Neil Thompson

Tal Shnitzer, Anthony Ou, Mírian Silva et al.

There is a rapidly growing number of open-source Large Language Models (LLMs) and benchmark datasets to compare them. While some models dominate these benchmarks, no single model typically achieves the best accuracy in all tasks and use cases. In this work, we address the challenge of selecting the best LLM out of a collection of models for new tasks. We propose a new formulation for the problem, in which benchmark datasets are repurposed to learn a "router" model for this LLM selection, and we show that this problem can be reduced to a collection of binary classification tasks. We demonstrate the utility and limitations of learning model routers from various benchmark datasets, where we consistently improve performance upon using any single model for all tasks.

Alexander K. Saeri, Jess Graham, Michael Noetel et al.

Artificial intelligence poses many risks, ranging from familiar present-day harms to unprecedented and potentially catastrophic ones. Effective risk management requires prioritization: we must understand which risks are most severe, who is most vulnerable, and who is most responsible for addressing them. We report results from a three-round Delphi study conducted late 2025 with 272 international AI experts. Experts rated 24 AI risks on harm probability and severity, sector and actor vulnerability, actor responsibility, and overall concern. Experts estimated the five most severe harms in the next 5 years were likely to come from dangerous capabilities, competitive dynamics, weapons & cyberattacks (including CBRNE), power centralization, and false information. In a business-as-usual scenario, experts judged 18 of 24 risks as having a more than 10% probability of catastrophic outcomes (e.g., more than 1 million deaths or more than USD 100B in financial loss) in the next 5 years (2025-2030). In a scenario where pragmatic mitigations are implemented, experts still judged five risks as having a more than 10% probability of catastrophic outcomes: dangerous capabilities, weapons & cyberattacks, environmental harm, inequality & unemployment, and power centralization. All 24 risks were judged as being more than 5% likely to cause catastrophic outcomes. AI users and the general public were judged the most vulnerable to these risks, but experts assigned the highest responsibility for addressing them to general-purpose AI developers and governance actors (including governments, regulators, and standards bodies). Across most risks, experts identified information, finance, and national security as the most vulnerable sectors. These findings can guide AI risk prioritization and clarify expert expectations about who should bear responsibility for mitigation.

Peter Slattery, Alexander K. Saeri, Emily A. C. Grundy et al.

The risks posed by Artificial Intelligence (AI) are of considerable concern to academics, auditors, policymakers, AI companies, and the public. However, a lack of shared understanding of AI risks can impede our ability to comprehensively discuss, research, and react to them. This paper addresses this gap by creating an AI Risk Repository to serve as a common frame of reference. This comprises a living database of 777 risks extracted from 43 taxonomies, which can be filtered based on two overarching taxonomies and easily accessed, modified, and updated via our website and online spreadsheets. We construct our Repository with a systematic review of taxonomies and other structured classifications of AI risk followed by an expert consultation. We develop our taxonomies of AI risk using a best-fit framework synthesis. Our high-level Causal Taxonomy of AI Risks classifies each risk by its causal factors (1) Entity: Human, AI; (2) Intentionality: Intentional, Unintentional; and (3) Timing: Pre-deployment; Post-deployment. Our mid-level Domain Taxonomy of AI Risks classifies risks into seven AI risk domains: (1) Discrimination & toxicity, (2) Privacy & security, (3) Misinformation, (4) Malicious actors & misuse, (5) Human-computer interaction, (6) Socioeconomic & environmental, and (7) AI system safety, failures, & limitations. These are further divided into 23 subdomains. The AI Risk Repository is, to our knowledge, the first attempt to rigorously curate, analyze, and extract AI risk frameworks into a publicly accessible, comprehensive, extensible, and categorized risk database. This creates a foundation for a more coordinated, coherent, and complete approach to defining, auditing, and managing the risks posed by AI systems.

Fraser Mince, Dzung Dinh, Jonas Kgomo et al.

