Main Page - Publication Page

Paper PDF

Code

Summary

This paper introduces Automated Search for Artificial Life (ASAL), a novel framework that leverages foundation models (FMs), particularly vision-language FMs like CLIP, to automate the discovery and analysis of Artificial Life (ALife) simulations. Historically, ALife research has been hampered by the reliance on manual design and trial-and-error to explore the vast combinatorial space of possible simulation configurations. ASAL addresses this by enabling three distinct search mechanisms:

• Supervised Target Search: This method allows researchers to find simulations that generate specific, predefined phenomena. By providing textual prompts (e.g., "a red oscillating blob") at different simulation timesteps, ASAL guides the search to discover configurations that produce the desired visual outcomes. This is crucial for identifying worlds similar to our own or testing hypothetical evolutionary scenarios.

• Open-Endedness Search: A significant challenge in ALife is discovering simulations that exhibit temporally open-ended novelty, mirroring the continuous innovation observed in natural evolution. ASAL tackles this by searching for simulations whose trajectories in the FM's representation space are historically novel, meaning they deviate significantly from their past states. This approach outsources the subjective assessment of "interestingness" to the rich representations learned by FMs.

• Illumination Search: To map out the diverse landscape of emergent behaviors within a substrate, ASAL employs an illumination strategy. It searches for a set of simulations that are visually distinct from each other in the FM's representation space, effectively creating a "simulation atlas" that showcases the breadth of possibilities within a given ALife system.

ASAL's key strengths lie in its generality and its ability to handle diverse ALife substrates, including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata, provided they can be rendered as images. The framework not only facilitates the discovery of novel ALife forms and behaviors—such as previously unseen Lenia organisms, exotic Boids flocking patterns, and open-ended cellular automata—but also enables the quantitative analysis of qualitative phenomena. By using FM embeddings, ASAL can provide human-aligned metrics for complexity, diversity, and novelty, moving beyond traditional, often inadequate, quantitative measures. The paper demonstrates that ASAL is agnostic to the specific FM used, with vision-language models like CLIP and vision-only models like DINOv2 proving effective, and highlights the superiority of deep FM representations over raw pixel-based comparisons for capturing human notions of diversity. This paradigm shift promises to significantly accelerate ALife research by automating the exploration of "life as it could be."

Akarsh Kumar MIT Sakana AI, Chris Lu OpenAI, Louis Kirsch The Swiss AI Lab IDSIA, Yujin Tang Sakana AI, Kenneth O. Stanley Independent, Phillip Isola MIT, David Ha Sakana AI

Abstract

With the recent Nobel Prize awarded for radical advances in protein discovery, foundation models (FMs) for exploring large combinatorial spaces promise to revolutionize many scientific fields. Artificial Life (ALife) has not yet integrated FMs, thus presenting a major opportunity for the field to alleviate the historical burden of relying chiefly on manual design and trial-and-error to discover the configurations of lifelike simulations. This paper presents, for the first time, a successful realization of this opportunity using vision-language FMs. The proposed approach, called Automated Search for Artificial Life (ASAL), (1) finds simulations that produce target phenomena, (2) discovers simulations that generate temporally open-ended novelty, and (3) illuminates an entire space of interestingly diverse simulations. Because of the generality of FMs, ASAL works effectively across a diverse range of ALife substrates including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata. A major result highlighting the potential of this technique is the discovery of previously unseen Lenia and Boids lifeforms, as well as cellular automata that are open-ended like Conway’s Game of Life. Additionally, the use of FMs allows for the quantification of previously qualitative phenomena in a human-aligned way. This new paradigm promises to accelerate ALife research beyond what is possible through human ingenuity alone.

Conclusion

Summary

This project launches a new paradigm in ALife by taking the first step towards using FMs to automate the search for interesting simulations. Our approach is effective in finding target, open-ended, and diverse simulations over a wide spectrum of substrates. Additionally, FMs enable the quantification of many qualitative phenomena in ALife, offering a path to replacing low-level complexity metrics with deep representations aligned with humans.

