Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

You Are What You Eat -- AI Alignment Requires Understanding How Data Shapes Structure and Generalisation

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper argues that AI alignment cannot be assured by evaluation alone; it requires a mathematical understanding of how structure in the training data becomes internal structure in models, and how that structure drives generalisation.

desk verdict A coherent, honest position paper whose central thesis is worth engaging, but whose concrete risk mechanism rests on an explicitly unverified SGD/Bayesian bridge. read the letter →

arxiv 2502.05475 v1 pith:KGYJUEL4 submitted 2025-02-08 cs.LG

Simon Pepin Lehalleur , Jesse Hoogland , Matthew Farrugia-Roberts , Susan Wei , Alexander Gietelink Oldenziel , George Wang , Liam Carroll , Daniel Murfet This is my paper · ORCID
classification cs.LG MSC 68T0762F15
keywords AIalignmentsingularlearningtheoryinternalmodelselectiongeneralisationunderspecificationmechanisticinterpretabilityinductivebiasdatadistribution
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

This position paper argues that current AI alignment practice—shaping the training data and hoping the model internalises the intended behaviour—cannot provide safety assurances, because two networks can match on training loss yet compute their outputs through different internal algorithms and generalise differently. The authors claim that evaluation and testing, however extensive, cannot distinguish those cases, so alignment needs new statistical foundations that connect structure in the data distribution to structure inside the model and to generalisation. They point to singular learning theory as a precise example: in Bayesian inference over singular model classes, a simpler but less accurate solution can be preferred over a complex but accurate one for finite data, and they argue this mechanism can make training prefer simple, misaligned shortcut solutions even when the data uniquely specifies aligned behaviour at optimality. If this picture is right, then reliable alignment requires understanding and ultimately engineering the chain from data patterns to internal structure, not just curating datasets and measuring benchmark scores.

What carries the argument

The load-bearing mechanism is internal model selection in singular learning theory. For a parameter region $U$ with local learning coefficient $\lambda(u)$ and empirical loss $\ell_n(u)$, the contribution to the Bayesian posterior is $-\log Z_n(U) = n\ell_n(u) + \lambda(u)\log n + O_p(\log\log n)$, where the local learning coefficient is a singularity-aware measure of effective model complexity. Comparing two regions, the posterior log-odds is roughly $\Delta \ell_n \cdot n + \Delta \lambda \cdot \log n$, so a region with higher loss but simpler structure (lower local learning coefficient) is preferred while $n/\log n$ is below $\Delta \lambda / (-\Delta \ell_n)$, after which the more accurate but complex solution wins. The paper invokes this trade-off to explain how a simple but misaligned shortcut solution can dominate during a finite training window, and cites an in-context linear regression transformer experiment in which SGD trajectories first approach a generalising solution (low local learning coefficient) before converging to a memorising solution (high local learning coefficient).

What would settle it

Instrument a training run where the data distribution uniquely specifies an aligned solution at optimality but a simpler misaligned shortcut exists, and track the estimated local learning coefficients and posterior log-odds over training; if the trajectory never spends an interval near the simpler solution, or if its preference between the two regions does not follow the n/log n crossover predicted by the expansion, then the internal-model-selection mechanism is not what drives shortcut preference in deep learning.

Watch

Extended reading notes

Core claim

The paper's central claim is that alignment is an inexact science because we do not understand the chain of influence from changes in the data distribution to changes in internal model structure to changes in generalisation behaviour. It argues that internal algorithmic structure determines how a model behaves outside its training distribution, that this structure is shaped by statistical patterns in the training data, and that standard evaluation cannot certify safety because underspecification allows many low-loss parameter regions that implement qualitatively different algorithms with different generalisation properties. Using results from singular learning theory, the paper shows that a Bayesian posterior can prefer a simpler, higher-loss solution over a more complex, lower-loss aligned solution when the sample size is finite, and it hypothesises that stochastic gradient descent behaves analogously with an effective sample size that grows over training. It concludes that moving beyond evaluation to a mathematical science of AI alignment requires a statistical theory of how data shapes structure, with interpretability and pattern engineering as the two main research directions.

Load-bearing premise

The argument depends on the assumption that stochastic gradient descent behaves like Bayesian posterior inference with an effective sample size that grows over training time, so the singular-learning-theory result on internal model selection applies to deep learning; if that correspondence fails, the specific mechanism by which simpler but misaligned solutions are preferred loses its theoretical backing.

Editorial extensions

If this is right

  • If alignment techniques work only by shaping the data distribution, their effect on internal structure is indirect and can be shallow, which is consistent with findings that safety fine-tuning can be undone and that alignment-related structures may not survive distribution shift.
  • Inductive biases toward simpler solutions can make training prefer misaligned shortcuts even when the data distribution uniquely specifies aligned behaviour at optimality.
  • Compression techniques such as pruning and quantisation, which lower model complexity, may preferentially sacrifice aligned behaviour, aggravating the shortcut preference.
  • A mathematical foundation for alignment requires linking mechanistic interpretability to Bayesian statistics and developing a science of pattern engineering for post-training.
  • For reinforcement learning agents, the data distribution is co-shaped by the model's own actions, so the data-to-structure relation must extend beyond the independent and identically distributed setting.

