AI Daily, Sept 9: The Builders Say There’s No Plan — and the Money Moves Down the Stack

The reckoning, on the record

• An Anthropic researcher puts a number on it. Pretraining researcher Jacob Coxon resigned and said the labs are “gambling with our lives”; a current Anthropic alignment researcher, Evan Hubinger, replied that he “personally” puts catastrophe at “>10% within the next decade” and that Anthropic “do not yet have a plan to solve alignment for superintelligence.” Not a coordinated walkout — one resignation, one colleague’s reply — but the internal counterpart to Pachocki’s institutional warning two days earlier. Our take →

Value moved to the last mile

• Google rents a deployment army for Gemini Enterprise. Google Cloud and Accenture formed a new business group staffed by 1,000 forward-deployed engineers to get Gemini Enterprise into production on-site — a bet that enterprise AI is won at deployment, not the model. (Announced Tuesday; figures are Accenture’s.) Our take →
• Instacart keeps the checkout. Its new Clementine assistant — plus the first live retailers on its white-label Cart Assistant — make Instacart the settlement layer for agentic grocery: the chatbot builds the cart, Instacart keeps the transaction, the payment credential, and the ads. Our take →

The app layer got repriced

• Harvey → $550M at $15.5B. The legal-AI company raised at a ~41% step-up over March — not to build a frontier model, but to own a post-trained checkpoint on open weights as a hedge while still renting the frontier. The wrapper thesis, maturing into optionality. Our take →
• Salesforce eyes the customer-understanding layer. Salesforce is reportedly in talks to buy AI customer-research startup Listen Labs for ~$2B (Business Insider; unconfirmed) — a bet that the scarce input for agents is continuous understanding of what customers want. Our take →
• And the GTM layer. Clay reportedly raised $115M at $7.1B, up from $3.1B a year ago — sales-and-data tooling repriced as an AI-native category. (The ~$7B pre-money had been reported in late August.)

The spend data caught up

• Ramp’s AI Index: Anthropic leads OpenAI in business. In Ramp’s card-and-bill-pay base, Anthropic is used by 43.8% of businesses vs OpenAI’s 39.8%; the effective price per million tokens fell ~41% from March; and frontier-model token share slipped to ~45% from ~53% as buyers routed volume to cheaper models. The barbell, visible in real spend. (Ramp’s base, a spend proxy — not the whole economy.) Our take →

Consumer agents shipped ahead of control

• Meta launches Muse. A cross-app personal agent with inbox, calendar, and payment authority — and, in Meta’s own documentation, a conceded susceptibility to prompt injection. Capability shipped to consumers ahead of the controls, with a bounty program standing in for the missing wall. Our take →

By the numbers

>10% Hubinger’s personal odds of catastrophe  ·  $550M / $15.5B Harvey  ·  1,000 Accenture forward-deployed engineers  ·  43.8% vs 39.8% Anthropic vs OpenAI (Ramp base)  ·  −41% price per million tokens since March  ·  ~$2B reported Salesforce–Listen Labs  ·  0 Anthropic S-1s filed (still watching)

The bottom line. The frontier’s own builders said out loud that they don’t have a plan — and the market answered by moving its money to where it thinks the value now lives: not the model, but the last mile that deploys it, the rails that distribute it, and the layer that understands the customer.