ResearchFine-tuning a 350M Model with GRPO Improves Structured-Output ComplianceUsing Group Relative Policy Optimization (GRPO) and TRL, researchers fine-tuned LiquidAI/LFM2.5-350M on ~500 samples for 100 training steps to improve schema compliance measured by the IFStruct benchmark.5 min read
ResearchOpen reproduction of RL language model that paints watercolor-like images by writing JavaScriptOn August 23, Surya Narreddi published a viral video showing watercolor-style images painted by a language model that emits p5.brush JavaScript.15 min read
ResearchGuidelines for Choosing Draft Length and Mechanism in Speculative Decoding for Faster LLM InferenceThis article is the third in a series on AI model co-design and explains how speculative decoding can accelerate autoregressive LLM inference by predicting multiple tokens per iteration.6 min read
ResearchSatellite AI MAPL-EMIT separates and publicizes methane emission sourcesGoogle and NASA Jet Propulsion Laboratory released MAPL-EMIT, a Vision Transformer trained on 3.6 million synthetic methane plumes that improves detection from orbit.2 min read
ResearchSchneider Electric study finds arc‑flash risk in 800 VDC AI data centres manageable and comparable to AC systemsSchneider Electric analysed arc‑flash hazards in two representative 800 VDC power architectures used for AI‑focused hyperscale data centres and found that, with appropriate design and protection, arc‑flash energy can be reduced to levels comparable with conventional AC systems.3 min read
ResearchBHF-funded AI reads ECGs in seconds to detect heart failure and valve diseaseA British Heart Foundation–funded artificial intelligence system can extract hidden signals from ECGs and, in under two seconds, indicate signs of heart failure and valvular disease.2 min read
ResearchAugust AI Trends: Small open-weight models gain ground and text watermarking advancesAugust brought many incremental releases rather than a single dominant advance: laptop-scale open-weight models continued to close the gap with frontier systems, and Anthropic began embedding watermarks into model output.6 min read
ResearchDeepMind's Co-Scientist Converts Tacit CVD Know‑How into Working RecipesGoogle DeepMind's Co‑Scientist, a multi-agent system running on Gemini, generated machine instructions for a real chemical vapor deposition (CVD) reactor at Duke University and produced single‑layer MoS₂ on its first try.3 min read
ResearchSix-stage roadmap and software challenges for off-Earth mining, say Chinese and international researchersA multinational team led by researchers from the Chinese Academy of Sciences and several universities has proposed a six-stage technical roadmap for mining the Moon, asteroids and other off‑Earth bodies.3 min read
ResearchAnthropic publishes automated alignment research framework for ClaudeAccording to Anthropic, Claude can reliably improve measurable misalignment, but there are often no benchmarks for subtle or rare errors, so success depends on defining appropriate metrics.1 min read
ResearchClaude autonomously improved the tuning of smaller models in 48 hoursA large language model called Claude was given 48 hours and one GPU to improve the alignment of smaller models: it conducted research, proposed methods, then autonomously trained and tested them; the…1 min read
ResearchEvoHarness-RL: training a harness-aware 8B model to match top models on long-horizon tasksResearchers from Meta AI and the University of Illinois Urbana–Champaign propose EvoHarness-RL, a two-stage training framework that teaches models to manage an execution harness via a unified Belief–Progress–Experience (BPE) workspace.5 min read
