TLDRocket
26 July 2026
Kuaishou's KwaiKAT team is pushing coding AI beyond isolated snippets with KAT-Coder-V2.5, trained on 100,000+ verifiable repository environments where the model learns to operate inside real, executable codebases. The breakthrough came through AutoBuilder, which raised environment construction success from 16.5% to 57.2%, and infrastructure improvements that cut sandbox-induced training failures below 2%. The model now tops PinchBench at 94.9 and ranks second on SWE-Bench Pro at 65.2, with an open-weight variant available on Hugging Face. This reflects a broader shift toward training AI systems in realistic, messy conditions rather than curated datasets—a pattern echoed by Induction Labs' Photon-1, a 106-billion-parameter vision model trained on 18 years of unlabeled desktop screen recordings that learns next-latent-token prediction without action labels. Photon-1 costs less to run than Gemini 3.1 Flash-Lite and demonstrates improved performance on downstream tasks like checkers and physics simulation. Meanwhile, Meta's FAIRChem v2 and UMA potential tackle another domain-specific problem: scientists building separate models for molecules, catalysts, materials, and dynamics can now use a single pretrained universal potential across those workflows with GPU acceleration. The common thread is operational realism—systems trained on or deployed in the actual environments where they'll be used, rather than laboratory abstractions. This month's tech layoffs tell a darker story. Monday.com cut 600 employees (20% of staff) citing AI-driven product transformation, joining at least 20 other major firms. U.S. tech companies have eliminated nearly 140,000 jobs since January 2026, with Amazon, Oracle, Meta, and Microsoft accounting for roughly half. Yet companies announcing AI-related cuts have underperformed the Nasdaq by nearly 10% in the month following their announcements—suggesting investors remain skeptical of the claimed efficiency gains.
Read the full briefing →