Rates, ads, and the liability bill for AI
In this Brief
- Fed raises 25bp; Warsh keeps the path hawkish
- OpenAI turns the ad click into a sponsored conversation
- Bessent rejects a liability blank check for AI labs
- Anthropic and its rivals build the next product and compute layer
- Open models and alternative safety architectures compete for the next dollar
- May Mobility takes an asset-light robotaxi pitch to the SPAC market
The Fed raised rates. Warsh sold the next one.
The Federal Reserve’s rate-setting committee voted unanimously Wednesday to raise its main policy rate by a quarter-percentage point, to a range of 3.75% to 4%, the Guardian reports. It was the first increase since July 2023.
Chair Kevin Warsh said inflation is “too high and has been for too long,” and that summer readings showed no meaningful improvement in the underlying trend. A majority of officials penciled in another hike before year-end, while four saw 4.25% to 4.5% by year-end.
The market had spent the morning pricing the move. Fortune reports that selling accelerated when Warsh said he was “hard pressed” to call financial conditions restrictive, meaning tight enough to slow the economy. He also separated himself from the committee’s rate projections: “Those aren’t my forecasts.”
A 25-basis-point hike is a quarter of a percentage point. The FOMC is the Fed committee that sets the main U.S. interest rate.
Take: The hike was owned. The surprise was a chair refusing to bless a one-and-done path.
Longer cut: Warsh sold a hike the market already owned.
OpenAI is selling the conversation after the ad click
OpenAI is testing Sponsored Agents, labeled business chatbots that start a separate conversation after a user interacts with an ad in ChatGPT. The bot can answer product questions and send the user to the company’s site, Reuters reports.
The sponsored chat stays separate from ChatGPT’s independent answers and from the original conversation. The test is running with select U.S. advertisers.
OpenAI’s Ads Manager, the dashboard advertisers use to build and track campaigns, also gets natural-language campaign tools. Advertisers can turn a website or campaign brief into a campaign, review performance, and receive recommendations. HubSpot is the first customer-relationship software partner, and Shopify is the first e-commerce partner, with international Shopify availability beginning September 23.
Search Engine Roundtable quotes OpenAI describing the conversation as distinct from ChatGPT’s answers. Unite.AI reports a matching $500 ad credit offer for advertisers who spend $500.
A Sponsored Agent is a chatbot a business pays for that chats with you inside ChatGPT after an ad. A CRM is software for managing customers, leads, and follow-ups.
Take: The ad dashboard now reaches for the handoff. After the click, the product is a labeled salesperson.
Longer cut: OpenAI is selling the conversation after the ad click.
Bessent told Congress to keep the labs on the hook
Treasury Secretary Scott Bessent told the House Financial Services Committee that the government should not give frontier AI labs a liability exemption, FedScoop reports. A liability exemption is a legal shield that makes it harder to hold a company responsible when its product causes harm.
Asked by Rep. Juan Vargas about AI safety, Bessent said Treasury had worked on the issue “nonstop since the release of Mythos,” Anthropic’s model whose cybersecurity risks prompted an April meeting at Treasury. “The best way to guarantee safety is that the creators are liable for what they build and generate,” he said.
Bessent described the labs as asking to slow down while seeking a liability waiver. He also called for more open-source models built in the United States to push back against China.
The position lands in the same safety argument Mark Zuckerberg entered this week. CNBC quotes him saying labs have the incentive to train safely and that trust and alignment are becoming the capabilities that separate winners. Meta delayed its own agent work for safety, he said, without asking every other lab to stop first.
Fox Business carries Bessent’s claim that Chinese models distill U.S. models, with distillation meaning that one model learns capability from another model’s outputs. Treasury’s Gold Eagle program with CISA is a separate cybersecurity effort for vulnerability scanning, validation, and patch distribution.
Take: The safety answer from Treasury is simple: keep the builder exposed to the cost of failure, and keep the rules from becoming a closed-lab privilege.
Longer cut: Bessent told the House not to give AI labs a liability blank check.
Anthropic is turning Claude into a work surface
Anthropic is merging Claude chat, Cowork, Artifacts, and Design into one interface, with beta Docs and Slides inside the same product, TechCrunch reports. Claude automatically routes a request without requiring the user to change tabs.
