--- title: "The Ends of Science" description: "Will Freudenheim, Imran Sekalala, and Darren Zhu propose studying scientific AI as a model superorganism: an experimental system whose internal operations can reveal mechanisms that its predictions alone leave unexplained." type: "reading-notes" authors: - "Will Freudenheim" - "Imran Sekalala" - "Darren Zhu" published: "2023-07-26" original_url: "https://endsofscience.org/" canonical_url: "https://antikythera.wiki/work/studio/ends-of-science" md_url: "https://antikythera.wiki/md/work/studio/ends-of-science" last_updated: "2026-09-06" site: "Antikythera Wiki" --- # The Ends of Science > Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below. - **Authors:** Will Freudenheim, Imran Sekalala, Darren Zhu - **Published:** 2023-07-26 - **Kind:** Studio projects · Studio paper - **Venue:** Planetary Computation Studio - **Original:** https://endsofscience.org/ - **Length:** 1.6k words - **This page:** https://antikythera.wiki/work/studio/ends-of-science - **Notes generated:** 2026-09-06 ## Summary [Will Freudenheim](https://antikythera.wiki/people/will-freudenheim), [Imran Sekalala](https://antikythera.wiki/people/imran-sekalala), and [Darren Zhu](https://antikythera.wiki/people/darren-zhu) propose studying scientific AI as a model superorganism: an experimental system whose internal operations can reveal mechanisms that its predictions alone leave unexplained. Their starting problem is the widening separation between producing a finding and understanding or verifying it. They distinguish theories that outrun verification from useful empirical results that lack causal explanations. Mechanistic interpretability, the investigation of how neural networks produce particular outputs, could help address both gaps. The proposed method combines large observational datasets, scientific foundation models, and experiments on those models. The title names a change in science's methods and purposes, rather than the exhaustion of discovery. ## The argument The authors begin with a mismatch between machine performance and human judgment. AlphaGo's celebrated move against Lee Sedol demonstrated an unfamiliar route to success; adversarial play against Go systems also exposed failures concealed by apparently superhuman competence. Scientific AI raises an analogous problem. A result can surprise its observers without making clear whether that surprise indicates discovery or error. Producing more hypotheses does not by itself improve the means of distinguishing them. This problem precedes machine learning. The essay connects expanding publication volume and disciplinary specialization to difficulties in evaluating research. Shinichi Mochizuki's proposed proof of the abc conjecture supplies an example of work whose unfamiliar conceptual apparatus complicates assessment by other specialists. The example establishes that a barrier to understanding can arise within human science. AI could intensify a difficulty already present in collective knowledge production. Two terms separate the resulting gaps. An epistemic overhang arises when a theory exceeds available means of verification or application. An epistemic underhang arises when an empirical result can be used without an adequate account of its mechanism. The distinction concerns the relation among claims, evidence, and explanation. Neither gap is guaranteed to disappear through the accumulation of more papers. The authors argue that unresolved gaps can damage scientific communities' ability to establish shared judgments. Their response draws on experimental biology. A model organism is useful because researchers can manipulate and inspect it, then investigate whether the resulting knowledge generalizes beyond that organism. A scientific neural network could serve a comparable purpose. Its value would extend beyond answering questions to becoming something on which explanatory experiments can be performed. Mechanistic interpretability supplies the proposed instruments for those experiments. The essay moves from demonstrations of interpretable neural computation to a larger research program. Train a scientific foundation model on extensive observations, then investigate the operations through which it organizes those observations. A network's unfamiliar solution could become an intelligible alternative to an established account. The proposal nevertheless leaves a substantive question open: discovering how a model computes a successful answer does not by itself establish how the corresponding natural process works. The authors offer examples and a methodological direction, without a general procedure for validating that transfer. Planetary sensor networks enlarge the program's possible scope. Environmental observations could feed models whose internal organization becomes a new object of science. Scientific infrastructure would thereby contribute to the planet's capacity to represent and investigate its own processes. The conclusion places artificial intelligence within an expanding history of inquiry rather than outside science as an inscrutable replacement for it. ## Section by section ### Introduction: Scientific Knowledge Production The introduction contrasts the traditional sequence of observation, hypothesis, and experiment with machine learning's capacity to derive candidate patterns from large datasets. Its concern is the evaluation burden created by scale and opacity. AlphaGo introduces the ambiguity of surprising