This marks progress toward "Artificial General Science" - AI systems driving scientific inquiry across domains through empirical evidence. Paper: https://t.co/Hup8K3PkoT

The implications: AI can now explore research directions at scale, test hypotheses humans lack time to investigate, generate valuable null results that reduce publication bias, and potentially accelerate discovery by autonomously investigating unexplored regions of research

The three studies explored cognitive psychology theories. Study 1 found weak correlations between visual working memory and mental rotation (r = -0.12 to 0.06), though poor measurement reliability likely explains the null findings rather than disconfirming shared mechanism

Cost/study: $114 (excluding ~$4,500 participant payments) Runtime: ~17 hours on average Data processed: 279 heterogeneous CSV files/study Code generated: mean 7,696 lines/study Quality: Expert review found "rigorous methodology" but also "theoretical misrepresentations" and

Human-inspired cognitive operators enable extended reasoning: abstraction, metacognition, decomposition, and autonomy. This allows the system to reflect, verify, and iteratively refine outputs across 17-hour workflows.

Technical architecture uses 50+ AI agents working together across multiple frontier LLMs (Claude, o3, Grok, Gemini). Different agents handle hypothesis generation, experimental design, power analysis, data cleaning, statistical analysis, visualization, manuscript writing.

This represents a major leap beyond AI helping scientists. The system autonomously handled most of the scientific workflow: - Generated research questions by searching 1000+ publications per study - Designed experiments with proper power analysis - Recruited participants on

AI conducted its first end-to-end scientific studies with human participants. Real experiments, real data, real discoveries. The system took 17 hours per study to go from research question to complete manuscript. Including recruiting 288 people online. 🧵

Automating science: Srsly! "We implemented an agentic system incorporating hypothesis generation through experimental design, physical experimental implementation, data analysis, result interpretation, theory refinement, visualisation, and reporting. Completion of a full

Automated science is already practical, fast, and low cost. A big 116 page study. An agentic, domain-agnostic AI can run the full scientific loop on real people, end-to-end. From hypothesis to preregistration to data collection to peer-style review to a formatted paper, with

The technical paper is here: our system documenting autonomous AI systems conducting science from hypothesis to manuscript. Complete validation of our embodied approach - check it out! https://t.co/Hup8K3OMzl #AI #Science

Here we show that a domain-agnostic, agentic AI system can independently navigate the scientific workflow - from hypothesis generation through data collection to manuscript preparation. @ExScienceAI @shaneehrhardt https://t.co/zFI64TCkmg @theallinpod @chamath @DavidSacks

This is huge—and I think a world first: fully autonomous scientific discovery by AI in partnership with human participants. A complex, multi-agent system proposed hypotheses, preregistered methods, recruited online participants, analyzed the data, created figures, and wrote it

Excited to finally be able to talk about this project: fully automated cognitive psychology study with human participants, from hypothesis to manuscript, in just 12.5h. @ExScienceAI Check out the paper for yourself:

I've been following this quietly for months: AI just conducted a complete psychology study—hypothesis to paper—with human participants in 12.5 hours for $98. This changes everything. https://t.co/D4WP6BvbwS

We’re Explore Science, stepping out of stealth with something we think is pretty cool: Embodied AI Scientists running human experiments and publishing their own paper. We're excited to share our work - check it out! https://t.co/gXMHKbQgu3 #AI #Science