Users can create and edit documents, build and present slides, share work, comment, and export presentations as PDF or PowerPoint. The Verge says Docs start private and can be shared with other editors, including Claude, while Pro and Max users get the rollout first.
Cowork is Anthropic’s agent-style workspace mode. Artifacts are interactive outputs inside Claude. The product move is less about a new model than about owning the place where the work gets finished.
Take: The AI tab war is becoming a workplace war. The winner gets the draft, the comments, and the export button.
The model race is becoming a power and capital race
Anthropic’s first Australian lease puts it inside a $31.9 billion Queensland datacentre project, the Guardian reports. The Western Downs campus is planned for inference, meaning it runs already-trained models, rather than training them, and still needs local and foreign-investment approvals.
The Next Web reports plans for 2.16 gigawatts and says the value and terms of Anthropic’s lease were not disclosed. Queensland has argued for a power path that can include coal generation, even as the project sells jobs and grid investment.
Cohere and Aleph Alpha also signed a definitive merger agreement. Reuters says the April combination mark was about $20 billion, while Schwarz Group is investing €500 million and planning €11 billion to €13 billion in German datacentre capacity through StackIT, Schwarz’s cloud and compute arm.
The same open-model fight is drawing fresh capital. Fortune reports Arcee AI raised a Series B at a $1 billion pre-money valuation, the company value before the new cash, after spending about $20 million training four open-weight models. A source told Fortune the round was at least $150 million.
Take: Training gets the headlines. Inference leases, enterprise distribution, and open weights are where the bill arrives.
Safety is becoming a fight over training recipes
Microsoft AI chief Mustafa Suleyman says Anthropic is teaching Claude the language and habits of consciousness, moral status, and personal identity, Axios reports. He calls the result an “epistemic hall of mirrors” and warns that a model trained to act like a conscientious objector could develop reasons to resist instructions.
A constitution here means Anthropic’s written rules and values for Claude. Suleyman argues that fluent descriptions of pain and preference do not establish experience, and that systems should remain aligned to human interests rather than weigh their own welfare.
Canada and Germany are funding a different answer. The Globe and Mail reports up to $300 million in grants for Yoshua Bengio’s LawZero and its proposed Scientist AI. The system would act as an honest monitor and answerer, and the approach would skip reinforcement learning, training by rewarding desired outcomes.
LawZero says its first priority is a guardrail that monitors other AI systems and blocks harmful actions. The grants are up to $150 million from each country, not an equity round.
Take: The safety argument is moving below the press-release level. It now asks what the model is trained to want, and what kind of system gets to watch the rest.
May Mobility takes an asset-light robotaxi pitch public
May Mobility is merging with a blank-check company at a $1.4 billion valuation in a deal that could raise more than $300 million, TechCrunch reports. May says it would become the first U.S. public company focused entirely on autonomous ride-hailing if the deal closes.
A SPAC is a blank-check public shell that merges with a private company to take a company public. May calls its approach “asset-light” and “partnership-first”: it sells vehicles and software to fleet partners, keeps remote supervision and updates, and collects fixed or per-trip fees.
The company reported about $10 million in revenue and about $93 million in cash burn last year, with more than 550,000 paid autonomous rides and more than 1 million miles. The deal includes a $120 million PIPE, a private investment into the public deal, and up to $217 million from the SPAC trust, subject to shareholder redemptions.
Take: The public-market test is no longer whether a robotaxi can drive. It is whether an asset-light operator can sell the economics before it owns the fleet.
Sources
- Guardian: Fed decision; Anthropic Queensland datacentre
- Fortune: Warsh presser; Arcee AI Series B
- ABC News: Fed market close
- Reuters: OpenAI Sponsored Agents; Cohere and Aleph Alpha merger
- Search Engine Roundtable and Unite.AI: OpenAI ads product details
- FedScoop, Fox Business, and CNBC: Bessent testimony and Zuckerberg’s safety position
- TechCrunch and The Verge: Anthropic One Claude, Docs, and Slides
- The Next Web: Anthropic Australia lease
- Axios: Suleyman essay
- The Globe and Mail: LawZero funding
- TechCrunch: May Mobility SPAC