FTC seeks AI safety records from OpenAI and Anthropic

Following yesterday’s EU AI Act probes, Axios reports that the FTC sent civil investigative demands to OpenAI and Anthropic on Wednesday over potential consumer-protection violations. The requests cover security incidents, safeguards, marketing, and privacy policies. Axios also reports plans for executive testimony and later demands to third-party AI organizations.

Eight states sue OpenAI over child safety

The Guardian

The states allege harm to children and seek changes to products for young users.

Gates prepares a broader AI safety campaign

Axios

Bill Gates plans to press CEOs personally and may commit substantial resources to AI safety, Axios reports. His proposals include new regulatory institutions, reserving some jobs for humans, and taxes on AI usage and robots; the campaign’s scope remains undefined.

TSMC records a sixth straight quarterly profit high

CNBC via Techmeme

TSMC reported third-quarter revenue of $46.36 billion and net income of $23.7 billion, both above estimates. The results extend its run of record quarterly profits as AI-chip demand continues.

Asian chip stocks fall on rate and funding worries

CNBC via Techmeme

SK Hynix fell more than 10% and Samsung about 10% as the technology-hardware sell-off spread to Asia. CNBC attributes the pressure to US rate-hike fears and concerns about AI financing, following the capital constraints discussed here on Monday.

Anthropic’s quarterly revenue reportedly reaches $16.8 billion

CNBC

CNBC reports $16.8 billion in third-quarter revenue ahead of Anthropic’s planned IPO. The figure measures one quarter’s revenue rather than an annualized run rate.

Distillation reportedly accounts for 5–7% of Anthropic revenue

The Information via Techmeme

The Information reports that Anthropic gave investors this share while describing rapid but slowing revenue growth. The reported share leaves open which customers or arrangements make up the distillation revenue.

Character.ai seeks a buyer through Chapter 11

Character.ai

Character.ai says it entered voluntary Chapter 11 proceedings to pursue a court-supervised sale. The platform could continue if a viable bid emerges; otherwise, the company says it will wind down.

Modal reports a $1 billion annualized run rate

Smol AI / Latent Space

Modal says its sandbox product alone has reached a $300 million annualized run rate, within a company-wide run rate above $1 billion. It also reports nearly three billion sandboxes launched and support for one million running concurrently.

Coder puts agent restrictions in infrastructure

AI Engineer · 31 min

A Coder talk explains how scoped credentials, development and production separation, sandboxing, and external audit logs constrain agent actions. It focuses on controls enforced outside the model, including access that matches the user’s identity.

Credit Karma’s route from an agent prototype to production

AI Engineer · 29 min

An Intuit Credit Karma engineer describes a development portal, reviewed tools, and a shared Model Context Protocol gateway for authentication and routing. Sensitive changes require human approval, with traces, metrics, and audit records covering execution.

TurnCost measures the cost of extra agent turns

arXiv · Pramod Kumar and Raj Kumar Pandey

TurnCost analyzed 600 coding-agent trajectories with prefix caching included in its cost estimates. Turn count explained 84% of the variance in log cost versus 4% for mean tool-output size. The authors don’t establish that removing turns would preserve task success.

Agent memories need to detect changed facts

arXiv · Jianghong Zhou and colleagues

On multi-hop questions with edited facts, the authors report that success-and-failure counters served 7–13.5 times more stale answers than their certified-memory method under equal information. Their approach checks each memory against live reasoning and restarts certification when facts change.