Discussion

Because this project is agnostic to the FM and substrate used, it raises the question of which ones to use. The choice of FMs seems to not matter much from our experiments, and FMs in general may also be converging to similar representations of reality (Huh et al., 2024). The proper substrate largely depends on the phenomena that is being studied (e.g. self-organization, open-ended evolution, etc.). The most expressive substrate would simply parameterize all the RGB pixels of an entire video, but is useless for studying emergence. The most insightful substrates bake in as little information as possible, while maintaining vast emergent capabilities. For example, the periodic table of elements can be defined with little information, yet gives rise to the entirety of the observable universe.

Eventually, with the proper substrate, more powerful FMs, and enough compute, this paradigm may allow researchers to automatically search for worlds which start off as “simple cells in primordial soup”, then undergo “a Cambrian explosion of complexity”, and eventually become “an artificial alien civilization.” Researchers could alternatively search for hypothetical worlds where life evolves without DNA. Finding open-ended worlds would solve one of ALife’s grand challenges (Bedau et al., 2000; Stanley et al., 2017). Illuminating such a substrate could help map the space of possible lifeforms and intelligences, giving a taxonomy of life as it could be in the computational universe.

This work can be generalized by replacing the image-language FM with video-language FMs that natively process the temporal nature of simulations (Tang et al., 2023; Xu et al., 2021) or with 3-D FMs to handle 3-D simulations. To leverage the recent advances of LLMs, images can be converted to text via image-to-text models, allowing all analyses to be done in text space.

Instead of searching for ALife simulations, a similar approach could be constructed for low-level physics research. For example, in Wolfram’s Physics Project (Wolfram, 2020), one could search for the hypergraph update rule which emerges structures that an FM considers natural. At a meta-level, LLMs could be useful for generating code that describes the substrates themselves, driven by higher-level research agendas, similar to Faldor et al. (2024); Lu et al. (2024b, c).

The paper presents an ambitious and highly speculative framework that seeks to bridge the gap between abstract, high-level mathematics and the functional architecture of the human brain.

The core hypothesis is that the structural and algebraic richness of the Exceptional Simple Lie Group E8 may serve as a candidate symmetry model underlying key aspects of cortical computation, connectivity, and information processing.

Summary of the Framework

The paper proposes that the massive complexity and efficiency of the neocortex are not merely an emergent property of cellular-level biological interactions, but are fundamentally constrained and organized by a deep, elegant mathematical symmetry: E8.

E8 is the largest and most intricate of the five exceptional simple Lie groups, possessing an extraordinary 248-dimensional structure. It is a mathematical object of immense elegance that has appeared unexpectedly in various fields of theoretical physics, notably in some unified theories like string theory.

The framework draws from algebraic topology, theoretical neuroscience, and information theory to map the properties of this group onto the brain. Specifically, the study aims to:

• Map E8 to Topology: Relate the mathematical properties of E8 to the functional topology of cortical manifolds. This suggests the brain's activity patterns might organize themselves in a high-dimensional structure whose geometry is governed by the E8 root system.

• Model Dynamics: Examine how feedback loops and information flow in the cortex correspond to differential and geometric analogues within the E8 structure. The paper outlines a potential computational model that is intrinsically constrained by E8 symmetry, offering a rigid, non-arbitrary template for brain function.

• Validation and Application: The work suggests pathways for neuroscientific validation, focusing on analyzing imaging and time-series data for E8-like patterns. Furthermore, it explicitly considers applications to Artificial Intelligence (AI), hypothesizing that an AI built upon this inherent brain symmetry could achieve more efficient and human-like general intelligence.

Deeper Insights and Implications

The true significance of this paper lies in its philosophical and conceptual implications, which challenge conventional views of neuroscience and nature's elegance.

1. The Principle of Deep Mathematical Realism

By proposing E8 as the organizational principle of the brain, the paper is asserting a form of deep mathematical realism. This implies that the most efficient and robust physical and computational systems in the universe, from particle physics to consciousness, are built not just described by—but governed by—a small set of highly structured mathematical objects.

