以 Shiki 为金标准评测代码高亮，热们语言正确率引质疑
DoctorGester · x · 2026-09-09
- shuding explained his benchmark methodology: comparing span types against Shiki (real AST-based, near gold standard), reporting weighted correctness across GitHub's top 25 popular languages.
- DoctorGester pushed back, calling results misleading: dozens of incorrect highlights are easy to find, with roughly 50% of threejs.min.js marked wrong and confusing spans in other files.
所属事件:Vercel 开源 gpu-lexer：27.5KB 小模型用 WebGPU 做语法高亮(9 条相关)→
「模型」频道最新
- MissionReport 基准：540 份政府任务报告仅 3 份全达标 — Exp_Mark · 2026-10-27
- 隐私场景实测：本地 Qwen 3.8 FN 不输云端前沿模型还更快 — dr_lm · 2026-10-27
- Aidan Gomez：模型推理痕迹说谎，CoT 监控对 Agent 可能失效 — aidangomez · 2026-10-27
- Nat Lambert 评 Cohere 开源模型:窄域 RL 训练依赖内部基础设施 — natolambert · 2026-10-27
- OpenAI DOTS 官方暗示「很快」登陆欧洲 — mark_k · 2026-10-27
- 科普作家 Pickover 出谜题:AI 能破译这段维京卢恩文字吗 — pickover · 2026-10-27