Pushing the boundaries of machine learning often requires exploring different hardware and software combinations. However, the freedom to experiment across different tooling stacks can be at odds with the drive for efficiency, which has produced increasingly specialized AI hardware and incentivized consolidation around a narrow set of ML frameworks. Exploratory research can be restricted if software and hardware are co-evolving, making it even harder to stray away from mainstream ideas that work well with popular tooling stacks. While this friction increasingly impacts the rate of innovation in machine learning, to our knowledge the lack of portability in tooling has not been quantified. In this work, we ask: How portable are popular ML software frameworks? We conduct a large-scale study of the portability of mainstream ML frameworks across different hardware types. Our findings paint an uncomfortable picture -- frameworks can lose more than 40% of their key functions when ported to other hardware. Worse, even when functions are portable, the slowdown in their performance can be extreme and render performance untenable. Collectively, our results reveal how costly straying from a narrow set of hardware-software combinations can be - and suggest that specialization of hardware impedes innovation in machine learning research.

Tamay Besiroglu, Sage Andrus Bergerson, Amelia Michael et al.

There are pronounced differences in the extent to which industrial and academic AI labs use computing resources. We provide a data-driven survey of the role of the compute divide in shaping machine learning research. We show that a compute divide has coincided with a reduced representation of academic-only research teams in compute intensive research topics, especially foundation models. We argue that, academia will likely play a smaller role in advancing the associated techniques, providing critical evaluation and scrutiny, and in the diffusion of such models. Concurrent with this change in research focus, there is a noticeable shift in academic research towards embracing open source, pre-trained models developed within the industry. To address the challenges arising from this trend, especially reduced scrutiny of influential models, we recommend approaches aimed at thoughtfully expanding academic insights. Nationally-sponsored computing infrastructure coupled with open science initiatives could judiciously boost academic compute access, prioritizing research on interpretability, safety and security. Structured access programs and third-party auditing may also allow measured external evaluation of industry systems.

Alexander K. Saeri, Sophia Lloyd George, Jess Graham et al.

Organizations and governments that develop, deploy, use, and govern AI must coordinate on effective risk mitigation. However, the landscape of AI risk mitigation frameworks is fragmented, uses inconsistent terminology, and has gaps in coverage. This paper introduces a preliminary AI Risk Mitigation Taxonomy to organize AI risk mitigations and provide a common frame of reference. The Taxonomy was developed through a rapid evidence scan of 13 AI risk mitigation frameworks published between 2023-2025, which were extracted into a living database of 831 AI risk mitigations. The mitigations were iteratively clustered & coded to create the Taxonomy. The preliminary AI Risk Mitigation Taxonomy organizes mitigations into four categories and 23 subcategories: (1) Governance & Oversight: Formal organizational structures and policy frameworks that establish human oversight mechanisms and decision protocols; (2) Technical & Security: Technical, physical, and engineering safeguards that secure AI systems and constrain model behaviors; (3) Operational Process: processes and management frameworks governing AI system deployment, usage, monitoring, incident handling, and validation; and (4) Transparency & Accountability: formal disclosure practices and verification mechanisms that communicate AI system information and enable external scrutiny. The rapid evidence scan and taxonomy construction also revealed several cases where terms like 'risk management' and 'red teaming' are used widely but refer to different responsible actors, actions, and mechanisms of action to reduce risk. This Taxonomy and associated mitigation database, while preliminary, offers a starting point for collation and synthesis of AI risk mitigations. It also offers an accessible, structured way for different actors in the AI ecosystem to discuss and coordinate action to reduce risks from AI.

Matthias Mertens, Natalia Fischl-Lanzoni, Neil Thompson

Do leading LLM developers possess a proprietary ``secret sauce'', or is LLM performance driven by scaling up compute? Using training and benchmark data for 809 models released between 2022 and 2025, we estimate scaling-law regressions with release-date and developer fixed effects. We find clear evidence of developer-specific efficiency advantages, but their importance depends on where models lie in the performance distribution. At the frontier, 80-90% of performance differences are explained by higher training compute, implying that scale--not proprietary technology--drives frontier advances. Away from the frontier, however, proprietary techniques and shared algorithmic progress substantially reduce the compute required to reach fixed capability thresholds. Some companies can systematically produce smaller models more efficiently. Strikingly, we also find substantial variation of model efficiency within companies; a firm can train two models with more than 40x compute efficiency difference. We also discuss the implications for AI leadership and capability diffusion.