Reading between the lines

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

  • Editorial inference: if the internal-model-selection mechanism is real, alignment behaviour should show a predictable crossover with dataset size, so experiments measuring aligned versus shortcut behaviour across increasing fine-tuning datasets could test the theory directly.
  • Editorial inference: the claim that capabilities generalise further than alignment could be restated as a difference in local learning coefficients, which would make alignment systematically more fragile than capability under distribution shift and would give a measurable target for safety assessments.
  • Editorial inference: the paper's position implies that mechanistic interpretability is not merely a descriptive tool but a necessary part of safety assurance, a stronger demand than current evaluation-based practices assume.
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

3 major / 5 minor

Summary. This position paper argues that standard evaluation-based assurance is insufficient for AI alignment because models with equal training performance can differ in internal algorithmic structure and hence in out-of-distribution generalisation. It proposes that alignment requires a mathematical understanding of the pipeline from statistical structure in training data, through internal structure in models, to generalisation. After reviewing mechanistic interpretability and singular learning theory, the paper develops a mechanism by which Bayesian internal model selection can prefer simpler but misaligned solutions at finite sample sizes, and argues that this creates an inductive-bias risk for alignment. It closes with research directions and responses to alternative views that empiricism may suffice or understanding may be intractable.

Significance. The paper is a useful, broad synthesis that connects singularity-theoretic model selection to alignment risk, and it gives a clearly stated research programme rather than a new theorem. Its strengths are that it labels the SGD-posterior bridge as a hypothesis, uses standard SLT results in Eqs. (2)-(3) in a way consistent with Watanabe's framework, and engages with concrete empirical work by the authors and others. If the central thesis is accepted, it motivates investment in singular learning theory and mechanistic interpretability as alignment foundations. The main caveat is that the paper's most concrete risk mechanism rests on an unproven correspondence between SGD trajectories and Bayesian posteriors, so the paper is best read as a research proposal whose specific alignment-risk conclusion is not yet established.

major comments (3)
  1. [§4.3 and Figure 2 caption] The specific alignment-risk mechanism is carried by the hypothesis that SGD's parameter preference evolves like a Bayesian posterior with an increasing effective sample size. The paper explicitly labels this as a hypothesis in the Figure 2 caption, and the only direct empirical support adduced is the single in-context linear regression study of Carroll et al. (2025). The other cited SLT applications establish that local-learning-coefficient-style quantities are estimable in deep networks, but not that SGD trajectories track posterior log-odds over time. If this correspondence fails, the claim in Section 5.1 that 'SLT provides a proof-of-concept' that training may find and remain at undesirable shortcuts is unsupported. Because this is the paper's most concrete argument for the SLT-based research agenda, the revision should either supply additional evidence for the SGD-posterior correspondence (for example, direct comparisons of trajectory frequencies with posterior probabilities, or tests in additional settings) or explicitly downgrade Section 5 to a speculative illustration.
  2. [§5.1] Even if the SGD-posterior correspondence holds, Eq. (3) implies that for sufficiently large training time (that is, large n) the posterior, and hence the analogue of SGD, will eventually prefer V. The clause 'and perhaps remain, at' U therefore requires an additional mechanism beyond internal model selection, such as saturation of the effective sample size, early stopping, or an SGD-specific inability to escape U once reached. The paper does not provide such a mechanism. Please either add one or soften the permanence claim to a claim about transient, finite-time preference for the simpler misaligned region.
  3. [§4.3] The empirical illustration with Carroll et al. (2025) has the opposite assignment of roles from the alignment narrative: the low-LLC region U is the generalising solution and the high-LLC region V is the memorising solution. This supports a Bayesian-style ordering of learning stages but does not by itself instantiate the simple-but-misaligned shortcut scenario of Section 5.1. The paper should state explicitly that this experiment tests only the SGD-posterior correspondence, not the alignment-risk claim, and should indicate what kind of experiment would test the latter.
minor comments (5)
  1. [§4.3] There is a grammatical agreement error: 'the phenomena described above does occur' should read 'the phenomenon described above does occur' or 'the phenomena described above do occur.'
  2. [Figure 1] The labels 'The Architecture' and 'The Optimiser' appear without explanation in the caption; consider clarifying how these boxes relate to the S4 correspondence of Wang et al. (2024).
  3. [§4.3, Eq. (2)] Watanabe's asymptotic expansion for the free energy contains a deterministic -(m-1) log log n term in addition to the Op(log log n) remainder shown in Eq. (2). The paper should state whether this term is intentionally absorbed into the remainder and whether its possible non-cancellation between U and V is assumed away.
  4. [§6.1] The sentence 'we have no current basis for confidence in our ability to align advanced AI systems' is stronger than the preceding argument supports; evaluation can provide partial assurance, and the paper's own position would be better served by saying that evaluation alone is insufficient rather than that there is no basis at all.
  5. [References] Several substantive claims are supported by unpublished blog posts or non-archival sources (for example, Murfet 2024 and Soares 2023); consider flagging these as non-archival or supplementing them with peer-reviewed material where available.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central argument is a position, SLT mathematics is external, and the SGD-to-posterior bridge is explicitly labeled a hypothesis rather than a derived prediction.

