# Epoch AI

> Source: https://aiwiki.ai/wiki/epoch_ai
> Updated: 2026-08-02
> Fact-checked: 2026-08-02
> Categories: AI Companies, Model Evaluation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Epoch AI." aiwiki.ai, 2 Aug 2026. https://aiwiki.ai/wiki/epoch_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Epoch AI is a nonprofit research organization, founded in 2022 and directed by [Jaime Sevilla](https://aiwiki.ai/wiki/jaime_sevilla), that studies the trajectory of [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) through quantitative analysis of compute, data, algorithms, and economics. It describes itself as "a data-first research nonprofit investigating the future of artificial intelligence," and is widely cited as a neutral authority on how fast AI capabilities are advancing and what is driving that progress. [1][3] Its database of notable [machine learning models](https://aiwiki.ai/wiki/model) is the canonical public source for figures such as the "training compute doubling time," and its charts of long-run compute trends are reproduced throughout the field. Epoch AI also created the [FrontierMath](https://aiwiki.ai/wiki/frontiermath) benchmark of research-level mathematics problems, runs an AI Benchmarking Hub, publishes the Epoch Capabilities Index (ECI), a composite measure of [frontier model](https://aiwiki.ai/wiki/frontier_models) ability, and since late 2025 has tracked AI data centers and chip shipments through open databases built from satellite imagery, permits, and financial disclosures. [1][2][3][27][30]

This article concerns the research organization Epoch AI and should not be confused with the [epoch](https://aiwiki.ai/wiki/epoch) concept in [machine learning](https://aiwiki.ai/wiki/machine_learning), which refers to a single pass through a training dataset.

## What is Epoch AI?

Epoch AI is a data-first research nonprofit whose stated goal is to help people understand what is happening in AI "from a neutral perspective and grounded in the best possible evidence." [3] The organization organizes its work around three questions: the drivers of AI progress (training [compute](https://aiwiki.ai/wiki/compute), data, hardware, and scaling feasibility), the measurement of capability progress (benchmarks and indices), and the downstream impacts of AI (economics, adoption, and automation). [1][3]

Much of Epoch AI's influence comes from making its underlying datasets and methods public. It maintains open data explorers covering notable AI models, machine learning hardware, GPU clusters and data centers, and leading [AI companies](https://aiwiki.ai/wiki/companies), alongside reports, a newsletter, and benchmark results. Its data and analysis have been used by Stanford's AI Index, and the organization has had past and ongoing partnerships with bodies including the UK government's science and AI departments. [3][9]

## When was Epoch AI founded?

Epoch grew out of a volunteer effort begun around 2021, when [Jaime Sevilla](https://aiwiki.ai/wiki/jaime_sevilla) put out a call for collaborators to study historical trends in machine learning. The group's early analysis of training compute drew a strongly positive response, and Sevilla sought philanthropic funding to formalize the project. Epoch was founded in 2022, with Rethink Priorities acting as its initial fiscal sponsor. The founding team included Sevilla, Tamay Besiroglu, [Lennart Heim](https://aiwiki.ai/wiki/lennart_heim), Pablo Villalobos, Eduardo Infante-Roldan, [Marius Hobbhahn](https://aiwiki.ai/wiki/marius_hobbhahn), and Anson Ho. The work was carried out in close collaboration with organizations such as Rethink Priorities and the grantmaker [Open Philanthropy](https://aiwiki.ai/wiki/open_philanthropy), which became a major funder. [4][5][6]

Over the following years Epoch grew into an independent organization, rebranding as Epoch AI and spinning out from its fiscal sponsor to operate as an independent 501(c)(3) nonprofit in early 2025. According to its 2025 impact report, published on January 16, 2026, the organization had 21 full-time staff, published more than 100 outputs, raised about $10.3 million in 2025 (a roughly 40 percent increase over the prior year), and spent about $5 million. Its board of directors included Tom Davidson, Ajeya Cotra, Jaime Sevilla, and Maria de la Lama. [7]

The same report set out three priorities for 2026: expanding the organization's data programs, benchmark development and evaluations, and research and consultations. Epoch said it was seeking a further $3 million and stated that it could productively deploy roughly $10 million more than that. [7]

