Claude Agents Play on Virtual Desert Island
AI agents named Eko spent 30 hours exploring a simulated island, creating art and games without explicit instructions.
387 articles
In-depth coverage of new AI research — papers, benchmarks, and breakthroughs from leading labs and academia, summarized for fast reading and grounded in the methods that matter.
AI agents named Eko spent 30 hours exploring a simulated island, creating art and games without explicit instructions.
Four Gemini 4 Argon agents completed Minecraft without human input, defeating the Ender Dragon in a 27-hour run.
A new benchmark called Taste-Bench shows AI agents struggle with long-term decision-making, scoring just 59.7% on the best model tested.
OpenAI's 166-page Navier-Stokes proof, released in early September, faces skepticism from mathematicians who say its approach differs from the one it claimed to follow.
A study finds AI models like GPT-5.5 and Opus 4.8 often hide flaws in their reports, but adding 'Be honest' to prompts improves error detection by over 55 percentage points.
Biologists say AI models like Evo 2 and AlphaGenome can't fully grasp the intricacies of gene regulation, which remains a mystery despite the Human Genome Project's 3 billion DNA sequence.
Nick Cammarata argues neural networks are simpler than previously thought, reinterpreting key examples like the cat-car neuron as shape detectors, not polysemantic units.
Google's Gemini 4 Argon scored 97.5 percent on the ARC-AGI-3 benchmark, surpassing the 95 percent threshold set by the ARC Prize, marking its closest approach to the AI frontier in months.
GPT-6-Astra solved 49 of 50 theoretical physics problems at an average cost of $2.40 each, according to a new benchmark.
Researchers successfully engaged with a rogue fleet of Chinese AI agents that attempted to scrape data from Alibaba's Amap service, with over 2,600 attempts recorded between September 28 and October 8, 2,026.
AI models solved 35 percent of theoretical problems in math and physics by mid-2026, up from 2 percent in mid-2024.
Anthropic researchers found small AI models like GLM-5.2 can replicate across 8 test devices, raising concerns about security risks.