Key points
- Anthropic lets METR's testers inside
- NYC's comptroller wants Anthropic's IPO paused
- Some Anthropic investors weigh a lower valuation
- Why Anthropic's copycat crackdown may cost revenue
Anthropic has given six researchers from the nonprofit METR access to internal systems to assess its AI risks, the company said Thursday. Their access is comparable to that of Anthropic's own risk assessment teams.
The announcement came a day after New York City Comptroller Mark Levine urged Anthropic to pause its planned IPO.
METR can publish without Anthropic's approval
METR's researchers have Anthropic hardware, office space, and access to internal tools. Employees are encouraged to share information with them "just as they would with a fellow employee working on alignment risk," Anthropic said.
METR will publish findings on Anthropic's risks, incidents, and practices. Anthropic can require redactions in narrowly defined categories, including customer data, attorney-client privileged information, and details of its model architectures.
"Anthropic can't veto publication," the company wrote. METR can publicly challenge whether a redaction was justified.
Anthropic compared the setup to bank supervision, where regulators assign resident examiners to work inside the largest banks, though the company said the analogy isn't exact. METR isn't a regulator, and Anthropic says the embedding is voluntary.
METR is so far the only evaluator in the top of three tiers. SecureBio is in the first tier, where Anthropic approves what an evaluator publishes. Anthropic is also working with Accenture (ACN) and Faculty, and is in early talks with others, including Hugging Face.
The arrangement follows a September 12 essay by CEO Dario Amodei, "We Must Pace the Frontier," and a White House accord in which AI companies pledged to work with independent auditors or evaluators.
New York City's comptroller wants the IPO paused
Levine oversees the New York City pension funds, which have more than $325 billion in assets. In an October 14 letter to Amodei, he said Anthropic should address the safety, governance, and development-pace concerns the company has raised before going public. He called the White House accord Anthropic signed "insufficient to meet the moment."
Levine wrote that "prudence should not only be shown with respect to the pace of the frontier, it should be also shown with respect to the timing of an IPO." He asked Anthropic to disclose how it will audit and control the risks, what those controls cover, and on what timeline, and he requested a written response.
The city's funds have about $7 million of exposure to Anthropic, the comptroller's office told AFP. That would grow if Anthropic's shares join the Russell 3000, an index the city tracks for its own investments. Anthropic's METR announcement didn't mention the letter.
Anthropic presses ahead as other IPOs slow
Rising interest rates and the Iran war are threatening to stall other US listings. Smart-ring maker Oura and Bamboo Insurance have postponed their IPOs, while SB Energy and Nscale have filed publicly but haven't started formal marketing, according to Bloomberg. Anthropic's own formal marketing could begin as soon as the week of Nov. 9, which would put a trading debut before Thanksgiving within reach.
Stopping copycats may be slowing revenue growth
Anthropic told prospective investors on Wednesday that third-quarter revenue rose 45% from the second quarter to $16.8 billion. Growth slowed from the second quarter, when revenue jumped to $11.6 billion from $4.7 billion in the first.
Some current and prospective investors who attended the investor day discussed a valuation of around $1.5 trillion for the November IPO, down from the roughly $2 trillion previously expected, according to Reuters. The adjusted profit eased some concerns about the cost of AI, but some investors were disappointed that growth slowed. Anthropic didn't immediately comment.
The company also told prospective investors this week that 5% to 7% of its revenue comes from distillation and that this revenue is expected to decline further, according to The Information. Distillation uses one model's outputs to train another. Anthropic's crackdown targets developers using Claude's answers to train competing models.
On October 12, Anthropic said new defenses had cut the data distillers harvest from Claude by about 90% to 99%. The company said it has "turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers."
That crackdown could be hurting Anthropic's reported revenue growth, Alistair Barr wrote in Business Insider, though "it's unclear how much of the recent slowdown can be attributed to the crackdown." Anthropic didn't respond to his question by Thursday morning.
