One landmark article per discipline — from interpretability to robots to climate. Ten technologies, zero overlap.
Nick Cammarata argues that the famous evidence for superposition was misread, and that most neurons in models up to trillions of parameters have one main job rather than an arbitrary tangle of unrelated concepts. The essay says the networks can be studied directly, and that the outlook for mechanistic interpretability is better than the field has assumed.
G2’s October 2026 survey of more than 1,000 B2B software buyers finds that 66% of organizations now run agents in limited or scaled production. The same buyers describe a distrust tax: review, corrections, and exception handling give back a large share of the speed those agents were supposed to buy.
Air Street’s July 2026 index, built with Zeta Alpha, finds NVIDIA chips in about 91% of tracked AI-paper citations, even as AMD, Huawei, and Apple move in the open literature. On the infrastructure side, Hopper is the installed base, while most tracked Grace-Blackwell capacity is still announced rather than running.
This October 2026 NEJM AI perspective, drawn from the May SAIL Year in Review, says clinical AI has shifted from speculative adoption toward measuring effects on patients, workflows, and society. It covers evaluation, workflow integration, direct-to-consumer tools, regulation, medical education, and foundational technology, and finds the gains uneven.
An Arthur D. Little report, covered on 20 October 2026, puts the physical-AI market near $18 billion against about $3 billion for humanoids. The authors say near-term value is in specialized machines in factories and warehouses, and that capable home robots are still 15 to 20 years off.
Zhang and Tang review 490 papers on text-to-image synthesis, updated through 15 October 2026, across GANs, autoregressive models, and diffusion. The survey organizes the recent work by controllability, editing, preference alignment, efficiency, safety, and models that both understand and generate images.
A survey of 2,350 professional developers, weighed against public seat counts, finds Claude Code the primary assistant for 29.4% and GitHub Copilot for 22.7%. Terminal and command-line agents rose to 38.6% of primary use, while editor extensions fell to 29.4%.
The second annual report, published 3 February 2026 and chaired by Yoshua Bengio, reviews the capabilities and risks of general-purpose AI. More than 100 experts wrote it, with backing from more than 30 countries and international organisations.
Caltech reports that a recurrent neural operator predicted a simulated low-cloud-loss tipping point more than two years ahead, with an error of about 11 days. The work, led by Anima Anandkumar, appears in PNAS as “Tipping Point Forecasting in Non-Stationary Dynamics on Function Spaces.”
A NeurIPS 2026 spotlight derives scaling laws for mixing text, vision, and audio when pretraining multimodal models, from 268 mixture-of-experts runs up to 8.3 billion parameters. Audio favors larger models over more tokens, repeating vision or audio past twice barely helps, and vision interferes with audio far more than the reverse.