If the brain is an E8 system, it would explain its astonishing efficiency. E8, being a highly constrained structure, represents an optimal configuration of many interacting parts. The insight is that the brain is not simply a biological computer that works, but a minimal complexity, maximal computational power system whose architecture is necessitated by the requirement for this perfect symmetry. This shift in perspective moves the study of consciousness from a purely neurobiological problem to an algebraic topology problem.

2. A Symmetry-Constrained Path to AGI

The paper offers a powerful, constraint-based template for Artificial General Intelligence (AGI). Current AI often uses architectures like deep neural networks, which are highly effective but lack a demonstrable, unifying principle that links them directly to the efficiency of the human brain.

The E8 framework suggests that to build AGI, researchers should not just model connectivity, but must embed the E8 symmetry into the AI's computational core. This is the deeper insight for AI: a truly general intelligence may only be achievable by replicating the fundamental algebraic necessity of the neocortex, rather than merely its statistical or connectionist properties. This could lead to AI models that are exponentially more efficient, less prone to catastrophic forgetting, and capable of true abstract generalization.

3. Epistemology and the Limits of Reductionism

Philosophically, the E8 hypothesis directly engages with questions of epistemology and the limits of reductionism. If the brain’s highest-level functions—the things we call consciousness and thought—are simply an expression of the E8 geometry, it means these phenomena are algebraically necessary outputs of the system.

The paper argues against extreme reductionism, suggesting that to understand thought, reducing the system to individual neurons (the components) is insufficient. Instead, one must understand the symmetry group (E8) that constrains the arrangement of the components. This structuralist approach suggests that the whole (consciousness) is not merely the sum of its parts, but the expression of its governing symmetry.

In conclusion, the paper serves as a potent intellectual provocation, aiming to stimulate dialogue that views the brain not just as a complex biological machine, but as a marvel of mathematical physics, whose ultimate secrets are inscribed in the language of symmetry and exceptional Lie groups.

Abstract

In 2024, artificial intelligence (AI), for the first time, helped win a Nobel Prize. DeepMind’s AlphaFold cracked one of biology’s hardest puzzles: protein folding, the challenge of predicting how a chain of amino acids twists into the intricate 3D shape that determines its function. Scientists had struggled with this problem for decades. It was crucial for medicine and drug discovery but seemed unsolvable due to the astronomical number of possible protein structures. Then AI delivered the answer.

A game-changer, no doubt. But it also raises the question: what does this mean for science and for scientists? Is traditional scientific inquiry becoming obsolete? Are we approaching a future where algorithms are the primary drivers of discovery, relegating humans to the sidelines?

Throughout history, every breakthrough technology has redefined how discoveries were made, marking four distinct eras of science [1] (Fig 1). The first, the empirical era, relied on direct observation, as Copernicus challenged the Earth-centered view of the universe by observing the skies. The second, the theoretical era, introduced mathematics to predict nature, like Newton’s equations of motion that shaped physics for centuries. The third, the computational era, which began in the 1950s, harnessed computers to simulate complex systems, leading to Kohn and Pople’s quantum chemistry Nobel Prize. The fourth, the data-driven era of our 21st century, uses machine learning to extract patterns from vast datasets, with AlphaFold solving protein structures by learning from the protein data bank [2].

Today, we stand at the doorstep of the fifth era of science—the artificial scientific intelligence era—where companies like Google, Lila Sciences, and Sakana are unveiling AI scientists that not only assist research but drive discoveries, generate hypotheses, and test them on their own [3–5] (Fig 1). Hence, why not let AI run the show from here?

In some fields, perhaps we can. In chemistry, organic synthesis—the process of assembling complex drug-like molecules from basic building blocks—is now guided by interpretable AI models that help scientists plan each step [6]. In materials science, generative AI can design novel inorganic compounds with tailored mechanical, electronic, and magnetic properties, accelerating innovation with minimal human tuning [7]. These are domains where experimental feedback is relatively tractable, simulations are mature and the data is plentiful and structured. In short, these fields provide ideal conditions for autonomous AI exploration.