full rationale

The paper does not claim to derive a novel quantitative prediction from first principles; its central claim is the position that alignment requires understanding how data shapes internal structure and generalization. The only mathematical chain in the paper is the internal-model-selection argument in Section 4.3. Equation (3) follows algebraically from Equation (2), and Equation (2) is attributed to Watanabe (2018), an external source, with the local learning coefficient concept taken from Lau et al. (2025). This is not a case where an output is defined in terms of an input or where a fitted parameter is renamed as a prediction. The paper explicitly flags the key SGD-to-Bayesian-posterior bridge as a hypothesis in the Figure 2 caption: 'Under the hypothesis that SGD finds parameters that are preferred by the posterior, with the preference of SGD at step t evolving as the Bayesian posterior for some n increasing with t...' Section 4.2 likewise states that 'the parameters that are found by SGD are not necessarily those that are the most probable according to the posterior.' The empirical support cited for this hypothesis, Carroll et al. (2025), is prior work by overlapping authors, but it is presented as an experimental observation about transformer training trajectories, not as a theorem whose conclusion is assumed by the present paper. That makes any weakness in the bridge an evidential or correctness concern, not a circularity. The self-citations to Wang et al. (2024) and Hoogland et al. (2024) for the 'S4 correspondence' and developmental structure are also empirical anchors, and the paper's broader thesis does not reduce to them; it is independently motivated by underspecification and the insufficiency of evaluation-based assurance. Overall, the derivation chain does not collapse into its own inputs, so the circularity score is low.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper contributes no new fitted parameters or entities. Its argument rests on domain assumptions about learning dynamics, especially the SGD-posterior correspondence, which is explicitly hypothetical, and on imported mathematical results from singular learning theory.

assumptions (5)
  • domain assumption SGD trajectories behave like Bayesian posterior inference with effective sample size increasing over training, so internal model selection applies to deep learning.
    Figure 2 caption and Section 4.3. The paper states this as a hypothesis and relies on it to extend Eq. (3) from Bayesian inference to SGD-trained transformers. Only a single toy experiment (Carroll et al. 2025) is cited as support.
  • domain assumption Underspecification: training data distributions admit many low-loss parameter settings that generalize differently.
    Section 2.3, citing D'Amour et al. 2022 and Yang et al. 2021. This premise is needed to argue that test performance alone cannot distinguish safe from unsafe models.
  • domain assumption The world has algorithmic structure that neural networks learn, and this learned structure is what drives generalization.
    Section 2.2, citing Solomonoff, Crutchfield, and interpretability work. The paper uses this to argue that internal structure, not just loss, is the object alignment must control.
  • standard math Watanabe's regularity conditions for the asymptotic expansion in Eq. (2) hold for the neural network loss landscapes under discussion.
    Section 4.3 imports Watanabe 2018 Theorem 11 and Lau et al. 2025. The paper does not verify the technical conditions for deep learning, treating them as background.
  • domain assumption Post-training alignment methods such as RLHF, DPO, and constitutional AI work by shaping the effective data distribution.
    Section 3.2. This framing is necessary to claim that alignment is a data-shaping problem and that understanding data-structure links is central.

how reviews work

0 comments
Cite this review

Pith. "Pith review of You Are What You Eat -- AI Alignment Requires Understanding How Data Shapes Structure and Generalisation." pith.science (2026). https://pith.science/paper/KGYJUEL4

@misc{pith2026250205475,
  author       = {Pith},
  title        = {Pith review of: You Are What You Eat -- AI Alignment Requires Understanding How Data Shapes Structure and Generalisation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGYJUEL4}},
  note         = {Machine review of arXiv:2502.05475}
}
read the original abstract

In this position paper, we argue that understanding the relation between structure in the data distribution and structure in trained models is central to AI alignment. First, we discuss how two neural networks can have equivalent performance on the training set but compute their outputs in essentially different ways and thus generalise differently. For this reason, standard testing and evaluation are insufficient for obtaining assurances of safety for widely deployed generally intelligent systems. We argue that to progress beyond evaluation to a robust mathematical science of AI alignment, we need to develop statistical foundations for an understanding of the relation between structure in the data distribution, internal structure in models, and how these structures underlie generalisation.

Figures

Figures reproduced from arXiv: 2502.05475 by the authors.

Figure 1
Figure 1. From data to model behaviour: Structure in data determines internal structure in models and thus generalisation. Current approaches to alignment work by shaping the training distribution (left), which only indirectly determines model structure (right) through the effects on shaping the optimisation process (middle left & right). To mitigate the limitations of this indirect approach, alignment requires a better under… view at source ↗
Figure 2
Figure 2. Perfect specification is not enough: The parameter re￾gion U has higher loss but is simpler (indicated by a broader basin) while the parameter region V has lower loss but is more complex. The posterior distribution could, in theory, prefer the higher loss U when the sample size n is low. Under the hypothesis that SGD finds parameters that are preferred by the posterior, with the pref￾erence of SGD at step t evolving… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Influence Dynamics and Stagewise Data Attribution

    cs.LG 2025-10 conditional novelty 7.0 of 10

    Using Bayesian influence functions and singular learning theory, the authors show that a sample's influence on a model varies non-monotonically over training, peaking and flipping sign at phase transitions.

  2. A Scalable Approach to Evaluating Moral Sensitivity in LLMs

    cs.CY 2026-07 conditional novelty 6.5 of 10

    Under morally irrelevant noise, eight LLMs preserve the semantic content of identified moral features above calibrated floors, despite significant changes in feature counts.

  3. Value Drifts: Tracing Value Alignment During LLM Post-Training

    cs.CL 2025-10 conditional novelty 6.0 of 10

    Value alignment in LLMs is set largely during supervised fine-tuning; standard preference-optimization datasets carry too little stance contrast to re-align it, but with engineered contrast algorithms differ (DPO ampl...

