URLs, combined with legal instructions, helps the system find relevant authority and use it in an answer. Anyone who has assembled a research memo knows the problem: finding a plausible case is only the start. A lawyer still has to establish whether it supports the argument. OpenAI is packaging that work into a foundation on which firms and legal software companies can build, with their own knowledge and review processes.
Astra for Law scores higher across weighted evaluation criteria. This measure differs from all-pass correctness. Chart: OpenAI.
Law firms already have methods worth preserving. Cooley GO Public applies the firm’s capital-markets expertise to drafting and reviewing filings for an initial public offering, or IPO. Sullivan & Cromwell brings its negotiating playbooks into agreement review, producing proposed contract edits and draft client advice. Ropes & Gray traces acquisition diligence findings back to their sources, including contract terms that could require a customer’s consent. These examples put the assistant inside a particular legal task. A lawyer can question its suggestions against the firm’s own standards and supporting documents, rather than trying to assess a polished answer on its own.
Cooley GO Public interface for IPO drafting and review. Image: Cooley / OpenAI.
There is evidence that the configuration helps. In OpenAI’s evaluation on 200 U.S. legal research questions, Astra for Law passed the overall correctness check 54.0% of the time, versus 38.7% for Astra with ordinary web search at the highest reasoning effort. That is a substantial improvement with considerable room left for error. Access initially goes to selected firms through Trusted Access in ChatGPT and Codex, with the API coming later; the legal offering also includes confidentiality controls. Firms will now need to test those research gains on their own matters, within their access rules and review processes.
Why it matters: A firm’s research habits and internal expertise can become part of the tool its lawyers use to prepare work. That could make AI assistance easier to examine and fit into practice, provided the firm keeps responsibility for the finished advice.
For product teams moving at AI speed.
AI makes it easier to ship anything, even bad ideas. The hard part is knowing which ideas are worth building.
Jira Product Discovery brings your ideas, customer insights, and priorities into one place, so your team can decide what to ship and move forward with confidence.
Capture ideas, prioritize with evidence, and build living roadmaps your team can rally around—all while staying connected to delivery in Jira, so everyone can see what’s being built and why.
Better product decisions in the AI era.
The chart: Returning to Astra for Law, this chart adds a cost comparison. On the legal research test, its lowest plotted effort setting achieves a 46.8% all-pass rate at $1.31 per answer, ahead of ordinary Astra with web search at its highest effort: 38.7% at $4.86. All-pass means meeting the evaluation’s overall correctness check.
The lesson: Buying more reasoning did less here than changing the research setup. The legal configuration stays above the baseline across the plotted settings, suggesting that task-specific tools and instructions deserve attention alongside raw computing budgets.
The caveat: This is OpenAI’s evaluation on 200 U.S. questions, comparing complete configurations; it does not isolate which component caused the gain. The dollar amounts are model costs, not a law firm’s total cost per matter, and even the best legal setting reaches only 54.0% all-pass.
🌏 Can AI Rivals Agree on Shared Risks?
⚡ Bottom line: AI governance is expected at next week’s planned Trump-Xi meeting, while both countries compete for an advantage.
💡 Why it matters: A dangerous capability developed in either country could create problems that the other cannot manage alone.
🔎 What it means: Watch for concrete channels to exchange risk information; a shared statement would leave harder verification questions unresolved.
AI safety is heading into a meeting between two governments determined to stay ahead of each other. AP reports that AI governance is expected on the agenda when Donald Trump and Xi Jinping meet in Washington next week. U.S. Treasury Secretary Scott Bessent told Axios the U.S. was open to discussing shared AI risks with China. That creates an opening for diplomacy, without establishing that either side has agreed on limits.
A visitor at Shanghai’s World AI Conference, July 17, 2026. Photo: AP / Ng Han Guan.
The difficulty is already visible. Washington treats leadership in advanced AI as a strategic advantage, and Trump has argued that stronger regulation could benefit China. Beijing objects to safety arguments it sees as a vehicle for containment. In its September 14 briefing, China’s foreign ministry advocated inclusive cooperation while rejecting confrontation. Both can support the language of safety while disagreeing about whose behavior needs to change.
A visitor tests Alibaba’s Qwen AI glasses in Shanghai, July 17, 2026. Photo: AP / Ng Han Guan.
China is also promoting international initiatives to spread its approach to AI. The same briefing described Xi’s call at BRICS, a group of emerging economies, for an open-source AI community and a shared digital cloud platform. Such initiatives make governance a contest over participation and influence as well as technical risk. For countries and companies outside the U.S. and China, the stakes include which systems they can adopt and whose standards accompany them.
Moonshot’s Kimi K3 exhibit in Shanghai, July 17, 2026. File photo: AP / Ng Han Guan.
A useful first outcome would be a dependable way to discuss dangerous incidents, with named officials and a process for follow-up. That is what we would look for; neither side has announced such an agreement in the reporting cited here. Coordinating a slowdown would demand much more: compatible definitions, reciprocal commitments and a way to establish whether promises are being kept. Next week’s test is whether the leaders create room for that work while continuing to compete.
The best marketing ideas come from marketers who live it. That’s what The Marketing Millennials delivers: real insights, fresh takes, and no fluff. Written by Daniel Murray, a marketer who knows what works, this newsletter cuts through the noise so you can stop guessing and start winning. Subscribe and level up your marketing game.
