Daily Brief
2026-08-19 · Evening · 8 items
-
GLM-5.3 API launches: scores 60 on AA index, ties for first among open-source models
Lowest per-task cost among frontier flagships, pricing unchanged from 5.2, weights open-sourced next Friday.公众号：智谱（GLM）
-
Claude designs protein binders: 26.7% hit rate, roughly double the industry norm
Mythos Preview hit 26.7% across 15 targets, versus the typical 10–15% industry range.Anthropic：Research（发表成果 · 网页）
-
Sentence Transformers v6.0 adds multi-vector retrieval support
v6.0 introduces a new multi-vector model type that loads ColBERT and PyLate weights directly.Hugging Face：Blog（RSS）
-
More agent memory isn't always better: 8-model benchmark shows dosage must match capability
Strong models gain most from full playbooks; weaker ones do better on a curated subset.Hugging Face：Blog（RSS）
-
Claude can now send Gmail and manage Google Drive files
Reply to threads and handle Drive files inside Claude, available on all paid plans.X：Claude (@claudeai)
-
Claude Tag serves as Anthropic's first responder for CI incidents
Writes the initial incident report for nearly every CI outage, usually within 15 minutes.Claude：Blog（网页）
-
StartupBench: top models solve only ~30% of real startup tasks
Drawn from shipped AI startup workflows, the best model clears just 30%.HuggingFace Daily Papers（社区热门论文）
-
Mojo language goes open source: compiler and toolchain fully released
Version 1.0 shipped last week under Apache 2.0 with the compiler and toolchain, though compiler pull requests stay closed until year-end.Modular Blog