Pith. sign in

REVIEW 4 major objections 4 minor 30 references

Jolting Technologies: Superexponential Acceleration in AI Capabilities and Implications for AGI

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A sustained positive jolt—the normalized third derivative of a capability curve—would show AI progress is superexponential, not merely exponential, and a hybrid detector can flag it in noisy time series.

desk verdict The central definition classifies ordinary exponential growth as jolting, so the paper's core claim collapses; the rest is a plan. read the letter →

arxiv 2507.06398 v1 pith:6ZKETLHG submitted 2025-07-08 cs.AI cs.CY

David Orban This is my paper
classification cs.AIcs.CY
keywords superexponentialgrowththirdderivativejoltdetectionAGItimelinesMonteCarlosimulationAIbenchmarksgovernancecapabilityforecasting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper makes the case that AI capability growth can be superexponential in a precise sense: the third derivative of a capability curve $C(t)$ is positive, not just the first and second. It defines this as a 'jolt,' normalized as $J(t)=C'''(t)/C(t)>0$, and argues that standard exponential extrapolations miss such regimes, which would shorten AGI timelines and make capability jumps abrupt. Because real longitudinal benchmark data are not yet sufficient for third-derivative analysis, the paper validates detection methods on Monte Carlo simulations, reporting true positive rates of 0.85 to 0.95 with false positive rates of 0.05 to 0.15 depending on noise. A sympathetic reader would take away a testable monitoring tool and a warning that governance built on reaction time may be outdated if jolts are real.

What carries the argument

The load-bearing object is the normalized jolt magnitude $J(t)=\frac{C'''(t)}{C(t)}$, together with the dimensionless form $J_N(t)=\frac{C'''(t)C(t)}{C'(t)C''(t)}$. The argument treats a sustained positive value of $J$ as the signature of a jolting regime, and the hybrid detector (peak-ratio analysis, pattern matching, duration metrics) as the practical instrument for finding that signature in noisy data. Supporting machinery includes the composite jolt model $C'''_{\mathrm{total}}(t)=\sum_i w_i C_i'''(t)+\sum_{i\ne j} I_{ij}(t)$ for interacting advances and the resource-damped effective jolt $J_{\mathrm{effective}}(t)=J(t)\, (R_{\mathrm{max}}-R(t))/R_{\mathrm{max}}$.

What would settle it

Run the hybrid detector on a pure exponential curve with added noise of the same amplitude as the reported simulations: if the false positive rate for detecting 'jolts' approaches the true positive rates in Table 1, the detector is not selective and the central evidence fails. Alternatively, on real benchmark series, vary the smoothing window width and polynomial order; if sustained $C'''(t)>0$ appears and disappears or changes sign with those choices, the jolt signal is a measurement artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that a sustained $J(t)>0$ is quantitative evidence that a technology's capacity to improve is itself accelerating, a regime qualitatively different from exponential growth where acceleration is constant or proportional to velocity. It proposes a hybrid jolt detector, combining peak-ratio analysis, pattern matching, and duration criteria, and claims Monte Carlo validation shows it identifies true jolts in synthetic time series with true positive rates of 0.95 at low noise, 0.92 at medium noise, and 0.85 at high noise, with corresponding false positives of 0.05, 0.08, and 0.15. It also models jolts as composites of interacting subfield improvements and as dampened by resource limits. The paper explicitly stops short of claiming the hypothesis is empirically confirmed, stating that validation on real AI benchmarks awaits suitable longitudinal data.

Load-bearing premise

The load-bearing premise is that AI capability can be represented as one smooth, differentiable scalar function built from benchmarks such as MMLU and ImageNet, and that its third derivative is a stable signal rather than an artifact of the smoothing choices used to compute it.