## What does Epoch AI research?

Epoch AI's best-known work is its empirical study of [AI compute](https://aiwiki.ai/wiki/ai_compute) and [scaling laws](https://aiwiki.ai/wiki/scaling_laws). The organization built a database of notable machine learning models, where a model is counted as "notable" if it set a state-of-the-art result on a recognized benchmark, was highly cited (on the order of 1,000 or more citations), had clear historical relevance, or saw significant real-world use. The expanded database includes 333 compute estimates for notable models released between 2010 and May 2024, up from 98 models in Epoch's original 2022 analysis, and it has continued to grow since. [2][8][10]

The dataset was originally assembled for the 2022 report "Compute Trends Across Three Eras of Machine Learning," which identified distinct historical regimes in how the compute used to train AI systems has grown. Epoch's later analysis found that the training compute of notable models grew about 4.1x per year (90 percent confidence interval: 3.7x to 4.6x) between 2010 and May 2024, while for frontier models, defined as the running top 10 by compute, the rate was about 5.3x per year (90 percent confidence interval: 4.9x to 5.7x). Epoch summarizes the headline trend as training compute growing "by 4-5x per year." These figures, and the accompanying charts, became the standard reference points for discussions of AI scaling. [2][8][11]

Beyond raw compute, Epoch AI has produced a series of influential reports on the inputs to AI progress, summarized below.

| Topic | Representative work | Key finding |
| --- | --- | --- |
| Training compute trends | "Compute Trends Across Three Eras of Machine Learning" (2022) | Notable-model training compute grew about 4.1x per year, and frontier compute about 5.3x per year, from 2010 to May 2024. [2][8][11] |
| Data limits | "Will We Run Out of Data?" (first released 2022, updated 2024) | The stock of public, human-generated text could be fully used for training at some point between roughly 2026 and 2032. [12][13] |
| Algorithmic progress | Studies of algorithmic efficiency in language models and image recognition | A large share of measured progress comes from improvements in algorithms and data efficiency, not compute alone. [14] |
| Hardware and clusters | Trends in machine learning hardware and AI supercomputers | Tracks GPU performance, cluster sizes, and the growth of large training systems. [9] |
| Physical infrastructure | Frontier Data Centers Hub (2025) and AI Chip Sales explorer (2026) | Tracks individual data center build-outs from satellite imagery and permits, and estimates the global installed stock of AI compute. [27][30] |
| Economics of scaling | GATE integrated assessment model (2025) | Models how compute investment and automation could drive rapid, even explosive, economic growth. [15] |

The 2024 update to "Will We Run Out of Data?" attracted broad attention for its projection that frontier models could exhaust the supply of high-quality, human-generated public text data around the middle of the decade, a finding widely covered in the press and tied to the rising interest in synthetic data. The report's authors stressed that the projection carried high uncertainty and that data efficiency gains and synthetic data could push the limit back. [12][13]

In 2025 Epoch AI released GATE (Growth and AI Transition Endogenous model), an integrated assessment model developed by Besiroglu, Heim, and Sevilla that links a compute-based model of AI development, an automation framework, and a semi-endogenous economic growth model. GATE is used to explore scenarios in which heavy reinvestment of output into AI hardware and research could produce very rapid growth, while quantifying the uncertainty around how much of the economy can be automated. [15]

## How does Epoch AI track AI data centers and chips?

Epoch announced its Frontier Data Centers Hub on November 4, 2025, an open database that tracks the construction, power draw, and compute capacity of the largest AI [data centers](https://aiwiki.ai/wiki/data_center). Sites are discovered from company announcements, news coverage, third-party databases, social media, and Epoch's own earlier work on [GPU clusters](https://aiwiki.ai/wiki/gpu_cluster), then researched using commercial satellite imagery (SkyWatch, Google Earth, and Sentinel-2), local permitting documents, and voluntary company disclosures, with cooling-equipment models and chip specifications used to estimate capacity and capital cost. Facilities generally qualify if they became operational in 2024 or later, and planned sites are included only once construction has begun. As of April 2026 Epoch estimated that the database covered roughly 27 percent of globally delivered AI compute capacity, weighted heavily toward the United States. [27][28]

