Microsoft Research's Webwright replaces step-by-step clicking with a terminal: the model writes and runs a re-runnable Python/Playwright script that completes the whole task in one file. On long-horizon tasks the same model jumps from 33.5% to 60.1% success, and each solve leaves a reusable skill behind. Rethinks what a browser agent should persist: code, not clicks.
Change one number in a grounded RAG answer (10,000 for 1,000) and every mainstream hallucination detector scores near chance: AUROC 0.51-0.59 for embedding similarity, entailment models and LLM judges, with MiniCheck barely better at 0.75. A warning for any production pipeline that trusts faithfulness checks on numeric answers.
Sentence Transformers now ships MultiVectorEncoder, letting you load any PyLate, ColBERT, or colpali-engine checkpoint through one familiar API. Late interaction keeps one vector per token and MaxSim scoring, trading index size for retrieval quality, including state-of-the-art OCR-free visual document search.
Grab reports AI agents now handle much of its analytics grunt work, dropping analysts' mechanical tasks from 44% to 30% between February and June and cutting time-to-answer for business questions. The setup pairs agent autonomy with certified data, context management, and human oversight, so self-service analytics absorbs metric, data, and SQL requests without analyst intervention.
OpenSearch 3.8 ships Base64 vector ingestion that cuts network payloads by 74% and boosts bulk throughput up to 4.16x, plus a redesigned radial-search traversal delivering 2.1x query throughput with higher recall. MCP integration now spans all agent types, and gRPC streaming inference lowers token latency.
Long-context sessions, agents, and RAG churn out KV caches and embeddings that are cheaper to store than recompute, so inference is becoming a state-management problem. Cached-token discounts of 90-99% prove the economics, and flash, not HBM, is where the battle moves. That redraws the memory-storage boundary in AI infrastructure.
DuckDB v2.0 will ship four new JSON scalar functions: json_merge_patch_diff (the inverse of RFC 7396 merge patch), json_deep_merge, json_normalize, and json_strip_nulls. They let pipelines reconcile entity state between systems in a single query, no Python detour; the post benchmarks them against Python on 500k CDC events.
NIFC publishes only a current snapshot of active wildfires and overwrites it, so past acreage and containment numbers vanish. This free CC BY 4.0 archive (static CSV/JSON, no signup) keeps the last read of each UTC day for every 1,000+ acre fire in NV/UT/AZ/CO/CA, preserving the growth curves the feed never shows. Started 2026-08-09; 39 incidents and growing.
You’re caught up.
That was all 8 stories in the 18 August issue.