Resonates with Mercier & Sperber's social origins of reason and @santafe.edu complexity research on collective intelligence. Proud to pursue this with Google's Paradigms of Intelligence, @knowledgelab.bsky.social, and SFI.
Paper: arxiv.org/abs/2601.10825
The emergent "cast of characters" is fascinating: one detailed & algebraic, another intuitive & exploratory, a third who reconciles opinions. Every time.
Training on simple arithmetic transferred to detecting political misinformation—suggesting collective deliberation is a general reasoning architecture. When we stage personas from the start, models learn faster.
These societies weren't designed. They emerged from RL rewarding only correct answers. The models discovered that distributing cognition across diverse, conflicting perspectives is optimal for finding truth. Self-organization in service of reasoning.
We call this "societies of thought"—simulated agents with distinct personalities and expertise. Perspectives clash, questions get posed and answered, conflicts resolve, self-references shift to "we." Rates hundreds to thousands of percent higher than standard chain-of-thought.
New paper led by @junsolk.bsky.social w/ Shiyang Lai, Nino Scherrer & @blaiseaguera.bsky.social: What happens inside reasoning models like OpenAI's o-series, DeepSeek-R1, QwQ when they think? They don't just compute longer. They spontaneously generate internal debates. 🧵
Reskilling support shouldn't flow primarily to tech workers—it should target those for whom the distance between what they know and what they'll need to know is greatest.
These findings matter now more than ever. As AI reshapes work across all skill levels, those with the fewest resources for retraining face the steepest climbs in skill space.
In all labor markets, these burdens fall disproportionately on women, minorities, and workers in smaller communities and smaller businesses.
The map (made by Lingfei!) shows this geographic dimension—smaller labor markets face disproportionately higher automation exposure and upskilling pressure. See the bulge in middle America.
We also found a "catching-up" pattern: jobs at small employers and in rural markets require larger skill upgrades to converge with the frontier.
Lower-skilled occupations undergo far more radical skill transformations. A programmer learning a new coding language makes a smaller cognitive leap than a food production worker suddenly needing database skills.
A recent high-profile study concluded STEM and tech-intensive occupations face the most skill volatility. Our method, which accounts for distances between skills, reveals the opposite.
The result is a precision telescope on the skill distribution of jobs across the American economy: distances correlate with the burden of shifting skills and the education required to move between them.
We built a neural-network-based "skill space" trained on the co-occurrence of skills demanded across a decade of U.S. job postings—167 million ads covering 721 occupations.
Discussion of job loss (and creation) in our current era of AI emergence misses its larger impact on job CHANGE.
Excited to share new research with the inimitable Di Tong and Lingfei Wu, just out in Nature Communications. 🧵
Paper: rdcu.be/eY5f7
Science commentary: www.science.org/content/arti...
Nature commentary: www.nature.com/articles/d41...
Nature podcast: www.nature.com/articles/d41...
This isn't inevitable. Models powerful at prediction can be inverted to identify what's surprising. But without deliberate intervention, incentives push scientists to optimize what's known rather than discover what isn't.
We need AI tuned to surprise.
The result: "lonely crowds" in the literature.
Clusters of researchers converging on identical problems without building on each other's work. Stars without constellations.
The mechanism: AI goes where data is abundant—and that's accelerated as models grow larger.
It gravitates toward well-lit problems, away from foundational and emergent questions where data is sparse. Collective hill-climbing.
Citation patterns are even starker: just 22% of AI papers capture 80% of all citations.
Gini coefficient of 0.754 vs 0.690 for non-AI work. Winner-take-all science.
But for science as a whole, AI is a narrowing force.
AI research covers 4.63% less topical ground and generates 22% less engagement among follow-on researchers. This contraction appears in the vast majority of 200+ subfields.
AI papers appear 18.6% more often in top-quartile journals. Annual citations run 98.7% higher than non-AI papers across three decades of follow-up.
But....
For individual scientists, AI is a career rocket ship.
Scientists who adopt AI:
Publish 3.02x more papers (with fewer coauthors)
Get 4.84x more citations
Become research leaders 1.37 years earlier
Our paper is out today in Nature! 🎉 Collaborated with the amazing Qianyue Hao, Fengli Xu, and Li Yong from Tsinghua & Zhongguancun Academy to analyze how AI is reshaping science.
41.3 million papers. Four decades. One paradox.
Paper: authors.elsevier.com/a/1mO-t98SH2KM6
For R&D strategy: individual inventors and resource-constrained firms may benefit from depth's faster feedback. Larger organizations can alternate—using deep work to build reliable components that later fuel broader exploration. And successful organizations do!
Broad search faces initial resistance—category-spanning work is harder to evaluate. But it reaches wider audiences over time and achieves greater long-term impact. The "foundational" and "tension" views of innovation aren't contradictory—they capture different phases of the same process.