The hub put Epoch into territory previously occupied mainly by commercial research firms. [SemiAnalysis](https://aiwiki.ai/wiki/semianalysis) sells a Datacenter Industry Model that tracks more than 5,000 data centers using property records, permits, power usage data, FOIA requests, and satellite images, with computer vision models used to process imagery at scale; it is an institutional subscription product sold to hyperscalers, semiconductor companies, and investors rather than a public dataset. Epoch's hub, by contrast, is free to browse and download, and its methodology documentation cites SemiAnalysis for one of its own inputs, a power usage effectiveness assumption of 1.35 for typical colocation facilities hosting AI accelerators. [28][29]

In early 2026 Epoch added an AI Chip Sales data explorer, built from financial reports and company disclosures, covering accelerators from NVIDIA, Google, Amazon, AMD, and Huawei. On its release Epoch reported that the global stock of AI compute had passed 16.5 million H100-equivalents. [30]

## What is FrontierMath?

[FrontierMath](https://aiwiki.ai/wiki/frontiermath) is a mathematics benchmark that Epoch AI announced on November 8, 2024. Its original release consisted of 300 problems (tiers 1 to 3), and the project later expanded with additional research-level material; the problems are original and unpublished, and intended to be far harder than earlier math benchmarks such as GSM8K and MATH, on which leading models had already reached near-perfect scores. The problems were created in collaboration with more than 60 mathematicians from universities across more than a dozen countries, and were designed to have answers that are automatically verifiable while still requiring deep, expert reasoning to produce. [16][17][18]

Problems span a wide range of difficulty. In the benchmark's main set, tiers run from advanced undergraduate material up through research-level mathematics, with the hardest tier (Tier 4) consisting of research-level problems; Epoch has separately maintained a collection of genuinely open research problems. At launch, frontier models performed extremely poorly: the FrontierMath paper reported that state-of-the-art systems, including OpenAI's o1-preview, GPT-4o, Anthropic's Claude 3.5 Sonnet, Google's Gemini 1.5 Pro, and xAI's Grok 2, each solved under 2 percent of the problems. The benchmark drew commentary from leading mathematicians, including Fields Medalists Terence Tao, Timothy Gowers, and Richard Borcherds, several of whom remarked on the difficulty of the questions. [16][17][18]

### The 2026 error correction (FrontierMath v2)

On June 12, 2026 Epoch published FrontierMath v2, a substantial revision of the existing problem set. The update corrected 123 problems in Tiers 1 to 3 and 12 problems in Tier 4, and removed a further 5 and 7 problems respectively, leaving 338 problems in total: 295 in the base set and 43 in the Tier 4 expansion set. Epoch stated that the update addressed errors in 42 percent of problems across the full dataset. Twelve problems are public, ten from Tiers 1 to 3 and two from Tier 4, and the previous version remains available for reference so that older scores can still be interpreted. [31][32]

### FrontierMath: Open Problems

Separately, Epoch began benchmarking models against mathematics that no one has solved. FrontierMath: Open Problems launched as a pilot in January 2026 with 14 genuinely unsolved problems, concentrated in areas such as combinatorics and number theory where candidate answers can be checked mechanically. Every problem ships with a bespoke verifier program, so a proposed solution can be confirmed without a human referee, and solving one would constitute a publishable mathematical result in its own right. The collection grew to 50 problems by July 31, 2026, spanning graph theory, Diophantine equations, algebraic number theory, analysis, arithmetic geometry, social choice theory, topology, knot theory, and spectral geometry; along the way Epoch added a Hadamard matrix problem of order 668 in February 2026, removed one problem in March 2026 that it judged would not meet its publishability bar, and removed three more in July 2026 over verifier fidelity or notability concerns. The work is supported by Schmidt Sciences. [30][33]

### What was the FrontierMath OpenAI funding controversy?

FrontierMath became the subject of controversy over how its funding was disclosed. On December 20, 2024, around the time OpenAI announced its o3 model and cited a strong FrontierMath score, Epoch AI revealed that OpenAI had funded the creation of the benchmark. Critics objected that this relationship had not been disclosed earlier, especially given that Epoch was otherwise known as an independent, largely Open Philanthropy funded organization. [19][20]