Wensu Li, Atin Aboutorabi, Harry Lyu et al.

This paper develops a unified framework for evaluating the optimal degree of task automation. Moving beyond binary automate-or-not assessments, we model automation intensity as a continuous choice in which firms minimize costs by selecting an AI accuracy level, from no automation through partial human-AI collaboration to full automation. On the supply side, we estimate an AI production function via scaling-law experiments linking performance to data, compute, and model size. Because AI systems exhibit predictable but diminishing returns to these inputs, the cost of higher accuracy is convex: good performance may be inexpensive, but near-perfect accuracy is disproportionately costly. Full automation is therefore often not cost-minimizing; partial automation, where firms retain human workers for residual tasks, frequently emerges as the equilibrium. On the demand side, we introduce an entropy-based measure of task complexity that maps model accuracy into a labor substitution ratio, quantifying human labor displacement at each accuracy level. We calibrate the framework with O*NET task data, a survey of 3,778 domain experts, and GPT-4o-derived task decompositions, implementing it in computer vision. Task complexity shapes substitution: low-complexity tasks see high substitution, while high-complexity tasks favor limited partial automation. Scale of deployment is a key determinant: AI-as-a-Service and AI agents spread fixed costs across users, sharply expanding economically viable tasks. At the firm level, cost-effective automation captures approximately 11% of computer-vision-exposed labor compensation; under economy-wide deployment, this share rises sharply. Since other AI systems exhibit similar scaling-law economics, our mechanisms extend beyond computer vision, reinforcing that partial automation is often the economically rational long-run outcome, not merely a transitional phase.

Matthias Mertens, Adam Kuzee, Brittany S. Harris et al.

We propose that AI automation is a continuum between: (i) crashing waves where AI capabilities surge abruptly over small sets of tasks, and (ii) rising tides where the increase in AI capabilities is more continuous and broad-based. We test for these effects in preliminary evidence from an ongoing evaluation of AI capabilities across over 3,000 broad-based tasks derived from the U.S. Department of Labor O*NET categorization that are text-based and thus LLM-addressable. Based on more than 17,000 evaluations by workers from these jobs, we find little evidence of crashing waves (in contrast to recent work by METR), but substantial evidence that rising tides are the primary form of AI automation. AI performance is high and improving rapidly across a wide range of tasks. We estimate that, in 2024-Q2, AI models successfully complete tasks that take humans approximately 3-4 hours with about a 50% success rate, increasing to about 65% by 2025-Q3. If recent trends in AI capability growth persist, this pace of AI improvement implies that LLMs will be able to complete most text-related tasks with success rates of, on average, 80%-95% by 2029 at a minimally sufficient quality level. Achieving near-perfect success rates at this quality level or comparable success rates at superior quality would require several additional years. These AI capability improvements would impact the economy and labor market as organizations adopt AI, which could have a substantially longer timeline.

Anson Ho, Tamay Besiroglu, Ege Erdil et al. · harvard, stanford

We investigate the rate at which algorithms for pre-training language models have improved since the advent of deep learning. Using a dataset of over 200 language model evaluations on Wikitext and Penn Treebank spanning 2012-2023, we find that the compute required to reach a set performance threshold has halved approximately every 8 months, with a 95% confidence interval of around 5 to 14 months, substantially faster than hardware gains per Moore's Law. We estimate augmented scaling laws, which enable us to quantify algorithmic progress and determine the relative contributions of scaling models versus innovations in training algorithms. Despite the rapid pace of algorithmic progress and the development of new architectures such as the transformer, our analysis reveals that the increase in compute made an even larger contribution to overall performance improvements over this time period. Though limited by noisy benchmark data, our analysis quantifies the rapid progress in language modeling, shedding light on the relative contributions from compute and algorithms.