Reference graph

Works this paper leans on

121 extracted references · 41 canonical work pages · cited by 3 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    B., and Misiakiewicz, T

    Abbe, E., Adsera, E. B., and Misiakiewicz, T. SGD learning on neural networks: leap complexity and saddle-to-saddle dynamics. In The Thirty Sixth Annual Conference on Learning Theory, pp.\ 2552--2623. PMLR, 2023

    2023

  3. [3]

    Critical learning periods in deep networks

    Achille, A., Rovere, M., and Soatto, S. Critical learning periods in deep networks. In International Conference on Learning Representations, September 2018. URL https://openreview.net/forum?id=BkeStsCcKQ

    2018

  4. [4]

    Adams, Z. P. and Mukherjee, S. Meta-posterior consistency for the bayesian inference of metastable system, August 2024. Preprint arXiv:2408.01868 https://arxiv.org/abs/2408.01868 [stat.ML]

    work page Pith review arXiv 2024

  5. [5]

    and Hernandez, D

    Amodei, D. and Hernandez, D. AI and compute. OpenAI Blog, May 2018. URL https://openai.com/index/ai-and-compute/

    2018

  6. [6]

    Concrete problems in AI safety, June 2016

    Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., and Mané, D. Concrete problems in AI safety, June 2016. Preprint arXiv:1606.06565 https://arxiv.org/abs/1606.06565 [cs.AI]

    arXiv 2016

  7. [7]

    S., Jenner, E., Casper, S., Sourbut, O., Edelman, B

    Anwar, U., Saparov, A., Rando, J., Paleka, D., Turpin, M., Hase, P., Lubana, E. S., Jenner, E., Casper, S., Sourbut, O., Edelman, B. L., Zhang, Z., G \"u nther, M., Korinek, A., Hernandez-Orallo, J., Hammond, L., Bigelow, E. J., Pan, A., Langosco, L., Korbak, T., Zhang, H. C., Zhong, R., h\'Eigeartaigh, S. O., Recchia, G., Corsi, G., Chan, A., Anderljung,...

    2024

  8. [8]

    A general language assistant as a laboratory for alignment, 2021

    Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al. A general language assistant as a laboratory for alignment, 2021

    2021

Show all 121 references
  1. [9]

    Novum Organum

    Bacon, F. Novum Organum. 1620. In Latin. Translated into English by fowler1878bacon

  2. [10]

    Training a helpful and harmless assistant with reinforcement learning from human feedback, April 2022 a

    Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., Joseph, N., Kadavath, S., Kernion, J., Conerly, T., El-Showk, S., Elhage, N., Hatfield-Dodds, Z., Hernandez, D., Hume, T., Johnston, S., Kravec, S., Lovitt, L.,...

    2022 arXiv

  3. [11]

    E., Fort, S., Lanham, T., Telleen-Lawton, T., Conerly, T., Henighan, T., Hume, T., Bowman, S

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., Chen, C., Olsson, C., Olah, C., Hernandez, D., Drain, D., Ganguli, D., Li, D., Tran-Johnson, E., Perez, E., Kerr, J., Mueller, J., Ladish, J., Landau, J., ...

    2022 arXiv

  4. [12]

    Open problems in machine unlearning for AI safety, January 2025

    Barez, F., Fu, T., Prabhu, A., Casper, S., Sanyal, A., Bibi, A., O'Gara, A., Kirk, R., Bucknall, B., Fist, T., Ong, L., Torr, P., Lam, K.-Y., Trager, R., Krueger, D., Mindermann, S., Hernandez-Orallo, J., Geva, M., and Gal, Y. Open problems in machine unlearning for AI safety,...

    2025 arXiv

  5. [13]

    Two models of double descent for weak features

    Belkin, M., Hsu, D., and Xu, J. Two models of double descent for weak features. SIAM Journal on Mathematics of Data Science, 2 0 (4): 0 1167--1180, 2020

    2020

  6. [14]

    Representation learning: A review and new perspectives

    Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 0 (8): 0 1798--1828, June 2012. doi:10.1109/TPAMI.2013.50

    2012 doi

  7. [15]

    N., Zhang, Y.-Q., Xue, L., Shalev-Shwartz, S., Hadfield, G., Clune, J., Maharaj, T., Hutter, F., Baydin, A

    Bengio, Y., Hinton, G., Yao, A., Song, D., Abbeel, P., Darrell, T., Harari, Y. N., Zhang, Y.-Q., Xue, L., Shalev-Shwartz, S., Hadfield, G., Clune, J., Maharaj, T., Hutter, F., Baydin, A. G., McIlraith, S., Gao, Q., Acharya, A., Krueger, D., Dragan, A., Torr, P., Russell, S., K...

    2023 arXiv

  8. [16]

    a , J., Johnson, C., Jolly, G., Katzir, Z., Khan, S. M., Kitano, H., Kr \

    Bengio, Y., Mindermann, S., Privitera, D., Besiroglu, T., Bommasani, R., Casper, S., Choi, Y., Fox, P., Garfinkel, B., Goldfarb, D., Heidari, H., Ho, A., Kapoor, S., Khalatbari, L., Longpre, S., Manning, S., Mavroudis, V., Mazeika, M., Michael, J., Newman, J., Ng, K. Y., Okolo...

    2025

  9. [17]

    A., Adeli, E., Altman, R., Arora, S., Arx, S

    Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., Arx, S. v., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J. Q., Demszky, D., Donahue, C., Doumbouya, M., ...