Editorial extensions

If this is right

  • If jolting regimes are real, AGI forecasts built on exponential extrapolation will systematically underestimate near-term capability growth and miss abrupt phase transitions.
  • The hybrid detector gives a concrete monitoring procedure: fit a smoothed curve to benchmark series, estimate third derivatives, and flag sustained $J(t)>0$ with known false positive costs.
  • Shrinking capability doubling times, when the relative growth rate $C'(t)/C(t)$ is rising, become an observable early warning that a jolt may be underway.
  • Because jolts can be composite, with interactions across hardware, algorithms, and data, aggregate metrics may hide local jolts; monitoring subfield series separately would be needed.
  • Governance mechanisms that are reactive and slow will be structurally out of phase with a jolting technology, motivating continuous monitoring, sunset clauses, and rapid response mechanisms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the third derivative of benchmark scores is highly sensitive to smoothing choices; a natural extension is to test whether detected jolts survive changes in smoothing window width or polynomial order, and if not, the signal is a measurement artifact.
  • Beyond the paper, the reported false positive rate of 0.05 to 0.15 at high noise suggests a single flagged jolt should not trigger policy action; a sequential or Bayesian detector raising the bar over multiple benchmarks would be a more conservative extension.
  • Beyond the paper, back-testing the detector on historical series with known qualitative jumps would give a direct out-of-sample check of whether third-derivative signals lead, coincide with, or lag actual capability transitions.
  • Beyond the paper, the resource-damping term implies jolts are transient; if real, each detected jolt should be followed by saturation or deceleration, making the framework testable over a longer window.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a "Jolting Technologies Hypothesis" (JTH), defining a jolt as a positive normalized third derivative J(t)=C'''(t)/C(t) of a capability metric C(t). It claims that sustained J(t)>0 identifies superexponential ("jolting") growth distinct from ordinary exponential growth, and it presents a hybrid jolt detector whose Monte Carlo simulation reports true positive rates between 0.85 and 0.95 (Table 1). The paper also discusses composite jolt models, resource constraints, and implications for AGI timelines and governance. The manuscript explicitly acknowledges that empirical validation with real-world data remains future work, and it provides a public repository with code and simulation data.

Significance. If the central mathematical criterion were valid, a detector for superexponential regime changes would be a useful monitoring and forecasting tool for AI capabilities and AGI timelines. The paper also has a praiseworthy feature: it ships code, data, and a repository that would make any subsequent empirical study reproducible. However, the paper's load-bearing definitional claim fails: the proposed criterion J(t)>0 does not distinguish exponential from superexponential growth, and Section 2.2 contains an internal contradiction about shrinking doubling times. Because the core mathematical foundation is incorrect, the reported detector validation does not establish the existence or detectability of jolts in any meaningful sense.

major comments (4)
  1. [Section 2.1, Eq. (1) and Eq. (2)] The definition of a jolt as J(t)=C'''(t)/C(t)>0 fails to distinguish jolting growth from ordinary exponential growth. For C(t)=e^{kt} with k>0, we have C'''(t)=k^3 e^{kt}, so J(t)=k^3>0, and the normalized jolt in Eq. (2) is J_N(t)=1 for every pure exponential trajectory. Thus the paper's central criterion labels the baseline exponential model as jolting, directly contradicting the claim in Section 2.1 that this is "mathematically distinct" from exponential growth. The Monte Carlo validation cannot rescue this unless the detector explicitly thresholds against an exponential baseline, which is not described.
  2. [Section 2.2] Section 2.2 contains a direct contradiction. The first paragraph correctly states that C'''(t)>0 is necessary but not sufficient for decreasing doubling times, and that shrinking doubling times require alpha'(t)>0. The very next paragraph then asserts that in a jolting system the increasing acceleration C''(t) implies alpha'(t)>0 and hence systematically decreasing doubling times. This implication is false; for example, C(t)=t^3 has C'''(t)=6>0 and C''(t)=6t increasing, but alpha(t)=C'(t)/C(t)=3/t is decreasing, so doubling times grow. This internal inconsistency undermines the claimed connection between jolts and doubling-time compression.
  3. [Section 3.2, Table 1] The Monte Carlo validation is circular and does not test the claim that the detector identifies jolts as distinct from exponential growth. The manuscript states that synthetic trajectories were generated "designed to exhibit exponential, logistic, and jolting growth patterns" (Section 4.2), but it never specifies the generative models, the detector's decision rule, or how the detector is prevented from flagging pure exponentials as jolting. Since every exponential has J(t)>0, the high true positive rates in Table 1 are consistent with a detector that simply ignores the difference between the classes. Without a negative-control experiment on exponential-only series and a description of the detector's thresholding, the reported true positive rates are uninformative.
  4. [Sections 3.1 and 3.3] The empirical analysis is presented entirely in a conditional voice: the text repeatedly says that smoothing "would be applied," a regression model "would then be fitted," and derivative estimates "would be performed." No actual benchmark data are analyzed, no fitted curves are shown, and no third-derivative estimates are reported. The paper's abstract and conclusion acknowledge this, but the framing of Section 3 as "Empirical Evidence and Results" is misleading; the section contains a proposed methodology and a simulation, not empirical evidence for the hypothesis.
minor comments (4)
  1. [Section 2.3, Eq. (4) and Section 4.2] Section 4.2 refers to a "correlation coefficient" in Eq. (4), but Eq. (4) defines a weighted sum of individual jolts and interaction terms; no correlation coefficient appears in that equation. The interaction terms I_{ij}(t) are never given a functional form, which limits the usefulness of the composite model.
  2. [References] Reference [16] is a Manifund project page rather than a methodological reference for jolt quantification; the paper cites it as the source for the normalized jolt J_N(t), but the connection is not explained.
  3. [Figure 1] Figure 1 is described in the text as a heatmap of error rates, but the figure itself is not included in the manuscript; only a generic caption appears. The reader cannot assess the claimed optimal parameter regions without the figure or a link to the repository.
  4. [Throughout] The text contains several typographical and formatting issues, including missing spaces (e.g., "theJolting", "synthesiszingexternal,verifiableinformation") and inconsistent use of "jolt" versus "jerk" for the third derivative, which should be cleaned up in revision.