Hans Gundlach, Alex Fogelson, Jayson Lynch et al.

Algorithms have been estimated to increase AI training FLOP efficiency by a factor of 22,000 between 2012 and 2023 [Ho et al., 2024]. Running small-scale ablation experiments on key innovations from this time period, we are able to account for less than 10x of these gains. Surveying the broader literature, we estimate that additional innovations not included in our ablations account for less than 10x, yielding a total under 100x. This leads us to conduct scaling experiments, which reveal that much of this efficiency gap can be explained by algorithms with scale-dependent efficiency improvements. In particular, we conduct scaling experiments between LSTMs and Transformers, finding exponent differences in their compute-optimal scaling law while finding little scaling difference for many other innovations. These experiments demonstrate that - contrary to standard assumptions - an algorithm's efficiency gains are tied to compute scale. Using experimental extrapolation and literature estimates, we account for 6,930x efficiency gains over the same time period, with the scale-dependent LSTM-to-Transformer transition accounting for the majority of gains. Our results indicate that algorithmic progress for small models has been far slower than previously assumed, and that measures of algorithmic efficiency are strongly reference-dependent.

Jeffery Li, Jayson Lynch, Liva Olina et al.

In nearly every discipline, scientific computations are limited by the cost and speed of computation. For example, the best-known exact algorithms for the canonical Traveling Salesman Problem would take centuries to run on an instance of size 1 million. A natural response to such limits is to try to find new algorithms or to parallelize existing ones, but many algorithms are already at their theoretically-optimal level and parallelization is often impossible or prohibitively expensive. Starting in the 1960's, computer scientists pursued another solution: allowing solutions to have a small amount of error (i.e. approximating them). In this paper, we survey 118 of the most important algorithm problems in computer science, quantifying the gains and tradeoffs from approximation that have been discovered over the history of the field. Overall, only $\approx$20\% of problems have benefited from approximation. However, those with good approximate algorithms can be dramatically faster to compute with little cost to accuracy. For example, a quarter of computationally intractable problems (e.g. those that take exponential time to compute) have polynomial time approximate algorithms. Approximation also increases the number of algorithms that can run in linear time by 23\%, opening up new computational opportunities for those working in the big data regime. This work also sheds light on what should be expected from progress in AI, where approximation is at the heart of how deep learning works.

Hans Gundlach, Hrvoje Kukina, Jayson Lynch et al.

Quantum computing technology is advancing rapidly. Yet, even accounting for these trends, a quantum leap would be needed for quantum computers to meaningfully impact deep learning over the coming decade or two. We arrive at this conclusion based on a first-of-its-kind survey of quantum algorithms and how they match potential deep learning applications. This survey reveals three important areas where quantum computing could potentially accelerate deep learning, each of which faces a challenging roadblock to realizing its potential. First, quantum algorithms for matrix multiplication and other algorithms central to deep learning offer small theoretical improvements in the number of operations needed, but this advantage is overwhelmed on practical problem sizes by how slowly quantum computers do each operation. Second, some promising quantum algorithms depend on practical Quantum Random Access Memory (QRAM), which is underdeveloped. Finally, there are quantum algorithms that offer large theoretical advantages, but which are only applicable to special cases, limiting their practical benefits. In each of these areas, we support our arguments using quantitative forecasts of quantum advantage that build on the work by Choi et al. [2023] as well as new research on limitations and quantum hardware trends. Our analysis outlines the current scope of quantum deep learning and points to research directions that could lead to greater practical advances in the field.

Emanuele Del Sozzo, Martin Fleming, Kenneth Flamm et al.