    2021 arXiv

  10. [18]

    Superintelligence: Paths, Dangers, Strategies

    Bostrom, N. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014

    2014

  11. [19]

    The scaling hypothesis, 2020

    Branwen, G. The scaling hypothesis, 2020. URL https://www.gwern.net/Scaling-Hypothesis. Accessed: 2023-08-05

    2020

  12. [20]

    Statistical modeling: The two cultures (with comments and a rejoinder by the author)

    Breiman, L. Statistical modeling: The two cultures (with comments and a rejoinder by the author). Statistical Science, 16 0 (3): 0 199--231, 2001

    2001

  13. [21]

    D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33: 0 1877--1901, 2020

    1901

  14. [22]

    H., Baker, B., Gao, L., Aschenbrenner, L., Chen, Y., Ecoffet, A., Joglekar, M., Leike, J., et al

    Burns, C., Izmailov, P., Kirchner, J. H., Baker, B., Gao, L., Aschenbrenner, L., Chen, Y., Ecoffet, A., Joglekar, M., Leike, J., et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. In International Conference on Machine Learning, 2023

    2023

  15. [23]

    Bury, R. G. Timaeus. Critias. Cleitophon. Menexenus. Epistles, chapter 1, pp.\ 1--254. Number 234 in Loeb Classical Library. Harvard University Press, 1929. English translation of timaeus

    1929

  16. [24]

    Thread: Circuits

    Cammarata, N., Olah, C., Schubert, L., Goh, G., Petrov, M., and Carter, S. Thread: Circuits. Distill, March 2020. doi:10.23915/distill.00024. https://distill.pub/2020/circuits

    2020 doi

  17. [25]

    Spectral bias and task-model alignment explain generalization in kernel regression and infinitely wide neural networks

    Canatar, A., Bordelon, B., and Pehlevan, C. Spectral bias and task-model alignment explain generalization in kernel regression and infinitely wide neural networks. Nature Communications, 12 0 (1): 0 2914, June 2020. doi:10.1038/s41467-021-23103-1

    2020 doi

  18. [26]

    Cardwell, D. S. L. From Watt to Clausius: The Rise of Thermodynamics in the Early Industrial Age . Cornell University Press, 1971

    1971

  19. [27]

    Dynamics of transient structure in in-context linear regression transformers, January 2025

    Carroll, L., Hoogland, J., Farrugia-Roberts, M., and Murfet, D. Dynamics of transient structure in in-context linear regression transformers, January 2025. Preprint arXiv:2501.17745 https://arxiv.org/abs/2501.17745 [cs.LG]

    2025 arXiv

  20. [28]

    K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., Wang, T

    Casper, S., Davies, X., Shi, C., Gilbert, T. K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., Wang, T. T., Marks, S., Segerie, C.-R., Carroll, M., Peng, A., Christoffersen, P. J., Damani, M., Slocum, S., Anwar, U., Siththaranjan, A., Nadeau, M., ...

    2023

  21. [29]

    Natural abstractions: Key claims, theorems, and critiques

    Chan, L., Lang, L., and Jenner, E. Natural abstractions: Key claims, theorems, and critiques. LessWrong, 2023. URL https://www.lesswrong.com/posts/gvzW46Z3BsaZsLc25

    2023

  22. [30]

    Stochastic collapse: How gradient noise attracts SGD dynamics towards simpler subnetworks

    Chen, F., Kunin, D., Yamamura, A., and Ganguli, S. Stochastic collapse: How gradient noise attracts SGD dynamics towards simpler subnetworks. Neural Information Processing Systems, 2023 a

    2023

  23. [31]

    Dynamical versus bayesian phase transitions in a toy model of superposition, October 2023 b

    Chen, Z., Lau, E., Mendel, J., Wei, S., and Murfet, D. Dynamical versus bayesian phase transitions in a toy model of superposition, October 2023 b . Preprint arXiv:2310.06301 https://arxiv.org/abs/2310.06301 [cs.LG]

    2023 arXiv

  24. [32]

    Model compression and acceleration for deep neural networks: The principles, progress, and challenges

    Cheng, Y., Wang, D., Zhou, P., and Zhang, T. Model compression and acceleration for deep neural networks: The principles, progress, and challenges. IEEE Signal Processing Magazine, 35 0 (1): 0 126--136, 2018. doi:10.1109/MSP.2017.2765695

    2018

  25. [33]

    A comprehensive survey on model compression and acceleration

    Choudhary, T., Mishra, V., Goswami, A., and Sarangapani, J. A comprehensive survey on model compression and acceleration. Artificial Intelligence Review, 53 0 (7): 0 5113--5155, 2020. doi:10.1007/s10462-020-09816-7

    2020 doi

  26. [34]

    What does the universal prior actually look like? Ordinary Ideas (Blog), 2016

    Christiano, P. What does the universal prior actually look like? Ordinary Ideas (Blog), 2016. URL https://ordinaryideas.wordpress.com/2016/11/30/what-does-the-universal-prior-actually-look-like/

    2016

  27. [35]

    Clarifying AI alignment

    Christiano, P. Clarifying AI alignment. LessWrong, 2018. URL https://www.lesswrong.org/posts/ZeE7EKHTFMBs8eMxn

    2018

  28. [36]

    What failure looks like

    Christiano, P. What failure looks like. LessWrong, 2019. URL https://www.lesswrong.com/posts/HBxe6wdjxK239zajf

    2019

  29. [37]

    A toy model of universality: Reverse engineering how networks learn group operations

    Chughtai, B., Chan, L., and Nanda, N. A toy model of universality: Reverse engineering how networks learn group operations. In International Conference on Machine Learning, 2023

    2023

  30. [38]

    K., Hutter, M., Bengio, Y., and Russell, S

    Cohen, M. K., Hutter, M., Bengio, Y., and Russell, S. RL, but don't do anything I wouldn't do, 2024. Preprint arXiv:2410.06213 https://arxiv.org/abs/2410.06213 [cs.LG]

    2024 arXiv

  31. [39]

    Copeland, B. J. (ed.). The Essential Turing: Seminal Writings in Computing, Logic, Philosophy, Artificial Intelligence, and Artificial Life: Plus The Secrets of Enigma, chapter 10, pp.\ 404--432. Oxford University Press, 2004. Restoration and republication of turing1948 , with...