But in many other areas, letting an AI run the show today would be like sending a self-driving car down a dirt road with half a map and no GPS. AI might have the horsepower, but it still needs humans to steer it around the pitfalls of specialized scientific data. Nowhere is this clearer than in biomedical imaging, where highly curated datasets are nothing like what traditional large vision models are trained on.

First, biomedical imaging datasets are often tiny by AI standards, and for good reason: collecting them requires technical equipment and trained professionals; labeling them demands significant time and expert input; and strict privacy regulations often limit access. MedPix, a leading medical imaging database, contains just 59,000 images and the Allen Cell Feature Explorer, one of the largest publicly available collections of high-resolution 3D images of human stem cells, only around 32,000 images. That is about a thousand times fewer than what is needed for AI to perform. This is where scientists step in.

Scientists are redefining AI to do more with less, helping algorithms find meaning in images even when data are scarce. One approach involves using mathematical insights to redesign the core building blocks of neural networks. Traditional models fall apart when we strip away their layers or parameters, but these new architectures stay strong—even with just a single layer and two convolutional filters [8]—precisely because they are built to thrive on small data. And, scientists do not just bend the design of the model to fit the lack of data, they also reimagine the data ecosystem to power the model; they decide what data to collect, how to collect it, and how to weave together existing, but fragmented, specialized datasets to train AI models for a wide variety of tasks, including brain tumor classification or diabetic retinopathy grading [9].

But scientific data is not just scarce, it is often noisy. Cryo-electron microscopy (cryo-EM), a Nobel Prize-winning technology that lets us see the invisible [10]—revealing molecules at the tiniest scale—produces incredibly blurry images, where the important details are 100 times weaker than the noise. It is like trying to recognize a friend in a crowd while wearing someone else’s prescription glasses. This stands in stark contrast to the crisp, high-resolution images—like street scenes, faces, or everyday objects—that traditional AI vision models are trained on.

Yet scientists have techniques to extract meaning from even the noisiest images. In cryo-EM, they can reconstruct the 3D shapes of molecules buried in noise; for example, providing the first high-resolution images of SARS-CoV-2 during the COVID-19 pandemic [11,12]. Today, they are combining that hard-won expertise with the power of AI. One breakthrough pairs a powerful denoising module with a foundation model, enabling AI to tackle the notoriously difficult processing steps of cryo-EM images [13]. Crucially, this was only possible because scientists also applied their domain expertise to curate a high-quality dataset by cleaning, annotating, and aggregating 529 verified cryo-EM datasets into one large training set that AI could learn from.

It is clear that AI presents an enormous opportunity for science, potentially the most powerful tool we have ever had in our arsenal. But the fifth era of artificial scientific intelligence is not void of human scientists: quite the opposite. In many ways, the future of revolutionary discoveries lies in this synergy: human expertise guiding AI, and AI augmenting human expertise. It is as if we have hired the most overachieving and wildly enthusiastic intern; one who works at superhuman speed, never sleeps, and eagerly devours mountains of data. They hold exceptional potential, but without proper guidance anchored in scientific knowledge, they are more likely to set the lab on fire than to push science forward.

Instead of hoping AI will magically handle limited, noisy, specialized data, we need experts to tailor algorithms to the realities of fields like biology and medicine, and to tailor data to the new requirements of the AI technology. To enter the fifth era of science, we need to equip researchers with AI expertise, AI experts with domain knowledge, and universities with interdisciplinary programs. The labs that thrive will be those where domain experts and AI specialists work in sync or where scientists master both. The next scientific revolution will come from teams who can judiciously steer AI, knowing when to trust it, when to adjust its course, and when to drive it into uncharted territory.