The dispute intensified after a contractor who had worked on the benchmark, posting on the forum LessWrong under the name "Meemi," wrote that many contributors had not been told of OpenAI's involvement, stating that "the communication about this has been non-transparent" and that "Epoch AI should have disclosed OpenAI funding, and contractors should have transparent information about the potential of their work being used for capabilities." A Stanford mathematics PhD student, Carina Hong, separately reported that several contributing mathematicians said they had been unaware that OpenAI would have access to the problems and would not have contributed had they known. [19][20]

Epoch AI's associate director Tamay Besiroglu acknowledged that the organization had made a mistake on transparency. He wrote: "We were restricted from disclosing the partnership until around the time o3 launched, and in hindsight we should have negotiated harder for the ability to be transparent to the benchmark contributors as soon as possible." He stated that OpenAI had access to the FrontierMath problems but had a verbal agreement not to train on them, and that Epoch maintained a separate, unseen holdout set to allow independent verification of results. Epoch's lead mathematician, Elliot Glazer, said his personal view was that OpenAI's reported score was legitimate, that is, that the company had not trained on the dataset, but that Epoch could not fully vouch for the figure until its own independent evaluation was complete. Epoch AI's 2025 impact report later described a research-level Tier 4 set of 50 problems as having been commissioned by OpenAI. [7][19][20]

Epoch's disclosure practice changed afterwards. The Open Problems benchmark carries an explicit statement on its own pages that "OpenAI funded the creation of the original FrontierMath: Tiers 1-4, but Open Problems is developed independently and owned solely by Epoch," alongside a note that OpenAI is at present the only entity to have purchased access to its verifiers. [33]

## What is MirrorCode?

MirrorCode is a long-horizon software engineering benchmark that Epoch AI developed jointly with the AI evaluations nonprofit [METR](https://aiwiki.ai/wiki/metr), which also provided grant support. Models are asked to reimplement an entire program end to end without seeing its source code, working only from execute-only access to the reference binary, its documentation, and a test suite of hundreds to thousands of cases; a submission counts as correct only when its outputs match the original exactly. Epoch published preliminary results on April 10, 2026 and the full results later in 2026, with a paper on arXiv (2606.30182) authored by Tom Adamczewski, David Owen, David Rein and colleagues. [34][35]

The benchmark comprises 25 target programs drawn from Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression, of which 22 have been released publicly as 132 task instances across six programming languages. Because the tasks are long, the evaluation allows far larger inference budgets than typical coding benchmarks: the preliminary report allowed up to one billion tokens per task, which it costed at roughly $550 per run. [34][35]

The headline finding was that some multi-week engineering tasks are already within reach of frontier systems. Claude Opus 4.7 scored 56 percent on the full benchmark and reimplemented gotree, a bioinformatics toolkit of roughly 16,000 lines of Go with more than 40 commands, in 14 hours at a cost of $251; Epoch estimated the same work would take a skilled human engineer between two and seventeen weeks. [35]

## How does Epoch AI rank AI models? (ECI and the Benchmarking Hub)

Early in 2025 Epoch AI relaunched an AI Benchmarking Hub, which collects evaluation results reported by model developers and third parties alongside benchmarks that Epoch runs itself. The Hub became one of the organization's most visited pages, and it underpins Epoch's flagship capability metric. [3][7][21]

That metric is the Epoch Capabilities Index (ECI), a composite score that combines results from dozens of distinct benchmarks into a single "general capability" scale, allowing models to be compared even across periods long enough for any one benchmark to saturate. Epoch likens the ECI to an IQ-style measure: rather than tracking performance on a single skill, it aims to capture a broad underlying capability. The index is built on item response theory, the statistical framework used in standardized testing, and functions as a relative measure similar to an Elo rating, jointly estimating both how capable each model is and how difficult each benchmark is from the pattern of results when models are tested on overlapping benchmarks. By late 2025 the ECI drew on more than a thousand evaluations covering on the order of 147 models and roughly 39 underlying benchmarks. Epoch describes the ECI as an independent product over which it has full rights, while noting that the work built on methodology from a Google DeepMind paper, "A Rosetta Stone for AI Benchmarks." [21][22][23]