    2004

  32. [40]

    Cox, D. R. Principles of Statistical Inference. Cambridge University Press, 2006

    2006

  33. [41]

    Crutchfield, J. P. Between order and chaos. Nature Physics, 8 0 (1): 0 17--24, 2012

    2012

  34. [42]

    D., et al

    D'Amour, A., Heller, K., Moldovan, D., Adlam, B., Alipanahi, B., Beutel, A., Chen, C., Deaton, J., Eisenstein, J., Hoffman, M. D., et al. Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23 0 (226): 0 1--61, 2022

    2022

  35. [43]

    DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., D...

    2025 arXiv

  36. [44]

    K., Aitchison, M., Orseau, L., Hutter, M., and Veness, J

    Delétang, G., Ruoss, A., Duquenne, P.-A., Catt, E., Genewein, T., Mattern, C., Grau-Moya, J., Wenliang, L. K., Aitchison, M., Orseau, L., Hutter, M., and Veness, J. Language modeling is compression. In International Conference on Learning Representations, 2023

    2023

  37. [45]

    Eddington, A. S. The Internal Constitution of the Stars. Cambridge University Press, 1926. Reissued in the Cambridge Science Classics series, 1988

    1926

  38. [46]

    AGI safety literature review

    Everitt, T., Lea, G., and Hutter, M. AGI safety literature review. In International Joint Conference on Artificial Intelligence, pp.\ 5441--5449. International Joint Conferences on Artificial Intelligence Organization, July 2018. doi:10.24963/ijcai.2018/768

    2018 doi

  39. [47]

    and Ganguli, S

    Fort, S. and Ganguli, S. Emergent properties of the local geometry of neural loss landscapes, October 2019. Preprint arXiv:1910.05929 https://arxiv.org/abs/1910.05929 [cs.LG]

    2019 arXiv

  40. [48]

    Bacon's Novum Organum

    Fowler, T. Bacon's Novum Organum. Clarendon Press, 1878. Annotated English translation of bacon

  41. [49]

    BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B , October 2023

    Gade, P., Lermen, S., Rogers-Smith, C., and Ladish, J. BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B , October 2023. Preprint arXiv:2311.00117 https://arxiv.org/abs/2311.00117 [cs.CL]

    2023 arXiv

  42. [50]

    Good, I. J. Speculations concerning the first ultraintelligent machine. Advances in Computers, pp.\ 31--88, 1966. doi:10.1016/S0065-2458(08)60418-0

    1966 doi

  43. [51]

    AI control: Improving safety despite intentional subversion

    Greenblatt, R., Shlegeris, B., Sachan, K., and Roger, F. AI control: Improving safety despite intentional subversion. In Forty-First International Conference on Machine Learning, June 2024

    2024

  44. [52]

    and Roos, T

    Grünwald, P. and Roos, T. Minimum description length revisited. International Journal of Mathematics for Industry, 11 0 (01): 0 1930001, March 2020. doi:10.1142/S2661335219300018

    2020 doi

  45. [53]

    Y., Joglekar, M., Wallace, E., Jain, S., Barak, B., Helyar, A., Dias, R., Vallone, A., Ren, H., Wei, J., Chung, H

    Guan, M. Y., Joglekar, M., Wallace, E., Jain, S., Barak, B., Helyar, A., Dias, R., Vallone, A., Ren, H., Wei, J., Chung, H. W., Toyer, S., Heidecke, J., Beutel, A., and Glaese, A. Deliberative alignment: Reasoning enables safer language models, December 2024. Preprint arXiv:24...

    2024 arXiv

  46. [54]

    and Schmidhuber, J

    Ha, D. and Schmidhuber, J. World models, March 2018. Preprint arxiv:1803.10122 https://arxiv.org/abs/1803.10122 [cs]

    2018 arXiv

  47. [55]

    Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks

    He, T., Doshi, D., Das, A., and Gromov, A. Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, June 2024

    2024

  48. [56]

    Introduction to AI Safety, Ethics, and Society

    Hendrycks, D. Introduction to AI Safety, Ethics, and Society . Taylor & Francis, 2025

    2025

  49. [57]

    Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., Patwary, M. M. A., Yang, Y., and Zhou, Y. Deep learning scaling is predictable, empirically, December 2017

    2017

  50. [58]

    and Schmidhuber, J

    Hochreiter, S. and Schmidhuber, J. Flat minima. Neural Computation, 9 0 (1): 0 1--42, 1997

    1997

  51. [59]

    The developmental landscape of in-context learning, February 2024

    Hoogland, J., Wang, G., Farrugia-Roberts, M., Carroll, L., Wei, S., and Murfet, D. The developmental landscape of in-context learning, February 2024. Preprint arXiv:2402.02364 https://arxiv.org/abs/2402.02364 [cs.LG]

    2024 arXiv

  52. [60]

    Are minimal circuits deceptive? LessWrong, 2019

    Hubinger, E. Are minimal circuits deceptive? LessWrong, 2019. URL https://www.lesswrong.com/posts/fM5ZWGDbnjb7ThNKJ

    2019

  53. [61]

    v., Mikulik, V., Skalse, J., and Garrabrant, S

    Hubinger, E., Merwijk, C. v., Mikulik, V., Skalse, J., and Garrabrant, S. Risks from learned optimization in advanced machine learning systems, June 2019. Preprint arXiv:1906.01820 https://arxiv.org/abs/1906.01820 [cs.AI]

    2019 arXiv

  54. [62]

    Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., Jermyn, A., Askell, A., Radhakrishnan, A., Anil, C., Duvenaud, D., Ganguli, D., Barez, F., Clark, J., Ndousse, K., Sachan, K., Sellitto, M., Sharma, M....