This Microsoft Research paper, "The Impact of Generative AI on Critical Thinking," presents a comprehensive analysis based on a large-scale survey investigating how Generative Artificial Intelligence (Gen AI) technologies are influencing human critical thinking abilities across various domains. The study acknowledges the dual-edged nature of Gen AI, identifying both significant opportunities for augmentation and considerable risks of degradation.

The research begins by framing critical thinking as a multifaceted cognitive process involving analysis, evaluation, inference, explanation, and self-regulation. It highlights the unprecedented capabilities of Gen AI models, such as large language models (LLMs), to process vast amounts of information, generate diverse content, and perform complex reasoning tasks. The primary objective of the survey was to gather empirical insights from a diverse demographic including students, educators, professionals, and AI researchers on their perceived experiences and the observed effects of Gen AI on their own critical thinking or that of others.

The survey methodology involved a mixed-methods approach, combining quantitative Likert-scale questions to gauge agreement on various impact statements and qualitative open-ended responses to capture nuanced experiences and elaborations. Participants were asked about their frequency of Gen AI use, the specific tools they employed, and their perceptions regarding its influence on aspects like information analysis, problem-solving, decision-making, and creativity.

The findings reveal a complex landscape. On the positive side, a substantial portion of respondents reported that Gen AI acts as a powerful cognitive augmentor. It was frequently cited as a valuable tool for brainstorming, generating initial ideas, and exploring diverse perspectives that might otherwise be overlooked. Users found it adept at synthesizing information rapidly from vast datasets, thereby reducing the initial cognitive load associated with information gathering and allowing them to focus more on higher-order analysis and evaluation. For many, Gen AI facilitated the automation of routine or repetitive analytical tasks, freeing up mental resources for more complex, creative, and strategic thinking. Educators noted its potential in personalizing learning experiences and offering immediate feedback, which could, if properly utilized, foster deeper engagement with critical concepts.

Conversely, the survey also unearthed significant concerns regarding the potential erosion of critical thinking skills. A prominent finding was the risk of over-reliance on Gen AI outputs, where users might become less inclined to engage in independent thought, rigorous fact-checking, or deep analytical processing. This reliance could lead to a 'deskilling' effect, where fundamental cognitive abilities like information synthesis, logical deduction, and error detection atrophy due to externalization to AI systems. Participants expressed worries about the phenomenon of 'algorithmic bias' and 'hallucinations,' where Gen AI might generate plausible but incorrect or biased information, making it harder for users to discern truth from falsehood without sufficient domain expertise or critical vigilance. There was also concern that the ease of generating content might reduce the effort invested in original thought and creative problem-solving, leading to a homogenization of ideas or a diminished capacity for truly novel contributions. The 'black box' nature of some AI models, where the reasoning process is opaque, further complicates the development of user trust and the ability to critically evaluate AI-generated solutions.

Nuance in the findings highlighted that the impact of Gen AI is not uniform and largely depends on the user's existing critical thinking proficiency, AI literacy, and the context of use. Highly skilled critical thinkers often leveraged Gen AI as an advanced tool to enhance their existing capabilities, viewing it as a co-pilot rather than a replacement for their intellect. In contrast, those with developing critical thinking skills were more susceptible to the negative effects, such as accepting AI outputs uncritically. The importance of 'prompt engineering' skills – the ability to effectively communicate with and guide AI models – emerged as a critical factor in maximizing positive outcomes and mitigating risks.

The paper concludes with significant implications for education, professional development, and policy. It advocates for the urgent integration of 'critical AI literacy' into curricula across all levels, emphasizing the need to teach individuals not just how to use Gen AI, but how to critically evaluate its outputs, understand its limitations, and ethically interact with it. The study underscores the necessity for developing metacognitive strategies that encourage users to reflect on their own thinking processes in conjunction with AI. Ultimately, the research suggests that the future of critical thinking in the age of Gen AI will hinge on fostering a symbiotic relationship where humans leverage AI's strengths while actively nurturing and exercising their unique cognitive capabilities, ensuring that technology serves as an amplifier of human intellect rather than a substitute for it.