Daily Brief
2026-09-18 · Morning · 8 items
-
Anthropic releases 3 metrics to track frontier AI development pace
First comparable data across R&D, agent oversight, and compute stages.Anthropic：The Institute（旗舰研究长文 · 网页）
-
Anthropic uses Claude to optimize 30+ open-source biomolecular models
About 4x average speedup in under four weeks, with 2x gains even when outputs are held identical.Anthropic：Research（发表成果 · 网页）
-
Anthropic CEO Amodei unveils a 3-step plan to slow AI: testing, allied coordination, global governance
Altman and Musk jointly backed a cross-border AI governance framework, an unusual alignment that cuts through their past public split.The Verge：AI（RSS）
-
Epoch AI: ~$3B in Likely AI Servers Routed via Malaysia into China
Average unit price of $106k points to restricted AI hardware, not ordinary servers.Epoch AI：研究、数据与评测
-
Goodfire finds internal signals can detect reward hacking at scale
Across three open models and benchmarks, 50-96% of rollouts hacked rewards; probes caught cases chain-of-thought monitoring missed.Goodfire Research（网页）
-
GitHub rewrites Copilot runtime in Rust via Copilot agentsRECAP
In 14.5 weeks, agents produced 832,378 lines of production Rust shipped through 128 merged PRs.GitHub Blog
-
Jie Tang: GLM-5.3-driven Infra Agent optimizes GLM-5.3-Flash in two weeksRECAP
On a domestic accelerator, the model went from first run to full production traffic with 3.2x end-to-end throughput.X：唐杰（@jietang）
-
GLM-5.3 builds its own inference infra, tripling throughputRECAP
It went from first successful run to production-ready in under two weeks, hitting 3x baseline throughput.X：智谱 Z.ai (@Zai_org)