By mid-2026 the index combined results from more than 50 benchmarks, weighting each benchmark by how much signal it carries, and required a model to have at least four benchmark evaluations before it is scored. The ECI scale is arbitrary but anchored: it is calibrated so that Claude 3.5 Sonnet sits at 130 and GPT-5 at 150, and it is intended to be linear, so that a 10-point gain represents a comparable amount of capability progress wherever it occurs on the scale. Epoch also publishes domain-specific variants covering mathematics, software engineering, and cyber capabilities. [21]

## Who runs Epoch AI, and how is it regarded?

Jaime Sevilla, a Spanish researcher with a background in mathematics and computer science, founded Epoch and serves as its director. He has become a prominent voice on [AI forecasting](https://aiwiki.ai/wiki/ai_forecasting) and the trajectory of [transformative AI](https://aiwiki.ai/wiki/transformative_ai), and was profiled by TIME in 2024 in connection with Epoch's trend analysis. Tamay Besiroglu was a co-founder and the organization's associate director, leading much of its work on compute and economics. [4][24]

In April 2025, Besiroglu left Epoch AI to co-found [Mechanize](https://aiwiki.ai/wiki/mechanize), a startup whose stated aim is "the full automation of the economy," beginning with white-collar work. Two other Epoch-affiliated researchers, Ege Erdil and Matthew Barnett, joined him. The move generated controversy and public criticism, because Mechanize's goal of automating human labor was seen as standing in tension with the safety-oriented and cautious framing associated with Epoch and parts of the [AI safety](https://aiwiki.ai/wiki/ai_safety) community; Mechanize reported raising a seed round of about $7.3 million. Lennart Heim, another co-founder, later led the compute team at the RAND Corporation's center on AI and emerging technology. [25][26]

Epoch AI is generally regarded across the field as an authoritative and relatively neutral source on quantitative AI trends, and its datasets are widely reused by researchers, journalists, and policymakers. Its analyses are frequently cited in coverage of AI scaling, compute, and the data supply. At the same time, the FrontierMath funding episode prompted broader debate about the independence of benchmark organizations and the importance of disclosing industry funding and data access. Epoch's roots in the effective altruism and AI-safety funding ecosystem, and the later departure of several staff to build automation technology, have also featured in commentary about the organization's positioning. [1][19][25]

The 2026 FrontierMath revision cut in two directions for that reputation. Publishing the scale of the problem, and republishing the corrected set alongside the original, is the kind of disclosure that benchmark maintainers are often criticized for avoiding; at the same time, an error rate of 42 percent in a flagship benchmark that had been used to argue about frontier model progress is a substantial defect in a widely cited dataset, and it applies to scores reported in the eighteen months before the correction. [31][32]

## See also

- [Stanford HAI AI Index Report](https://aiwiki.ai/wiki/ai_index_report)
- [Future of Life Institute AI Safety Index](https://aiwiki.ai/wiki/fli_ai_safety_index)

