New briefs daily around 7 AM Eastern
LangChain tested TypeSafe AI's Jev, a typed classifier rather than a text-generating LLM, as an agent-eval judge. It matched a human reviewer on every one of 500 repeated decisions while costing a fraction of a cent per call, far less than Claude's LLM-judge cost.
LangChain tested TypeSafe AI's Jev, a typed classifier rather than a text-generating LLM, as an agent-eval judge. It matched a human reviewer on every one of 500 repeated decisions while costing a fraction of a cent per call, far less than Claude's LLM-judge cost.
Filip Makraduli's FlashNorm folds a transformer's norm layer into its projection weights and overlaps the remaining divide on a separate CUDA stream, cutting norm-plus-projection cost by a third with no retraining needed.
A dated brief from the vibe-coding frontier. Today’s Intel.
A rare vLLM bug corrupted about one in a thousand prompts with no error. The cause: a scheduler race let decode run before prefill for Jamba's Mamba layers, computing a fresh request over a stale prior state.
Ben Swerdlow ran 171 real-time StarCraft matches between Codex, Claude, and Grok models, surfacing concrete failure modes in continuous, multi-unit agent control that single-turn benchmarks miss.
PlanetScale released Tin, a GA Postgres extension with boolean, phrase, fuzzy, and BM25-ranked search that keeps correct transactional visibility, cutting a common reason teams bolt on Elasticsearch.
Hacktron's fuller HEIF Heist disclosure shows the libheif bugs behind Tuesday's OpenAI account takeover also reach Slack, Meta, GitHub Enterprise, Rails, and Next.js through indirect dependencies.