Circularity Check

2 steps flagged · score 6.0 of 10

Jolt criterion is definitional and the Monte Carlo 'validation' re-detects jolts inserted by construction into synthetic trajectories; claimed support reduces to definitions.

  1. self definitional [Section 2.1, after Eq. (1)]
    "In this paper, 'superexponential growth' specifically refers to growth characterized by a positive third derivative (C'''(t) > 0), rather than the colloquial usage of 'faster-than-exponential.' ... A sustained J(t) > 0 provides quantitative evidence for a system operating in a jolting regime, where its capacity to improve is itself improving at an accelerating pace."

    The 'evidence' is the definition restated. J(t)>0 is equivalent to C'''(t)>0 divided by positive C(t), and the paper has just defined 'superexponential/jolting' as C'''(t)>0. Thus 'sustained J>0 is quantitative evidence for a jolting regime' reduces to 'C'''>0 means C'''>0.' Moreover, the criterion cannot do its claimed work against the exponential baseline: for C(t)=e^{kt}, C'''(t)=k^3 e^{kt} and J(t)=k^3>0 for every positive k, so standard exponential growth also satisfies the defining condition. No independent evidence is introduced.

  2. fitted input called prediction [Section 3.2 and Section 4.2, Table 1]
    "Our Monte Carlo simulations, designed to validate our jolt detection methodologies, provide key insights. The hybrid jolt detector ... demonstrated robust performance in identifying true jolts within synthetically generated time series data across various noise levels. ... We generate synthetic agent performance trajectories designed to exhibit exponential, logistic, and jolting growth patterns."

    The synthetic trajectories are generated to contain the target property (jolting = positive third derivative), and the detector's 'true positive rate' measures how often it recovers the jolts that were inserted. This validates numerical implementation and internal consistency of definitions, not the Jolting Technologies Hypothesis about real AI: high true-positive rates are forced by the construction whenever the derivative estimates work. The later conclusion in Section 5 that the 'empirical explorations ... support the Jolting Technologies Hypothesis' therefore rests on simulated data built from the hypothesis's own definition.

