A tech blog worth bookmarking: SWE-bench MM v2 flaws, 232x kernel speedup, prompt caching
dejavucoder · x · 2026-11-17
A recommended bookmark: sankalp's blog, praised by people at Prime Intellect, Epoch AI, OpenAI and GDM. Highlights:
- "SWE-bench Multimodal v2 is not multimodal" — a critique of a frontier benchmark.
- Auto-research with Codex: a 232x faster-than-baseline kernel on GPU Mode's qrv2 problem, #1 on Hacker News.
- How prompt caching works — paged attention, prefix caching and practical tips (75k unique readers).
- A Guide to Claude Code 2.0 plus a two-week experience post (HN top 5, 200k unique readers).
- Also: agent automation usage patterns and the Tracking Singularity incidents site.
Great for developers into agent evals, coding workflows and inference optimization.
More from coding & agent
- 知识图谱入门:用实体-关系网络为 AI 提供结构化上下文 — robertonappe · 2026-11-17
- "Summarize what you did in simple terms": the one AI coding prompt devs swear by — js_craft_hq · 2026-11-17
- 39 open-weight models, 11 days, 8,199 runs: 75.7% of agent failures came from runs that did call tools — Green_Competition_21 · 2026-11-17
- Gemini 4 Carbon vibe check: boldest UI designs, least reliable finishes — every · 2026-11-17
- gputaskman: an open-source htop for your GPU that breaks down VRAM by process and LLM KV-cache — alex_verem · 2026-11-17
- Matt Pocock's /constrain Idea: Cage Agents With Brutal Lint Rules — mattpocockuk · 2026-11-17