performance, while the discussion of publication growth and specialization places that ambiguity within an institutional problem. Mochizuki's lengthy proposed proof gives the argument a human precedent: originality and communicability can pull apart even among trained peers. The authors treat the controversy as an example of difficult verification, rather than resolving the proof's mathematical status. ### Epistemic Overhangs and Underhangs The second section adapts the language of AI capability overhangs to scientific knowledge. Theories may outrun verification, while applications may outrun explanation. The authors use acetaminophen as an example of the latter, invoking uncertainty about mechanisms despite established use. They also interpret disputes over SARS-CoV-2 origins and ivermectin through these categories. Those interpretations belong to the essay's argument about scientific disagreement; the essay does not supply clinical evidence establishing ivermectin's efficacy. Its cited ivermectin study concerns laboratory findings, so the example cannot establish that an effective treatment merely awaits explanation. The methodological claim is that new kinds of inquiry require new means of scrutiny. The authors invoke microscopy and experimental design as historical precedents for changes in what scientists can observe or verify. Their analogy assigns interpretability a comparable role in scientific AI. The section does not specify institutional arrangements for resolving disputes once interpretability produces additional explanations. ### Model Superorganism The final section develops the biological analogy through peas, yeast, fruit flies, and mice. Manipulation and observation make these organisms scientifically productive; the proposed AI equivalent requires access to mechanisms within and across models. The authors cite artificial curve detectors and research on primate visual cortex as evidence that computational and biological investigations can inform one another. A transformer trained on modular addition supplies a different example: inspection revealed a solution using trigonometric identities, showing that a learned procedure can depart from an observer's expected implementation. AlphaFold provides a prospective application. The authors suggest investigating how its architecture uses sequence information to predict protein structures. They also cite experiments in which language models generate and score explanations of another model's neuron activations. Their imagined machine counterpart to Mochizuki extends this possibility into mathematical communication. These are proposals of differing maturity, not a demonstrated general system for explaining scientific discoveries. The section closes by connecting metagenomic, bioacoustic, and hyperspectral observations to planetary research. Its proposed sequence is observation, foundation-model training, and interrogation through interpretability tools. Darwin's concluding image of evolving life supplies a final analogy for open-ended discovery. The references connect this program to metascience, model-organism research, machine learning, and histories of scientific instruments. ## Key concepts - **[AI Epistemic Overhang](https://antikythera.wiki/terms/ai-epistemic-overhang)** — A gap in which theories or discoveries exceed available verification or application. The essay distinguishes it from AI capabilities that merely lack adoption. - **Epistemic underhang** — An empirical finding that permits application while its causal mechanism remains inadequately understood. - **Model superorganism** — The proposed combination of scientific inquiry and AI, treated as an experimental system whose internal organization can be studied for explanations. - **Mechanistic interpretability** — Investigation of the internal computations producing a model's behavior, proposed here as an instrument for scientific discovery and explanation. - **[Model Epistemology](https://antikythera.wiki/terms/model-epistemology)** — A related corpus concept for knowledge mediated by computational models. This essay gives it a specific methodological form: investigate a trained model to recover explanations from observed patterns. ## Connections [After Alignment](https://antikythera.wiki/work/journal/afteralignment) directly credits this project and its three authors for the epistemic-overhang concept and the experimental-superorganism proposal. Bratton extends their methodological argument into a claim about alignment: discoveries that exceed existing human categories can be valuable disclosures to which culture should adapt. A difference of emphasis remains. This essay seeks tools that make gaps tractable, while Bratton stresses preserving the discoveries that create those gaps. [Speculative Philosophy of Planetary Computation](https://antikythera.wiki/work/journal/spoc) agrees that computational mediation can disclose realities inaccessible to unaided perception. Its discussion of model epistemology and the simulation-led observation of black holes provides a broader framework for this essay's proposed scientific instruments. The studio paper concentrates that framework on interrogating learned models, where explaining computational operations and explaining nature remain connected but distinct tasks. ## Related works - [After Alignment](https://antikythera.wiki/work/journal/afteralignment) - [Speculative Philosophy of Planetary Computation](https://antikythera.wiki/work/journal/spoc)