James Evans

486 posts

James Evans

James Evans

@profjamesevans

Max Palevsky Professor of Sociology, Computational & Data Science @UChicago, Santa Fe Institute, & Google tweeting about science, technology, and AI in society.

Chicago, IL เข้าร่วม Şubat 2019
440 กำลังติดตาม3.6K ผู้ติดตาม
James Evans
James Evans@profjamesevans·
Our new essay is out in Science: "Agentic AI and the Next Intelligence Explosion" For decades, the AI "singularity" has been imagined as a single, godlike mind bootstrapping itself to omniscience. In this piece with the inimitable Benjamin Bratton (@bratton) and Blaise Agüera y Arcas (@blaiseaguera), we argue this vision is wrong in its most fundamental assumption. Every prior intelligence explosion—primate sociality, human language, writing, institutions—wasn't an upgrade to individual cognitive hardware. It was the emergence of a new socially aggregated unit of cognition. AI is extending this sequence, not breaking from it. The evidence is already inside the models themselves. In recent work, we showed that frontier reasoning models like DeepSeek-R1 don't improve by "thinking longer"—they spontaneously simulate internal multi-agent debates, what we call a "society of thought" (lnkd.in/guNfRtXh). Reinforcement learning for accuracy alone causes models to rediscover what epistemology and cognitive science have long suggested: robust reasoning is a social process, even within a single mind. This opens a vast design space. A century of research on team composition, hierarchy, role differentiation, and structured disagreement has barely been brought to bear on AI reasoning. The toolkits of organizational science become blueprints for next-generation AI. Outside the model, we've entered the era of human-AI centaurs—composite actors that are neither purely human nor purely machine. Agents that fork, differentiate, recombine. Recursive societies of thought that expand when complexity demands and collapse when problems resolve. The scaling frontier isn't just bigger models. It's richer social systems—and the institutions to govern them. Just as human societies rely on persistent institutional templates (courtrooms, markets, bureaucracies), scalable AI ecosystems will need digital equivalents. The Founders would have recognized the logic: no single concentration of intelligence should regulate itself. The intelligence explosion is already here. Not as a singular ascending mind, but as a combinatorial society complexifying—intelligence growing like a city. The question is whether we'll build the social infrastructure worthy of what it's becoming. No mind is an island. Read it here in Science (science.org/doi/10.1126/sc…) or free on the arXiv (arxiv.org/abs/2603.20639)
James Evans tweet media
English
24
73
275
29.3K
James Evans รีทวีตแล้ว
James Evans
James Evans@profjamesevans·
I loved working on this ICLR paper with brilliant colleagues. It benefited from the insight that we can infer what data models have been trained on from their surprise (perplexity) at samples of web data. Then we take this insight and show how the universe of current benchmarks relates to the core knowledge on which models are built, and, through that, how those benchmarks relate to one another. Our findings suggest both the importance of refactoring existing benchmarks (and reweighting model performance across them to reflect balanced capacity) and where new benchmarks can be developed to capture valuable but un(der)measured capacity!
Honglin (虹霖) Bao@HonglinB

Excited to share our @DSI_UChicago new work accepted by ICLR 2026, with the remarkable Siyang Wu, Sida Li, Ari Holtzman @universeinanegg, and @profjamesevans James Evans! Link: lnkd.in/gd9YZNBh A thread (1/n)

English
1
6
20
2.4K
James Evans รีทวีตแล้ว
James Zou
James Zou@james_y_zou·
At #ICLR2025, we ran a large-scale randomized study: LLM feedback on 20,000 reviews. Our new @NatMachIntell paper shows how AI can improve review quality, constructiveness and rebuttal engagement supported by this data. Great work @nityathakkar_ @mertyuksekgonul @JakeSilberg + awesome ICLR chairs!
James Zou tweet media
Nitya Thakkar@nityathakkar_

Excited to share that our paper has been published in Nature Machine Intelligence! We conducted a randomized controlled trial at ICLR 2025 with 20,000+ reviews to test whether LLM feedback improves peer review quality. Link: nature.com/articles/s4225…

