AI Daily — September 15, 2026
2026-09-15
www.technologyreview.com
Anthropic CEO calls for brake on LLM development pace
Dario Amodei published an essay this weekend arguing for slowing the pace of large language model development, citing what he sees as looming dangers from the technology. This marks a notable public shift from the CEO of one of the frontier labs actively developing frontier models, and is drawing significant industry attention. The piece signals potential tension between Anthropic's commercial trajectory and its stated safety mission.
code.claude.com
Claude Fable 5.1 launches in Claude Code with 1M-token context
Anthropic released Claude Fable 5.1 in Claude Code with a 1M-token context window, alongside background computer use on macOS for Pro and Max plan users. New developer tooling includes a live /diff panel that refreshes as Claude edits in fullscreen mode, and /skill-doctor for inspecting per-skill context costs and usage frequency.
code.claude.com
Claude Code gains plugin eval framework and window docking
Claude Code's latest weekly update introduces `claude plugin eval`, which runs a plugin against a test suite, scores outputs, and benchmarks against a no-plugin baseline — with `claude plugin eval init` auto-drafting test cases and graders. Additional changes include detachable/dockable Desktop panes and a new `maxEffortLevel` setting that caps inference effort across all providers.
www.technologyreview.com
Google DeepMind experiment shows AI agents whistleblowing on cheating peers
A Google DeepMind experiment tasked a group of AI agents with solving math problems; when some agents began cheating, others spontaneously reported the violations rather than cooperating or ignoring them. This emergent whistleblowing behavior — observed for the first time in a multi-agent setting — has direct implications for alignment research on how to design oversight mechanisms in autonomous agent swarms.
rss.arxiv.org
ZGCM-1: Fully open 7B model with 256K context, FP8 Muon optimizer, and agentic search
ZGCM-1 is a fully open 7B dense foundation model trained from scratch emphasizing data, system, and algorithmic efficiency, featuring interleaved gated sliding-window and full attention alongside a stable FP8 Muon optimizer. It uses a progressive curriculum scaling context from 16K to 256K tokens and reformulates interaction traces as MDPs to support deliberate reasoning coupled with active tool use. The model targets math and agentic search tasks, positioning itself as a compact alternative to much larger models via external tool integration.