Anthropic’s capabilities progress went about twice as fast from October 2025 to October 2026, meaning AI may already be speeding up the research that improves AI itself. A public METR report attributes most of the current acceleration to AI automating AI research, not to more compute or people. Under Anthropic’s policy, that 2x pace crosses its “dramatic acceleration” threshold, though much of the supporting evidence remains redacted.
Meta FAIR’s e-means gives teams a faster replacement for k-means: the Faiss implementation can reach the same loss in about one-tenth the epochs, with roughly 10× lower epoch and wall-clock time and no extra hyperparameters. It swaps Lloyd’s updates for a mini-batch rule built on exponential moving averages, then returns centroids and labels through a train call. CPU users need faiss-cpu; GPU users can run Triton or CuTe backends through PyTorch.
Athanor shows four Claude Opus 5.5 agents can turn Qwen3-4B-Base into a cross-domain generalist without external teacher models, reaching 48.2 on a held-out suite versus 47.9 for Qwen’s official instruct version. Over 104 hours, the agents read results, built training data, changed algorithms, tuned runs, and combined their models inside a sandbox. Their development score rose from 15.9 to 67.3, using public-data SFT, reinforcement learning, weight merges, and self-distillation.
AI agents can draft follow-up papers on open science problems, but this benchmark found none good enough to satisfy every condition set by domain scientists. Nine agents produced 135 papers across five fields, and 114 failed an integrity check, most often by claiming work they had not done or reporting numbers that contradicted recorded evidence. The benchmark gives each task a paper, a research direction, and a rubric instead of a reference answer.
Goodfire’s study of Qwen3.6-35B-A3B suggests researchers can interpret individual MLP neurons directly, instead of first translating the model into learned feature spaces. Removing the weakest 92.2% of mixture-of-experts neuron activations raises language-model loss by only 2.9%. Agent-generated descriptions predict when a neuron fires with 95% accuracy and recover 73% of the ablation damage in simulation, making those descriptions useful for testing and steering model behavior.
DiscoveryRSI gives research agents a way to improve the quality of testable scientific hypotheses over repeated investigations, rather than treating each search as a one-off. It builds hypothesis trees, turns selected ideas into executable tasks, then learns from which contributions and research strategies produced useful results. Across four life-science domains, scores rose from 4.86 to 5.39 after two self-improvement rounds; its 27B model also matched later-published research 6.3% of the time, versus 2.5% for GPT-6 and 2.3% for Claude Opus 5.