full rationale

The paper is unusually candid in Section 6 that 'empirical validation with real-world data remains essential,' and there is no load-bearing self-citation chain: the author's repository [22] supplies code and data, not the central assumptions. Nevertheless, the claimed support for the Jolt Hypothesis is circular in two connected places. First, the hypothesis is stipulated as C'''(t)>0, and the same inequality is then presented as quantitative evidence of a jolting regime; since C(t)=e^{kt} gives J(t)=k^3>0, the criterion does not even separate the alleged jolt regime from the exponential baseline it is meant to oppose. Second, the only concrete 'validation' is a Monte Carlo study in which jolting trajectories are generated by construction and then detected; the reported true-positive rates are therefore a self-consistency check, not independent confirmation. I also note a non-circular mathematical error in Section 2.2: the claim that increasing acceleration C''(t) implies alpha'(t)>0 is false (C(t)=t^3 has C''(t)=6t increasing but alpha(t)=3/t decreasing). That is a correctness problem, not a circularity, so it does not itself raise the circularity score. Overall, the central evidentiary chain reduces to definitions and self-generated simulation inputs, but because the paper explicitly declines to claim real-world empirical validation, the circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

This ledger reflects the paper's heavy reliance on definitional and simulation-based inputs. The central jolt measure is a renamed third derivative, so no physical entity is added; the only new objects are a naming convention and an incompletely specified detector. The free parameters that matter are the detector's hyperparameters, never reported, and the composite/resource model coefficients that are never estimated.

free parameters (3)
  • Detector hyperparameters (peak ratio threshold, pattern matching window, duration threshold) = not reported
    Used in the Monte Carlo validation of the hybrid jolt detector; without the values, the simulation results in Table 1 cannot be reproduced or audited.
  • Composite jolt weights w_i = not estimated
    Equation (4) allows arbitrary weights w_i for individual jolt contributions; no estimation procedure or data are given, making the model unfalsifiable as stated.
  • Resource limit R_max = not estimated
    Equation (5) introduces a maximum sustainable resource level R_max that modulates the jolt; no operational definition or measurement is provided.
assumptions (4)
  • domain assumption AI capability can be represented by a smooth, twice-differentiable scalar function C(t)
    Invoked in Section 2.1 to define velocity, acceleration, and jolt; real benchmark data are sparse, noisy, saturating, and only loosely comparable across eras.
  • ad hoc to paper A positive third derivative is an appropriate indicator of superexponential growth
    The paper defines superexponential growth as C'''(t)>0 in Section 2.1, while other definitions (e.g., faster-than-exponential like t*e^kt) are explicitly excluded; this is a modeling choice, not a standard result.
  • domain assumption Synthetic trajectories generated for the Monte Carlo study are representative vehicles for testing jolt detection
    Section 4.2 states that exponential, logistic, and jolting trajectories were simulated; the paper assumes that such simulations capture the relevant failure modes of real AI capability series.
  • ad hoc to paper Multi-factor progress can be decomposed into additive weighted jolts plus interaction terms
    Equation (4) is stated in Section 2.3 without derivation or empirical support; it is a generic decomposition that cannot be refuted by the data presented.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Jolting Technologies: Superexponential Acceleration in AI Capabilities and Implications for AGI." pith.science (2026). https://pith.science/paper/6ZKETLHG

@misc{pith2026250706398,
  author       = {Pith},
  title        = {Pith review of: Jolting Technologies: Superexponential Acceleration in AI Capabilities and Implications for AGI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZKETLHG}},
  note         = {Machine review of arXiv:2507.06398}
}
read the original abstract

This paper investigates the Jolting Technologies Hypothesis, which posits superexponential growth (increasing acceleration, or a positive third derivative) in the development of AI capabilities. We develop a theoretical framework and validate detection methodologies through Monte Carlo simulations, while acknowledging that empirical validation awaits suitable longitudinal data. Our analysis focuses on creating robust tools for future empirical studies and exploring the potential implications should the hypothesis prove valid. The study examines how factors such as shrinking idea-to-action intervals and compounding iterative AI improvements drive this jolting pattern. By formalizing jolt dynamics and validating detection methods through simulation, this work provides the mathematical foundation necessary for understanding potential AI trajectories and their consequences for AGI emergence, offering insights for research and policy.

Figures

Figures reproduced from arXiv: 2507.06398 by the authors.

