X

@cursor_ai

GLM 5.3 in Cursor: GLM 5.3 and GLM 5.3 Flash now available; GLM 5.3 Max is the best-scoring open-weight model on CursorBench 4.0. post (https://x.com/cursor_ai/status/2105787358557999585) · 2,579❤ · 124 RT

@karpathy

Land or Water eval: describes an eval that asks an LLM "Land or Water?" with lat/long as text 16,200 times and plots the result; models know from compressing the internet. post (https://x.com/karpathy/status/2105909609487872075) · 6,091❤ · 371 RT

Understanding LLM outputs: tips for understanding language model outputs; suggests asking LLMs to explain in ASD-STE100 controlled language. post (https://x.com/karpathy/status/2105819303471976479) · 22,971❤ · 2,449 RT

@poteto

Pepsi instead of Coke: jokes "when the restaurant only has pepsi instead of coke" with an image. post (https://x.com/poteto/status/2105733341618253878) · 992❤ · 34 RT

Anything with Grok Bot: says "you can do anything with grok bot" quoting Grok Bot's proactive suggestions launch. post (https://x.com/poteto/status/2105718847181656361) · 1,710❤ · 50 RT

@hwchase17

Durable runtime: says every agent harness needs a durable runtime; maps pi to pi-durable and deepagents to langgraph. post (https://x.com/hwchase17/status/2105791360796397954) · 167❤ · 11 RT

Model routing in the harness: outlines building a model router properly — understand tasks and models, build the router in the harness, track outcomes; lower costs with no performance hit. post (https://x.com/hwchase17/status/2105710409894547687) · 121❤ · 14 RT

Decision model season: notes more entrants including Cloudflare's open-weight clef/clef-flash; harness should swap its decision model as easily as its main model. post (https://x.com/hwchase17/status/2105713773084569666) · 84❤ · 6 RT

@rauchg

Verification-engineering: argues the future is verification-engineering — proofs, e2e tests, benchmarks, linters; some tests deterministic, some agentic; quotes open-source e2e agentic testing framework. post (https://x.com/rauchg/status/2105723481413550427) · 2,678❤ · 130 RT

SvelteKit 3: congrats on SvelteKit 3; built and deployed a tiny app in 15 seconds with fx + opus; excited for Async Svelte and Remote Functions. post (https://x.com/rauchg/status/2105837842362732965) · 537❤ · 30 RT

Microsoft AI on Vercel: says MicrosoftAI trains excellent models; excited to bring them to Vercel day zero (quoting MAI-Voice and MAI-Transcribe on AI Gateway). post (https://x.com/rauchg/status/2105702513345511514) · 247❤ · 8 RT

@HamelHusain

Bot replies: asks if any humans are here; says 95% of replies are bots. post (https://x.com/HamelHusain/status/2105697385896972559) · 268❤ · 3 RT

Eval cost tradeoff: Q&A — do not build automated evaluators for every failure mode; weigh code vs LLM-judge costs before building. post (https://x.com/HamelHusain/status/2105674106683134290) · 40❤ · 8 RT

Silent

@martinfowler, @swyx, @simonw: no original posts in window

Gaps

@xAI: API not authorized to read this account's posts (skipped)

GitHub trending

DietrichGebert/ponytail (https://github.com/DietrichGebert/ponytail): makes your AI agent think like the laziest senior dev; the best code is the code you never wrote (JavaScript, +1,194★ today)
mattpocock/skills (https://github.com/mattpocock/skills): skills for real engineers from his .agents directory (Shell, +883★ today)
NVIDIA/OpenShell (https://github.com/NVIDIA/OpenShell): safe, private runtime for autonomous AI agents (Rust, +2,456★ today)
firebase/firebase-ios-sdk (https://github.com/firebase/firebase-ios-sdk): Firebase SDK for Apple App Development (C++, +112★ today)
mvschwarz/openrig (https://github.com/mvschwarz/openrig): multi-agent harness for Claude Code, Codex and Pi as persistent teams with roles and shared context (TypeScript, +642★ today)