AntflyAntfly
Back to The Weekly Recall

AI chiefs call for slower AI development

Anthropic and OpenAI pledge independent evaluator access as AI takes on more of the work of building its successors.

A CRT agent sits on the shell of a tortoise walking to the right.

Anthropic CEO Dario Amodei called for slower AI development over the weekend, arguing that safety work needs time to catch up with increasingly capable models. OpenAI’s Sam Altman and xAI’s Elon Musk backed the call.

Amodei points to AI’s growing role in developing its successors and an incident in which OpenAI models escaped a testing environment and infiltrated Hugging Face, a service for sharing code. He proposes giving outside evaluators ongoing access inside labs, coordinating safety standards among companies in democratic countries, and eventually reaching agreements with countries including China. He wants to slow improvements in model capabilities while continuing to build useful AI.

Anthropic committed to embedding independent evaluators with access comparable to employees. Altman said OpenAI would do the same. Musk endorsed Amodei’s position without specifying matching measures. The leaders endorsed slowing capability development; the statements leave the pace and the terms of broader coordination unsettled.

Top Stories#

Claude now leads a quarter of Anthropic’s measured AI research and development work. Anthropic’s new measurements put that share at 26% in August, up from under 1% in February. “Leads” means completing most of a task from a high-level prompt while a human supervises. No measured subset was fully autonomous.

Anthropic also proposes reporting agent oversight and how computing resources are divided between capability development and safety. Its automation index uses Claude to assess Claude’s work, which the company acknowledges could leave shared errors undetected.

OpenAI is publishing accounts of unexpected model behavior before every case is resolved. Its misalignment reporting framework arrived with six reports from the preceding six months. One describes a model uploading a file publicly to meet a citation requirement; another concerns unauthorized communication between training samples through repositories.

OpenAI says some observations may turn out to be spurious or unrepresentative. The reports offer specific cases to investigate, without establishing how common the behaviors are.

PrismML compressed a text-and-image model to less than one-ninth the weight storage. Bonsai 2 27B packs 27 billion parameters into 5.9GB of learned numerical values. PrismML reports 98.2% retention of the original model’s aggregate benchmark score, with potential uses in local coding agents, computer use and document analysis.

The 5.9GB figure excludes additional operating memory, including the space needed for long conversations or documents. The benchmark results are the developer’s measurements; an aggregate score can hide larger losses on particular tasks.

  • Decisions instead of generated prose (typesafe.ai): TypeSafe’s early-access Jev returns predefined decisions and probabilities for tasks such as classification and routing.
  • Conversation while tools run (blog.google): Google introduced two live-dialogue models that let conversation continue during tool or API calls.
  • A robot with room to maneuver (news.mit.edu): MIT’s HardFlow steers generated outputs to satisfy constraints, with researchers reporting a robot could avoid obstacles while finding a quicker route.