    2024 arXiv

  55. [63]

    The platonic representation hypothesis

    Huh, M., Cheung, B., Wang, T., and Isola, P. The platonic representation hypothesis. In International Conference on Machine Learning, 2024

    2024

  56. [64]

    Adversarial examples are not bugs, they are features

    Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., and Madry, A. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019

    2019

  57. [65]

    Johnson, M. H. and de Haan, M. D. Developmental Cognitive Neuroscience: An Introduction. John Wiley & Sons, June 2015

    2015

  58. [66]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models, January 2020. Preprint arXiv:2001.08361 https://arxiv.org/abs/2001.08361 [cs.LG]

    2020 arXiv

  59. [67]

    Kiebel, S. J. and Friston, K. J. Free energy and dendritic self-organization. Frontiers in Systems Neuroscience, 5, July 2011. doi:10.3389/fnsys.2011.00080

    2011 arXiv

  60. [68]

    A sketch of an AI control safety case, January 2025

    Korbak, T., Clymer, J., Hilton, B., Shlegeris, B., and Irving, G. A sketch of an AI control safety case, January 2025. Preprint arXiv:2501.17315 https://arxiv.org/abs/2501.17315 [cs.AI]

    2025 arXiv

  61. [69]

    The learning theoretic alignment agenda

    Kosoy, V. The learning theoretic alignment agenda. LessWrong, 2018. URL https://www.lesswrong.com/posts/5bd75cc58225bf0670375575

    2018

  62. [70]

    D., Pfau, J., and Krueger, D

    Langosco, L., Koch, J., Sharkey, L. D., Pfau, J., and Krueger, D. Goal misgeneralization in deep reinforcement learning. In International Conference on Machine Learning, pp.\ 12004--12019. PMLR, 2022

    2022

  63. [71]

    The local learning coefficient: A singularity-aware complexity measure

    Lau, E., Furman, Z., Wang, G., Murfet, D., and Wei, S. The local learning coefficient: A singularity-aware complexity measure. In The 28th International Conference on Artificial Intelligence and Statistics, 2025. URL https://openreview.net/forum?id=1av51ZlsuL. To appear. Prepr...

    2025 arXiv

  64. [72]

    Reinforcement learning and control as probabilistic inference: Tutorial and review, May 2018

    Levine, S. Reinforcement learning and control as probabilistic inference: Tutorial and review, May 2018. Preprint arXiv:1805.00909 https://arxiv.org/abs/1805.00909 [cs.LG]

    2018 arXiv

  65. [73]

    Visualizing the loss landscape of neural nets

    Li, H., Xu, Z., Taylor, G., Studer, C., and Goldstein, T. Visualizing the loss landscape of neural nets. Advances in Neural Information Processing Systems, 31, 2018

    2018

  66. [74]

    Does circuit analysis interpretability scale? evidence from multiple choice capabilities in chinchilla, July 2023

    Lieberum, T., Rahtz, M., Kramár, J., Nanda, N., Irving, G., Shah, R., and Mikulik, V. Does circuit analysis interpretability scale? evidence from multiple choice capabilities in chinchilla, July 2023. Preprint arXiv:2307.09458 https://arxiv.org/abs/2307.09458 [cs.LG]

    2023 arXiv

  67. [75]

    The Lever of Riches: Technological Creativity and Economic Progress

    Mokyr, J. The Lever of Riches: Technological Creativity and Economic Progress. Oxford University Press, 1992

    1992

  68. [76]

    The queen’s dilemma: A paradox of control

    Murfet, D. The queen’s dilemma: A paradox of control. LessWrong, 2024. URL https://www.lesswrong.com/posts/ui3AztyGKCpHA5ZhM

    2024

  69. [77]

    Deep double descent: Where bigger models and more data hurt

    Nakkiran, P., Kaplun, G., Bansal, Y., Yang, T., Barak, B., and Sutskever, I. Deep double descent: Where bigger models and more data hurt. Journal of Statistical Mechanics: Theory and Experiment, 2021 0 (12): 0 124003, 2021

    2021

  70. [78]

    Progress measures for grokking via mechanistic interpretability

    Nanda, N., Chan, L., Lieberum, T., Smith, J., and Steinhardt, J. Progress measures for grokking via mechanistic interpretability. In International Conference on Learning Representations, 2023

    2023

  71. [79]

    AGI safety from first principles

    Ngo, R. AGI safety from first principles. Alignment Forum, 2020. URL https://www.alignmentforum.org/s/mzgtmmTKKn5MuCzFJ

    2020

  72. [80]

    Visualizing representations: Deep learning and human beings, January 2015

    Olah, C. Visualizing representations: Deep learning and human beings, January 2015. URL https://colah.github.io/posts/2015-01-Visualizing-Representations/

    2015

  73. [81]

    Zoom in: An introduction to circuits

    Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., and Carter, S. Zoom in: An introduction to circuits. Distill, 5, March 2020. doi:10.23915/distill.00024.001. https://distill.pub/2020/circuits/zoom-in

    2020 doi

  74. [82]

    In-context learning and induction heads

    Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Johnston, S., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Cla...

    2022

  75. [83]

    Omohundro, S. M. The basic AI drives. In Artificial Intelligence Safety and Security, pp.\ 47--55. Chapman and Hall/CRC, July 2018. doi:10.1201/9781351251389-3

    2018 doi

  76. [84]

    OpenAI, :, Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Carney, A., Iftimie, A., Karpenko, A., Passos, A. T., Neitz, A., Prokofiev, A., Wei, A., Tam, A., Bennett, A., Kumar, A., Saraiva, A., Vallone, A., Duberstein...

