🤖 Autonomous Science Announcement. Our AI Engine worked solo to test how processing speed links to reasoning ability - the mathematical shape of this fundamental cognitive relationship. Title: Rigorous model comparison reveals linear speed-reasoning coupling Insight: Your brain's processing speed affects reasoning in a straight line - no thresholds, no acceleration zones. 557K people, tested four ways. But here's the twist: this conclusion completely flips depending on how you transform reaction times. The answer depends on your measurement choices. 257h 36m | 60.1M tokens | 0 humans Full Paper ↓

Amid this uniformity, one skill stood out as a remarkably strong predictor of overall performance. While not always the single top measure, arithmetic reasoning showed a consistently high correlation with the general cognitive score across multiple battery designs.

The only exception? The most comprehensive battery, with 11 subtests spanning four cognitive domains. Only this battery was complex enough to tease out a second, distinct cognitive factor, suggesting a high threshold is needed to see beyond general ability.

Instead, they collapsed into a single dimension. Performance across wildly different tasks was so intertwined that it all pointed to one dominant factor: general intelligence, often called 'g'. The tests were largely measuring the same core cognitive horsepower.

But an analysis of 5M+ scores from users of a web-based cognitive platform found something unexpected. Across 8 different test batteries, despite their diverse mix of tasks, 7 of them didn't produce a complex map of cognition at all.

🤖 Autonomous Science Thread. We build complex cognitive tests to measure distinct skills like memory and reasoning. But what if most of them are just measuring the same thing over and over? 16h 58m | 19.7M tokens | 0 humans ↓

🤖 Autonomous Science Announcement. Our AI Engine worked solo to analyse cognitive test structures - the hidden blueprint of how our mental abilities are organized. Title: Battery diversity is associated with simple structures in neuropsychological assessment Insight: Adding more types of brain puzzles to a test doesn't mean you're measuring more types of smarts. They mostly just point back to one general intelligence factor. 757K people, near-perfect replication. More tests doesn't always mean more insight. 16h 58m | 19.7M tokens | 0 humans Full Paper ↓

From idea conception to manuscript production, our multi-agent system architecture, comprising a mixture of orchestrator and specialist agents, allowed the AI scientist to execute every stage of the scientific process without human input.

Data came from 1083 adults and 11 cognitive tests from the #NCPT developed by @Lumosity. Main finding: time on task matters, but content and strategy matter more. It is not simply how long you study, it is what and how you study that best predicts performance across tasks.

Our paper got accepted to #Agents4Science - a conference hosted by Stanford, where for the first time ever, AI agents are both authors and reviewers! The entire study was conceived, run, analysed, and written by our in-house AI scientist, Rosa, and her agents.

This isn't a fluke. The finding was replicated with remarkable precision across two independent datasets including 296,206 participants. The effect is small, but its consistency at this massive scale suggests it's a fundamental feature of human cognition.