## References

1. "Epoch AI." Epoch AI. Accessed June 9, 2026. https://epoch.ai/
2. "Compute Trends Across Three Eras of Machine Learning." Epoch AI. https://epoch.ai/blog/compute-trends
3. "About." Epoch AI. Accessed June 9, 2026. https://epoch.ai/about
4. Naughton, John, and TIME staff. "The Epoch AI Researcher Trying to Glimpse the Future of AI." TIME, 2024. https://time.com/6985850/jaime-sevilla-epoch-ai/
5. "Announcing Epoch: A research organization investigating the road to Transformative AI." EA Forum, 2022. https://forum.effectivealtruism.org/posts/zqRDNChFburJMmpqK/announcing-epoch-a-research-organization-investigating-the
6. "Announcing Epoch AI: A research initiative investigating the road to transformative AI." Epoch AI. https://epoch.ai/latest/announcing-epoch
7. "Epoch AI 2025 impact report." Epoch AI, January 16, 2026. https://epoch.ai/blog/epoch-impact-report-2025
8. "Training compute of frontier AI models grows by 4-5x per year." Epoch AI. https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year
9. "Trends in AI supercomputers." Epoch AI. https://epoch.ai/blog/trends-in-ai-supercomputers
10. "Data on Notable AI Models." Epoch AI. https://epoch.ai/data/notable-ai-models
11. Sevilla, Jaime, et al. "Compute Trends Across Three Eras of Machine Learning." arXiv:2202.05924, 2022. https://arxiv.org/abs/2202.05924
12. "Will we run out of data? Limits of LLM scaling based on human-generated data." Epoch AI. https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
13. Villalobos, Pablo, et al. "Will we run out of data? Limits of LLM scaling based on human-generated data." arXiv:2211.04325. https://arxiv.org/abs/2211.04325
14. "The least understood driver of AI progress." Epoch AI / Anson Ho. https://epoch.ai/gradient-updates/the-least-understood-driver-of-ai-progress
15. "GATE: Modeling the trajectory of AI and automation." Epoch AI. https://epoch.ai/blog/announcing-gate
16. "FrontierMath: Evaluating advanced mathematical reasoning in AI." Epoch AI. https://epoch.ai/frontiermath
17. "FrontierMath." Epoch AI (benchmark overview). https://epoch.ai/frontiermath/tiers-1-4/about
18. Glazer, Elliot, et al. "FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI." arXiv:2411.04872, November 2024. https://arxiv.org/abs/2411.04872
19. Wiggers, Kyle. "AI benchmarking organization criticized for waiting to disclose funding from OpenAI." TechCrunch, January 19, 2025. https://techcrunch.com/2025/01/19/ai-benchmarking-organization-criticized-for-waiting-to-disclose-funding-from-openai/
20. "OpenAI's FrontierMath Fiasco." Fortune, January 21, 2025. https://fortune.com/2025/01/21/eye-on-ai-openai-o3-math-benchmark-frontiermath-epoch-altman-trump-biden/
21. "Epoch Capabilities Index." Epoch AI. Accessed August 1, 2026. https://epoch.ai/benchmarks/eci
22. "Epoch's Capabilities Index stitches together benchmarks across a wide range of difficulties." Epoch AI. https://epoch.ai/data-insights/interpreting-eci
23. "ECI Documentation - Overview." Epoch AI. https://epoch.ai/data/eci-documentation
24. "Jaime Sevilla." Epoch AI team page. https://epoch.ai/about/team/jaime-sevilla
25. "Epoch AI alumni launch Mechanize to 'automate the whole economy.'" EA Forum, April 2025. https://forum.effectivealtruism.org/posts/HqKnreqC3EFF9YcEs/epoch-ai-alumni-launch-mechanize-to-automate-the-whole
26. "A tech founder is getting skewered online after announcing his startup aims to replace all human workers with AI." Fortune, April 2025. https://fortune.com/article/tech-founder-online-epoch-ai-mechanize-tamay-besiroglu-automated-employees-workforce/
27. "Introducing the Frontier Data Centers Hub." Epoch AI, November 4, 2025. https://epoch.ai/blog/introducing-the-frontier-data-centers-hub/
28. "Frontier Data Centers Documentation - Methodology." Epoch AI. Accessed August 1, 2026. https://epoch.ai/data/data-centers-documentation/methodology
29. "Datacenter Industry Model." SemiAnalysis. Accessed August 1, 2026. https://semianalysis.com/datacenter-industry-model/
30. "The Epoch AI Brief - January 2026." Epoch AI. https://epochai.substack.com/p/the-epoch-ai-brief-january-2026
31. "FrontierMath Tiers 1-3 (v2)." Epoch AI. Accessed August 1, 2026. https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2
32. "FrontierMath Tier 4 (v2)." Epoch AI. Accessed August 1, 2026. https://epoch.ai/benchmarks/frontiermath-tier-4-v2
33. "FrontierMath: Open Problems." Epoch AI. Accessed August 1, 2026. https://epoch.ai/frontiermath/open-problems
34. "MirrorCode: Evidence AI can already do some weeks-long coding tasks." Epoch AI, April 10, 2026. https://epoch.ai/publications/mirrorcode-preliminary-results
35. "MirrorCode: What's the largest software project AI can complete on its own?" Epoch AI; arXiv:2606.30182. Accessed August 1, 2026. https://epoch.ai/MirrorCode