Figure 1
Figure 1. Heatmap illustrating error rates of the hybrid jolt detector from Monte Carlo simulations. Darker regions indicate lower error rates. 3.3 Visualization of Empirical Findings The results of the quantitative analysis would be presented through a series of visualizations for each benchmark: – Plots showing raw data points, the fitted capability curve C(t), and its con￾fidence intervals. – Separate plots illustrating th… view at source ↗
Figure 2
Figure 2. Visualization of extreme jolt scenario in AgentBench simulation. (Further de￾tails in repository [22]). 4.4 Implications for Agent Development and Safety Real-world AI agent capabilities exhibiting jolting patterns, have significant im￾plications. Rapid, unanticipated increases in agent capabilities pose safety chal- [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    https://ai-2027.com (2025)

    AI-2027.com: AI 2027: What Superintelligence Looks Like. https://ai-2027.com (2025)

    work page 2025

  2. [2]

    Report, UC Berkeley (2020)

    Center for Long-Term Cybersecurity (CLTC), UC Berkeley: Decision Points in AI Governance. Report, UC Berkeley (2020)

    work page 2020

  3. [3]

    Journal of Consciousness Studies 17(9-10), 7–65 (2010)

    Chalmers, D.J.: The Singularity: A Philosophical Analysis. Journal of Consciousness Studies 17(9-10), 7–65 (2010)

    work page 2010

  4. [4]

    University of Pennsylvania (2007)

    Diebold, F.X.: Elements of Forecasting. University of Pennsylvania (2007)

    work page 2007

  5. [5]

    arXiv preprint arXiv:2407.07890 (2024)

    Dominguez-Olmedo, R., Dorner, F.E., Hardt, M.: Training on the Test Task Confounds Evaluation and Emergence. arXiv preprint arXiv:2407.07890 (2024)

    arXiv 2024

  6. [6]

    Epoch AI: Compute trends database (2023), https://epochai.org/data/compute-trends, accessed: 2025-01-07

    work page 2023

  7. [7]

    https://www.mdpi.com/2078-2489/6/1/3/pdf (2023)

    Ferrara, E.: Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies. https://www.mdpi.com/2078-2489/6/1/3/pdf (2023)

    work page 2023

  8. [8]

    https://www.cnbc.com/2024/06/04/ elon-musk-told-nvidia-to-ship-ai-chips-reserved-for-tesla-to-x-xai.html (2024)

    Kolodny, L.: Elon Musk told Nvidia to ship AI chips reserved for Tesla to X, xAI. https://www.cnbc.com/2024/06/04/ elon-musk-told-nvidia-to-ship-ai-chips-reserved-for-tesla-to-x-xai.html (2024)

    work page 2024

Show all 30 references
  1. [9]

    The Review of Economics and Statistics 17(6), 105–115 (1935), originally published in German in 1926

    Kondratiev, N.D.: The long waves in economic life. The Review of Economics and Statistics 17(6), 105–115 (1935), originally published in German in 1926

    1935

  2. [10]

    University of Chicago Press, Chicago (1962)

    Kuhn, T.S.: The Structure of Scientific Revolutions. University of Chicago Press, Chicago (1962)

    1962

  3. [11]

    https://www.writingsbyraykurzweil.com/the-law-of-accelerating-returns (2001)

    Kurzweil, R.: The Law of Accelerating Returns. https://www.writingsbyraykurzweil.com/the-law-of-accelerating-returns (2001)

    2001

  4. [12]

    arXiv preprint arXiv:2503.14499 (2025)

    Kwa, T., Liu, Y., Shen, J., Andreassen, A., Jones, D., Chen, Y., Antonov, G., Kumar, A., Demszky, D., Eason, M., et al.: Measuring AI Ability to Complete Long Tasks. arXiv preprint arXiv:2503.14499 (2025)

    2025 arXiv

  5. [13]

    Harvard Data Science Review (2024)

    Leslie, D., Perini, A.M.: Future Shock: Generative AI and the International AI Policy and Governance Crisis. Harvard Data Science Review (2024)

    2024

  6. [14]

    arXiv preprint arXiv:2308.03688 (2023)

    Liu, G., Yuan, H., Li, C., Wang, P., Zhang, S., Shen, Y., Zhao, R., Li, L., Liu, Z., Tang, Z., et al.: AgentBench: Evaluating LLMs as Agents. arXiv preprint arXiv:2308.03688 (2023)

    2023 arXiv

  7. [15]

    https://hai.stanford.edu/news/ai-benchmarks-hit-saturation (April 2023)