Graphics Processing Units (GPUs) are the state-of-the-art architecture for essential tasks, ranging from rendering 2D/3D graphics to accelerating workloads in supercomputing centers and, of course, Artificial Intelligence (AI). As GPUs continue improving to satisfy ever-increasing performance demands, analyzing past and current progress becomes paramount in determining future constraints on scientific research. This is particularly compelling in the AI domain, where rapid technological advancements and fierce global competition have led the United States to recently implement export control regulations limiting international access to advanced AI chips. For this reason, this paper studies technical progress in NVIDIA datacenter GPUs released from the mid-2000s until today. Specifically, we compile a comprehensive dataset of datacenter NVIDIA GPUs comprising several features, ranging from computational performance to release price. Then, we examine trends in main GPU features and estimate progress indicators for per-memory bandwidth, per-dollar, and per-watt increase rates. Our main results identify doubling times of 1.44 and 1.69 years for FP16 and FP32 operations (without accounting for sparsity benefits), while FP64 doubling times range from 2.06 to 3.79 years. Off-chip memory size and bandwidth grew at slower rates than computing performance, doubling every 3.32 to 3.53 years. The release prices of datacenter GPUs have roughly doubled every 5.1 years, while their power consumption has approximately doubled every 16 years. Finally, we quantify the potential implications of current U.S. export control regulations in terms of the potential performance gaps that would result if implementation were assumed to be complete and successful. We find that recently proposed changes to export controls would shrink the potential performance gap from 23.6x to 3.54x.

Hans Gundlach, Jayson Lynch, Matthias Mertens et al.

Language models have seen enormous progress on advanced benchmarks in recent years, but much of this progress has only been possible by using more costly models. Benchmarks may therefore present a warped picture of progress in practical capabilities per dollar. To remedy this, we use data from Artificial Analysis and Epoch AI to form the largest dataset of current and historical prices to run benchmarks to date. We find that the price for a given level of benchmark performance has decreased remarkably fast, around $5\times$ to $10\times$ per year, for frontier models on knowledge, reasoning, math, and software engineering benchmarks. These reductions in the cost of AI inference are due to economic forces, hardware efficiency improvements, and algorithmic efficiency improvements. Isolating out open models to control for competition effects and dividing by hardware price declines, we estimate that algorithmic efficiency progress is around $3\times$ per year. Finally, we recommend that evaluators both publicize and take into account the price of benchmarking as an essential part of measuring the real-world impact of AI.

Ana Trišović, Alex Fogelson, Janakan Sivaloganathan et al.

We present the first large-scale analysis of AI foundation model usage in science - not just citations or keywords. We find that adoption has grown rapidly, at nearly-exponential rates, with the highest uptake in Linguistics, Computer Science, and Engineering. Vision models are the most used foundation models in science, although language models' share is growing. Open-weight models dominate. As AI builders increase the parameter counts of their models, scientists have followed suit but at a much slower rate: in 2013, the median foundation model built was 7.7x larger than the median one adopted in science, by 2024 this had jumped to 26x. We also present suggestive evidence that scientists' use of these smaller models may be limiting them from getting the full benefits of AI-enabled science, as papers that use larger models appear in higher-impact journals and accrue more citations.

Hans Gundlach, Jayson Lynch, Neil Thompson

The past decade has seen incredible scaling of AI systems by a few companies, leading to inequality in AI model performance. This paper argues that, contrary to prevailing intuition, the diminishing returns to compute scaling will lead to a convergence of AI model capabilities. In other words, meek models (those with limited computation budget) shall inherit the earth, approaching the performance level of the best models overall. We develop a model illustrating that under a fixed-distribution next-token objective, the marginal capability returns to raw compute shrink substantially. Given current scaling practices, we argue that these diminishing returns are strong enough that even companies that can scale their models exponentially faster than other organizations will eventually have little advantage in capabilities. As part of our argument, we give several reasons that proxies like training loss differences capture important capability measures using evidence from benchmark data and theoretical performance models. In addition, we analyze empirical data on the capability difference of AI models over time. Finally, in light of the increasing ability of meek models, we argue that AI strategy and policy require reexamination, and we outline the areas this shift will affect.