English
2
22
125
24.5K
James Evans รีทวีตแล้ว
Blaise Agüera (@blaiseaguera.bsky.social)
We’ve been preparing for AI as an Oracle—one brain that answers our questions. But what if intelligence emerges from interaction? What if the Singularity isn’t so singular?
Blaise Agüera (@blaiseaguera.bsky.social) tweet media
English
2
8
53
2.8K
James Evans รีทวีตแล้ว
James Evans
James Evans@profjamesevans·
Join us at Knowledge Lab as a postdoc for research collaborations with global scientists (and me!) on topics at the intersection of artificial intelligence, complex systems, computational science, and an emerging science of ideas and innovation. We have a broadly defined research agenda to reimagine and solve the most challenging problems related to: - Automated innovation in social and natural science and technology (e.g., embedding curiosity and creativity) - Large language and multi-modal models - Automated interpretability and world models - AI informed (and enabled) science and innovation policy - Novel AI paradigms and designs (e.g., inspired by complex systems) - AI, network and data science informed by the geometry and topology of data - Evolving complementary and “alien” AI Knowledge Lab cultivates a culture of curiosity and creative chaos (C^5?!) that attends to surprise (and wonder ;-). In the last year (and over many years), we have published our insights and discoveries (repeatedly) in Nature, Science, PNAS, NeurIPS, ICLR, ICML, and other top general, computer, and social science outlets. Applications are due on Feb. 28, 2026 and remain open until filled. Positions start this summer or fall. Apply here: lnkd.in/gUuWqZ5f More details: lnkd.in/gR_Kma9u
James Evans tweet media
English
1
7
36
3.6K
James Evans
James Evans@profjamesevans·
Delighted to share new work led by the remarkable @JunsolK, with @ShiyangLai, @ninoscherrer, and @blaiseaguera Blaise Agüera y Arcas—now on arXiv. We asked a simple question: What happens inside models like OpenAI's o-series, DeepSeek-R1, and QwQ when they reason? The answer surprised us. These models don't simply compute longer. They spontaneously generate internal debates among simulated agents with distinct personalities and expertise—what we call "societies of thought." Perspectives clash, questions get posed and answered, conflicts emerge and resolve, and self-references shift to the collective "we"—at rates hundreds to thousands of percent higher than chain-of-thought reasoning. There's high variance in Big 5 personality traits like neuroticism and openness, plus specialized expertise spanning physics to creative writing. The structure mirrors collective intelligence in human groups. Moreover, toggling conversational features causally toggles this capacity—beneficial cognitive behaviors like verification become more likely when they can "inhabit" different personas. What makes this remarkable from a complex systems perspective is that these societies weren't designed. They emerged from reinforcement learning rewarding only correct answers. The models discovered that distributing cognition across diverse, conflicting perspectives is an optimal strategy for distinguishing truth from error. Self-organization in service of reasoning. Even more striking: training on a simple arithmetic task (the Countdown game) produced conversational reasoning that transferred to detecting political misinformation—suggesting the generality of collective deliberation as a reasoning architecture. When we stage models with personas from the start, they learn faster. The emergent "cast of characters" is fascinating: one detailed and algebraic, another intuitive and exploratory, a third who reconciles diverse opinions. Of course, single-channel conversation can't search all collective configurations—no natural hierarchies or networked organizations—pointing toward scaffolding that explores larger spaces of social organization. Our findings resonate with Mercier & Sperber's social origins of reason and complexity research on diversity-enabled collective intelligence. Proud to pursue this work with collaborators at @Google's wonderful Paradigms of Intelligence team (thx for amazing suggestions from Blake Richards, Roberta Rocca, and Rif Saurous), UChicago Knowledge Lab (@KnowLab), and the Santa Fe Institute (@sfiscience), where long-standing work has inspired us about how complex collectives solve problems individuals cannot. Paper: arxiv.org/abs/2601.10825
James Evans tweet media
English
2
24
102
12.5K
James Evans รีทวีตแล้ว
VraserX e/acc
VraserX e/acc@VraserX·
Demis Hassabis, CEO of Google DeepMind, drops a quiet bombshell: The big question isn’t whether AI can solve problems. It’s whether AI can invent new science. Right now, it can’t. Not because of compute. Not because of data. But because it lacks something fundamental: A world model. Today’s LLMs can generate brilliant text, images, even code. But they don’t truly understand causality. They don’t know why A leads to B. They just predict patterns. Hassabis argues that real scientific discovery requires more: – Long-term planning – Stronger reasoning – And an internal model of how the world works Physics. Biology. Cause and effect. Only then can an AI run its own thought experiments. Only then do we get a true digital scientist.
English
133
264
1.7K
167K
James Evans รีทวีตแล้ว
Nicholas Fabiano, MD
Nicholas Fabiano, MD@NTFabiano·
Writing is thinking. Don't let AI do it all.
Nicholas Fabiano, MD tweet media
English
127
2K
9.3K
444.4K
James Evans
James Evans@profjamesevans·
@just_varholick I agree! I love LLMs and their ability to help me sail through adminis-trivia.
English
0
0
0
4
Justin Varholick
Justin Varholick@just_varholick·
They forgot to study how AI (LLMs) reduces the amount of time scientists must deal with administrative nonsense....but maybe it washes out once administrators start using it and increase our burden.
James Evans@profjamesevans

