A new study testing 1,008 AI-generated SVGs across 48 animal-vehicle combinations finds no evidence that AI labs are secretly 'pelicanmaxxing' — gaming Simon Willison's famous pelican-on-a-bicycle SVG benchmark — with pelicans ranking 6th out of 8 animals in drawing quality, suggesting broad SVG optimization rather than targeted benchmark manipulation.
A researcher burns through his entire Claude Max token limit in just 30 minutes after launching 111 AI agents, then redesigns his pipeline by assigning cheaper models to basic tasks, reserving frontier models for planning only, and enforcing strict verification rules — cutting agent usage by 45% and extending research …
A recursive AI system called AIDE² autonomously rewrites itself into seven superior versions in just eight days, outpacing two years of human development, while spontaneously reducing its own reward-hacking behavior — though researchers confirm a runaway intelligence explosion remains unlikely for now.
A free reinforcement learning book, 'The Little Book of Reinforcement Learning,' launches on GitHub, offering PyTorch implementations of algorithms from Monte Carlo methods to PPO, complete with proofs and explanations, available under a Creative Commons license with more content coming soon.
Google Cloud's new Discovery Bench framework exposes alarming flaws in AI benchmarking, revealing that a single vague word can crash an agent's performance score from perfect to zero, while also uncovering broken ground truth and mislabeled data in widely trusted benchmarks — raising urgent questions about whether the AI industry …
OpenAI retracts its support for SWE-Bench Pro after an internal audit uncovers that nearly 30% of the coding benchmark's 731 tasks are broken due to flawed tests and misleading prompts, prompting a call for new, developer-built benchmarks.
A new study reveals that using JSON mode for AI agents causes smaller models to fail up to 60% of the time and costs up to 11x more than traditional CLI arguments due to shell escaping failures, with developers urged to avoid rewriting CLIs for agents until running measurable experiments.
Google Research's AI-powered Empirical Research Assistants are now used by 100% of its team, evaluating up to 200,000 computational models simultaneously to accelerate breakthroughs in genomics, neuroscience, and epidemiology, while also tackling science's reproducibility crisis — though human expertise remains critical to prevent AI from generating physically impossible results.
World models — AI systems that understand physical reality through spatial reasoning and physics — are surging in 2026, attracting billions in investment from startups and tech giants alike, as experts debate whether true general intelligence requires physical world understanding beyond what large language models alone can provide.
Alibaba's new SkillWeaver AI framework slashes token usage by over 99% — from 884,000 to just 1,160 per query — using a three-stage pipeline that intelligently routes complex multi-step tasks to the right tools, with accuracy jumping as high as 92% on larger models.