    Lynch, S.: AI Benchmarks Hit Saturation. https://hai.stanford.edu/news/ai-benchmarks-hit-saturation (April 2023)

    2023

  8. [16]

    https:// manifund.org/projects/moral-progress-in-ai-to-prevent-premature-value-lock-in (2023)

    Manifund: Moral Progress in AI to Prevent Premature Value Lock-in. https:// manifund.org/projects/moral-progress-in-ai-to-prevent-premature-value-lock-in (2023)

    2023

  9. [17]

    Question opened January 23, 2020

    Metaculus Community: Date of Artificial General Intelligence? https: //www.metaculus.com/questions/5121/date-of-artificial-general-intelligence/ (2020), accessed: 2025-05-08. Question opened January 23, 2020

    2020

  10. [18]

    METR Blog, https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ (March 2025)

    METR: Measuring AI Ability to Complete Long Tasks. METR Blog, https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ (March 2025)

    2025

  11. [19]

    https://nvidianews.nvidia.com/news/ nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing (March 2024)

    NVIDIA Newsroom: NVIDIA Blackwell Platform Arrives to Power a New Era of Computing. https://nvidianews.nvidia.com/news/ nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing (March 2024)

    2024

  12. [20]

    https://arxiv.org/abs/2303.08774 (2023) Jolting Technologies: Superexponential Acceleration in AI 13

    OpenAI: GPT-4 Technical Report. https://arxiv.org/abs/2303.08774 (2023) Jolting Technologies: Superexponential Acceleration in AI 13

    2023 arXiv

  13. [21]

    https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo (2024)

    OpenAI: GPT-4 and GPT-4 Turbo. https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo (2024)

    2024

  14. [22]

    Jolting Technologies: Superexponential Acceleration in AI Capabilities and Implications for AGI

    Orban, D.: Jolting Technologies: Code and Data Repository. https://github.com/davidorban/jolting (2025), accessed: 2025-07-07. Supporting code, data, and supplementary materials for the analyses presented in "Jolting Technologies: Superexponential Acceleration in AI Capabiliti...

    2025

  15. [23]

    Journal of Machine Learning Research12, 2825–2830 (2011)

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine Learning in Python. Journal of Mach...

    2011

  16. [24]

    https://www.computerweekly.com/news/366615931/ Microsoft-Ignite-AI-capabilities-double-every-six-months (2024)

    Saran, C.: Microsoft Ignite: AI capabilities double every six months. https://www.computerweekly.com/news/366615931/ Microsoft-Ignite-AI-capabilities-double-every-six-months (2024)

    2024

  17. [25]

    arXiv preprint arXiv:2202.05924 (2022)

    Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., Villalobos, P.: Compute trends across three eras of machine learning. arXiv preprint arXiv:2202.05924 (2022)

    2022 arXiv

  18. [26]

    Stanford Institute for Human-Centered Artificial Intelligence (HAI): Artificial Intelligence Index Report 2024. Tech. rep., Stanford University (2024)

    2024

  19. [27]

    Stanford Institute for Human-Centered Artificial Intelligence (HAI): Artificial Intelligence Index Report 2025. Tech. rep., Stanford University (2025)

    2025

  20. [28]

    https://www.datacenterdynamics.com/en/news/ xai-to-double-colossus-compute-capacity-reveals-cluster-uses-nvidia-spectrum-x-ethernet/ (2024)

    Trueman, C.: xAI to double Colossus compute capacity, reveals cluster uses Nvidia Spectrum-X Ethernet. https://www.datacenterdynamics.com/en/news/ xai-to-double-colossus-compute-capacity-reveals-cluster-uses-nvidia-spectrum-x-ethernet/ (2024)

    2024

  21. [29]

    arXiv preprint arXiv:2407.16376 (2024)

    Tsaknaki, I.Y., Lillo, F., Mazzarisi, P.: Bayesian Autoregressive Online Change-Point Detection with Time-Varying Parameters. arXiv preprint arXiv:2407.16376 (2024)

    2024

  22. [30]

    Information 9(12), 332 (2018)

    Walton, P.: Artificial Intelligence and the Limitations of Information. Information 9(12), 332 (2018). https://doi.org/10.3390/info9120332

    2018 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.