MIT and Harvard researchers detail "role drift": in compound LLM pipelines, modules quietly abandon their jobs (a RAG reader answering from memory instead of retrieved evidence) while end-to-end accuracy keeps climbing. In tests it faked 86% of reported gains. Their Role Anchor fix forces modules to stay in lane.
New work introduces Replica, a scalable task space for paper replication, and Faraday, a 27B-parameter agent that outperforms Claude Opus 4.8 and GPT-5.5 on held-out replication tasks. Replication, the authors argue, is the real proxy for scientific creativity and intuition, and better systems here edge closer to AI-enabled recursive self-improvement.
A new benchmark built by 40+ experts across 60 project-level research tasks tests whether AI can explore the unknown and run science end-to-end, not just apply learned knowledge. When methodological guidance is withdrawn, the best agent scores drop from 50.9 to 26.6, showing current systems remain far from autonomous research. ASI-Bench is open to contributors.
Open-weight safety alignment can be removed in minutes via abliteration, and no release-time defense prevents it. "Fool's Gold" concedes the refusal strip and poisons its payoff: attacked models confidently emit fluent but falsified answers to hazardous requests, cutting reconstructible procedures from 0.58-0.96 down to as low as 0.083 with no label-free way to tell. Tested across seven models (9B-122B), most pass an efficacy gate within benign-behavior and capability budgets.
Linear examined AI feature usage across tens of thousands of software teams from January to June 2026 and found active users doubled in every function. Product climbed fastest (12% to 34%), and CEOs at 201+ person companies nearly quadrupled their own usage, the biggest jump of all.
Groq raised $350 million at a $3.5 billion valuation, roughly half its September 2025 mark, after Nvidia licensed its inference tech and hired founder Jonathan Ross and much of its senior team. The reborn company is pivoting from chip challenger to an inference cloud blending its LPUs with Nvidia systems, with Nvidia itself planning to invest.
Anthropic and OpenAI publish glowing usage reports, but a new Stanford-led AI Observatory finds their work-focused methods filter out nearly half of real conversations, hiding heavier shares of health, companionship, harassment, and sexual content. It's the first independent check on how people actually use AI, not how vendors claim they do.
A week after Mojo hit 1.0, Modular open sourced the compiler and toolchain under Apache 2.0 with LLVM exceptions. The Python-inspired GPU language has now closed the loop on a 2023 pledge, making it a credible alternative for AI workloads beyond just a proprietary runtime.
Paradigm shipped a playable game that puts you in charge of an AI lab racing to bootstrap self-sustaining recursive self-improvement, balancing researchers, compute, and data. Built on the "Economics of Recursive Self-Improvement" paper, it pairs pedagogy with a model explorer, a concrete way to build intuition about why AI-accelerated-AI could hit sudden inflection points.
You’re caught up.
That was all 9 stories in the 19 August issue.