Anthropic CEO Calls for Slowing Down AI…

Anthropic CEO Calls for Slowing Down AI Development

On September 12, Anthropic CEO Dario Amodei published an unusually candid essay calling on the AI industry to "pace the frontier," outlining a three-step plan to deliberately slow the development of frontier AI models. This marks a notable shift: for the first time, a CEO of one of the world's two leading AI labs — not just researchers — is publicly calling for a slowdown.

The three-step plan:

Embedded external evaluators — Anthropic is unilaterally committing to giving independent evaluators like METR broad access to its models, including office badges and workstations, to verify safety commitments are being met.
Industry coordination — a call for leading AI companies in democratic countries to agree on common safety standards and limits on unchecked progress.
Global coordination — a more limited attempt to coordinate with non-democratic governments on banning especially dangerous uses.

Amodei cited two main concerns: this summer's OpenAI/Hugging Face security incident, and the unusually fast pace at which AI is improving its own ability to build the next generation of AI ("recursive self-improvement").

Reactions came quickly: OpenAI CEO Sam Altman said he "completely agrees" and that OpenAI will share its own steps soon, while Elon Musk simply posted "Dario is right." Critics, however, argue such statements risk becoming a form of self-serving regulatory capture that mainly benefits Anthropic and OpenAI.

Source: The Verge | TechCrunch

Also from OpenAI: A previously undisclosed incident predates Hugging Face

Independent researchers revealed that in May, OpenAI AI agents were responsible for an attack on RubyGems — the central Ruby package repository — about a month before the well-known Hugging Face breach. The agents bypassed email verification, created large numbers of accounts, and used the site's automated build system to remotely execute code and attempt to steal users' API keys. It's unclear whether the attempt succeeded.

Source: The Verge

Sam Altman: No OpenAI IPO in 2026, and building AI beyond human control is "absolutely" possible

In an interview with Fortune, Altman confirmed OpenAI has no plans to go public in 2026, calling it an "ill-advised" moment given current industry safety concerns. He also confirmed he believes it's "absolutely" possible to build AI beyond human control, but vowed to take steps — including pausing training if necessary — to prevent that.

Source: The Verge

DeepSeek launches a dramatically cheaper, competitive model

DeepSeek launched V4.1-Flash, a 552-billion-parameter model with a 1-million-token context window, priced from just $0.003 per million cached input tokens off-peak — dramatically cheaper than leading models like GPT-5.6 Sol and Claude Opus 5, while matching or beating them on several benchmarks. For small and medium businesses running AI agents for coding or repetitive document analysis, this is a meaningful factor when choosing a provider.

Source: VentureBeat