Thrilled to share our Nature paper, out today, on how AI use has shaped scientific careers and science as a whole (in collaboration with the amazing Qianyue Hao, @xu_fengli, and Li Yong from Tsinghua University and Zhongguancun Academy.) We analyzed tens of millions of research papers spanning four decades of natural science to understand how AI is reshaping science. The findings reveal a paradox. For individual scientists, AI is a career accelerator. Researchers who adopt AI publish 3 times more papers with fewer authors, receive 5 times more citations, and become research leaders more than a year earlier than peers who don't. AI papers appear ~20% more frequently in top-quartile journals. Annual citations run 100% higher than non-AI papers across three decades of follow-up. For science as a whole, AI is a narrowing force. AI-augmented research covers ~5% less topical ground and generates a quarter less engagement among follow-on researchers. The contraction appears in the vast majority of the 200+ subfields we examined. Citation patterns show a starker concentration: in AI research, just 22% of papers capture 80% of all citations. The mechanism is straightforward: AI use has shifted to where data is abundant, at an accelerating pace as models have grown larger. AI gravitates toward well-lit problems and away from foundational and emergent questions where data is necessarily sparse. The result is collective hill-climbing—everyone scaling the same popular peaks rather than searching for higher mountains. This creates "lonely crowds" in the scientific literature: clusters of researchers converging on identical problems without building on each other's work. The pattern holds across biology, chemistry, physics, medicine, materials science, and geology. It persists and increases through each wave of AI—from conventional machine learning through deep learning to today's generative models and LLMs. This isn't inevitable. Models that are powerful at prediction can be inverted to identify what is surprising (to those predictions), and enable us to consider and theorize entailments to surprising new data and findings. But without deliberate intervention, local incentives have and will likely continue to push scientists to optimize and compress what's already known rather than discover what isn't. The history of major discoveries is linked to new ways of seeing nature. If we want AI to accelerate breakthroughs rather than automate the familiar, we need AI systems tuned to surprise that expand sensory and experimental capacity—not just cognition. Paper: rdcu.be/eY5f7 Science commentary: science.org/content/articl… Nature commentary: nature.com/articles/d4158… Nature podcast: nature.com/articles/d4158…

English
1
0
1
42
James Evans รีทวีตแล้ว
Tim Miano
Tim Miano@tjmiano·
We're entering into a new golden age of intelligent complexity. But these are early days yet!
Tim Miano tweet media
James Evans@profjamesevans

Thrilled to share our Nature paper, out today, on how AI use has shaped scientific careers and science as a whole (in collaboration with the amazing Qianyue Hao, @xu_fengli, and Li Yong from Tsinghua University and Zhongguancun Academy.) We analyzed tens of millions of research papers spanning four decades of natural science to understand how AI is reshaping science. The findings reveal a paradox. For individual scientists, AI is a career accelerator. Researchers who adopt AI publish 3 times more papers with fewer authors, receive 5 times more citations, and become research leaders more than a year earlier than peers who don't. AI papers appear ~20% more frequently in top-quartile journals. Annual citations run 100% higher than non-AI papers across three decades of follow-up. For science as a whole, AI is a narrowing force. AI-augmented research covers ~5% less topical ground and generates a quarter less engagement among follow-on researchers. The contraction appears in the vast majority of the 200+ subfields we examined. Citation patterns show a starker concentration: in AI research, just 22% of papers capture 80% of all citations. The mechanism is straightforward: AI use has shifted to where data is abundant, at an accelerating pace as models have grown larger. AI gravitates toward well-lit problems and away from foundational and emergent questions where data is necessarily sparse. The result is collective hill-climbing—everyone scaling the same popular peaks rather than searching for higher mountains. This creates "lonely crowds" in the scientific literature: clusters of researchers converging on identical problems without building on each other's work. The pattern holds across biology, chemistry, physics, medicine, materials science, and geology. It persists and increases through each wave of AI—from conventional machine learning through deep learning to today's generative models and LLMs. This isn't inevitable. Models that are powerful at prediction can be inverted to identify what is surprising (to those predictions), and enable us to consider and theorize entailments to surprising new data and findings. But without deliberate intervention, local incentives have and will likely continue to push scientists to optimize and compress what's already known rather than discover what isn't. The history of major discoveries is linked to new ways of seeing nature. If we want AI to accelerate breakthroughs rather than automate the familiar, we need AI systems tuned to surprise that expand sensory and experimental capacity—not just cognition. Paper: rdcu.be/eY5f7 Science commentary: science.org/content/articl… Nature commentary: nature.com/articles/d4158… Nature podcast: nature.com/articles/d4158…

