Daily AI News | Anthropic CEO Dario Amodei calls for slowing the frontier of AI development
# Daily AI News | Anthropic CEO Dario Amodei calls for slowing the frontier of AI development
## Lead Story
### Anthropic CEO Dario Amodei calls for slowing the frontier of AI development
Anthropic CEO Dario Amodei published “We must pace the frontier,” arguing that leading labs should slow progress at the AI frontier to create time for safety measures, international coordination, and social adaptation. TechCrunch subsequently reported the proposal; it is a public policy position, not an industry agreement already in force.
**AI take:** The proposal moves AI-safety debate toward development pace. The next test is whether labs make verifiable commitments and regulators respond.
## More News
### Codex Astra fixes over-triggering of older Skills and completes a usage reset
Codex product lead Tibo said the team fixed a quality issue in which Astra over-triggered older Skills and interfered with the model’s own checks, and issued a usage reset for affected users. He later confirmed that the reset had fully propagated.
**AI take:** Affected users can rerun abnormal or failed tasks. Older Skills should still be retested to confirm stable triggering.
### Chinese world model Motus 2 launches with a focus on self-improving dexterous robotics
New AI Era reported the release of the Motus 2 world model for dexterous robot manipulation, using feedback during training and execution to improve its policies. The report frames the change as a move from one-shot action generation toward continued self-improvement, although no independent official technical page verifying all metrics was available in the window.
**AI take:** Embodied-AI teams should watch the training loop. Cross-hardware and cross-task reproduction matters more than a single demo.
### Zhejiang University open-sources a pluggable agent-evaluation base with CLI and Skills
QbitAI reported that a Zhejiang University team open-sourced a pluggable agent-evaluation foundation with CLI and Skills integration. It is designed to combine different agents, tasks, and evaluation modules with less repeated scaffolding work.
**AI take:** Shared evaluation infrastructure can standardize experiments. Teams should verify task coverage, isolation, and reproducibility before adoption.
### Qoder CLI v1.1.50 adds custom-model support for free and paid individual users
Qoder’s official announcement says that, starting with CLI v1.1.50, free and paid individual users in both the global and China editions can configure custom models. Developers can therefore choose model endpoints based on cost, privacy requirements, or task performance.
**AI take:** Custom models ease integration with existing stacks. Teams should verify authentication, data flow, and parameter compatibility.
### Qwen3.8-27B becomes available on Cerebras for high-speed hosted inference
Qwen’s official account confirmed that Qwen3.8-27B is now running on Cerebras and available to try, adding a high-speed hosted-inference option. This is a new platform-availability stage rather than the model’s original release.
**AI take:** The endpoint enables latency and throughput comparisons. Check pricing, limits, and API compatibility before migrating workloads.
## Rumors
### Tibo teases Images 2.5, GPT-Live-1, Agents API, and more for this week
Codex product lead Tibo said a group of Astra-powered products may ship this week, naming Images 2.5, GPT-Live-1, Agents API, Data Agent, and ChatGPT for Financial Services. No individual launch pages, specifications, pricing, or complete availability details appeared in the window, so the names and timing remain a teaser.
**AI take:** The list is a near-term watchlist, not proof of launch. Wait for official documentation and account or regional availability.
