Thursday, August 27, 2026

RLVR with task expertise beats human on text-to-SQL

A model trained with RLVR on Tinker achieved state-of-the-art on BIRD, scoring above the human mark of 92.96%. The work, a guest post by UIUC and Bridgewater researchers, folds expert judgment into every part of RLVR.

OpenAI rolls out GPT-5.6 Sol to all paid ChatGPT users

GPT-5.6 Sol now powers all chats for paid users, including Instant, unifying the experience. In high-stakes factuality tests across finance, medicine, and law, it produced 68% fewer factual errors than GPT-5.5 Instant.

Kimi K3: 2.8T-parameter open-source model with 1M context

Anthropic brings Claude Mythos 5 to Claude Security scans

Claude Security scans now run on Claude Mythos 5, available in public beta for all Claude Enterprise customers. Anthropic also launched a $35M fund to secure open-source software and plans to expand its Cyber Verification Program.

Accelerated Understanding launches neural operator AI model

Startup co-founded by Caltech professor Anima Anandkumar and Benedikt Jenik launches a 4D AI model built on neural operators instead of transformers, processing up to 5 trillion data points in a single prompt. Scaled to 1 trillion parameters in pre-training, it targets physics simulation for energy, chip design, robotics, and weather.