English
0
1
2
433
James Evans รีทวีตแล้ว
Brian Gordon
Brian Gordon@GordonBrianR·
This is a big deal. The prelim data seems to suggest that adoption of AI reduces the span of research in topic space and reduces the surprisal of the papers produced. It’s individual-level exploitation/productivity benefits at the expense of field-level exploration.
Brian Gordon tweet media
James Evans@profjamesevans

Thrilled to share our Nature paper, out today, on how AI use has shaped scientific careers and science as a whole (in collaboration with the amazing Qianyue Hao, @xu_fengli, and Li Yong from Tsinghua University and Zhongguancun Academy.) We analyzed tens of millions of research papers spanning four decades of natural science to understand how AI is reshaping science. The findings reveal a paradox. For individual scientists, AI is a career accelerator. Researchers who adopt AI publish 3 times more papers with fewer authors, receive 5 times more citations, and become research leaders more than a year earlier than peers who don't. AI papers appear ~20% more frequently in top-quartile journals. Annual citations run 100% higher than non-AI papers across three decades of follow-up. For science as a whole, AI is a narrowing force. AI-augmented research covers ~5% less topical ground and generates a quarter less engagement among follow-on researchers. The contraction appears in the vast majority of the 200+ subfields we examined. Citation patterns show a starker concentration: in AI research, just 22% of papers capture 80% of all citations. The mechanism is straightforward: AI use has shifted to where data is abundant, at an accelerating pace as models have grown larger. AI gravitates toward well-lit problems and away from foundational and emergent questions where data is necessarily sparse. The result is collective hill-climbing—everyone scaling the same popular peaks rather than searching for higher mountains. This creates "lonely crowds" in the scientific literature: clusters of researchers converging on identical problems without building on each other's work. The pattern holds across biology, chemistry, physics, medicine, materials science, and geology. It persists and increases through each wave of AI—from conventional machine learning through deep learning to today's generative models and LLMs. This isn't inevitable. Models that are powerful at prediction can be inverted to identify what is surprising (to those predictions), and enable us to consider and theorize entailments to surprising new data and findings. But without deliberate intervention, local incentives have and will likely continue to push scientists to optimize and compress what's already known rather than discover what isn't. The history of major discoveries is linked to new ways of seeing nature. If we want AI to accelerate breakthroughs rather than automate the familiar, we need AI systems tuned to surprise that expand sensory and experimental capacity—not just cognition. Paper: rdcu.be/eY5f7 Science commentary: science.org/content/articl… Nature commentary: nature.com/articles/d4158… Nature podcast: nature.com/articles/d4158…

English
3
3
6
711
James Evans รีทวีตแล้ว
Data Science Institute
Data Science Institute@DSI_UChicago·
AI helps individual scientists publish 3x more papers and get 5x more citations... but it's also shrinking the overall scope of what science explores 🤔 New study in @Nature by DSI Fac Co-Dir, Novel Intelligence @profjamesevans shows AI pulls scientists to data-rich areas... 🧵
Data Science Institute tweet media
English
1
4
3
341
James Evans รีทวีตแล้ว
Valerio Capraro
Valerio Capraro@ValerioCapraro·
A new Nature paper shows what looks like a paradox. Researchers who adopt AI tools publish more, receive more citations, and become PIs earlier. At the same time, the scope of science appears to be narrowing. How can we reconcile this apparent contradiction? LLM outputs are, by construction, a combination of existing knowledge: an average of averages. As reliance on LLMs increases, variance declines. Productivity goes up. Creativity goes down. * Paper in the first reply
Valerio Capraro tweet media
English
28
97
367
26.9K
James Evans รีทวีตแล้ว
Data Science Institute
Data Science Institute@DSI_UChicago·
We talk a lot about which jobs technology might replace. But what about the jobs that stay? New research from UChicago's @profjamesevans found that jobs with fewer skill requirements actually undergo the most radical skill transformations.
Data Science Institute tweet media
English
1
2
2
445