    2024 arXiv

  77. [85]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and Lowe, R. Training language models to f...

    2022

  78. [86]

    Measurements of three-level hierarchical structure in the outliers in the spectrum of deepnet hessians

    Papyan, V. Measurements of three-level hierarchical structure in the outliers in the spectrum of deepnet hessians. In International Conference on Machine Learning, pp.\ 5012--5021. PMLR, January 2019

    2019

  79. [87]

    Plato. Timaeus. c. 360 BCE. In Ancient Greek. Translated into English by bury1929timaeus

  80. [88]

    Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, October 2023

    Qi, X., Zeng, Y., Xie, T., Chen, P.-Y., Jia, R., Mittal, P., and Henderson, P. Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, October 2023

    2023

  81. [89]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. Neural Information Processing Systems, 2023

    2023

  82. [90]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21 0 (140): 0 1--67, October 2019

    2019

  83. [91]

    Position: Understanding LLMs requires more than statistical generalization

    Reizinger, P., Ujváry, S., Mészáros, A., Kerekes, A., Brendel, W., and Huszár, F. Position: Understanding LLMs requires more than statistical generalization. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning ...

    2024

  84. [92]

    Stochastic Complexity in Statistical Inquiry, volume 15

    Rissanen, J. Stochastic Complexity in Statistical Inquiry, volume 15. World Scientific, November 1989. doi:10.1142/0822

    1989 doi

  85. [93]

    Rogers, T. T. and McClelland, J. L. Semantic Cognition. MIT Press, June 2004. doi:10.7551/mitpress/6161.001.0001

    2004 doi

  86. [94]

    Human Compatible: Artificial Intelligence and the Problem of Control

    Russell, S. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin, October 2019

    2019

  87. [95]

    M., McClelland, J

    Saxe, A. M., McClelland, J. L., and Ganguli, S. A mathematical theory of semantic development in deep neural networks. Proceedings of the National Academy of Sciences, 116 0 (23): 0 11537 -- 11546, October 2018. doi:10.1073/pnas.1820226116

    2018 doi

  88. [96]

    Learning complex, extended sequences using the principle of history compression

    Schmidhuber, J. Learning complex, extended sequences using the principle of history compression. Neural Computation, 4 0 (2): 0 234--242, 1992

    1992

  89. [97]

    Sejnowski, T. J. The unreasonable effectiveness of deep learning in artificial intelligence. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30033--30038, December 2020. ISSN 0027-8424, 1091-6490. doi:10.1073/pnas.1907373117

    2020 doi

  90. [98]

    J., Casper, S., Tegmark, M., Saunders, W., Bau, D., Todd, E., Geiger, A., Geva, M., Hoogland, J., Murfet, D., and McGrath, T

    Sharkey, L., Chughtai, B., Batson, J., Lindsey, J., Wu, J., Bushnaq, L., Goldowsky-Dill, N., Heimersheim, S., Ortega, A., Bloom, J., Biderman, S., Garriga-Alonso, A., Conmy, A., Nanda, N., Rumbelow, J., Wattenberg, M., Schoots, N., Miller, J., Michaud, E. J., Casper, S., Tegma...

    2025 arXiv

  91. [99]

    and Tishby, N

    Shwartz-Ziv, R. and Tishby, N. Opening the black box of deep neural networks via information, March 2017. Preprint arXiv:1703.00810 https://arxiv.org/abs/1703.00810 [cs.LG]

    2017 arXiv

  92. [100]

    Deep deceptiveness

    Soares, N. Deep deceptiveness. LessWrong, 2023. URL https://www.lesswrong.org/posts/XWwvwytieLtEWaFJX

    2023

  93. [101]

    Solomonoff, R. J. A formal theory of inductive inference. Part I . Information and Control, 7 0 (1): 0 1--22, 1964. doi:10.1016/s0019-9958(64)90223-2

    1964 doi

  94. [102]

    and Mukherjee, S

    Su, L. and Mukherjee, S. Large deviation asymptotics and bayesian posterior consistency on stochastic processes and dynamical systems, June 2021. Preprint arXiv:2106.06894 https://arxiv.org/abs/2106.06894 [math.ST]

    2021 arXiv

  95. [103]

    The bitter lesson, March 2019

    Sutton, R. The bitter lesson, March 2019. URL http://incompleteideas.net/IncIdeas/BitterLesson.html. Incomplete Ideas (Blog)

    2019

  96. [104]

    L., McDougall, C., MacDiarmid, M., Freeman, C

    Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., Pearce, A., Citro, C., Ameisen, E., Jones, A., Cunningham, H., Turner, N. L., McDougall, C., MacDiarmid, M., Freeman, C. D., Sumers, T. R., Rees, E., Batson, J., Jermyn, A., Carter, S., Olah, C., and H...

    2024

  97. [105]

    LLM circuit analyses are consistent across training and scale

    Tigges, C., Hanna, M., Yu, Q., and Biderman, S. LLM circuit analyses are consistent across training and scale. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, November 2024

    2024

  98. [106]

    Intelligent machinery

    Turing, A. Intelligent machinery. Technical report, National Physical Laboratory, 1948. Restored and republished by copeland2004turing1948

    1948

  99. [107]

    and Nasufi, J

    Urdshals, E. and Nasufi, J. Structure Development in List Sorting Transformers . In NeurIPS 2024 Workshop on Scientific Methods for Understanding Deep Learning , November 2024

    2024

  100. [108]

    Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient

    Wang, G., Hoogland, J., Van Wingerden, S., Furman, Z., and Murfet, D. Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient . In The Thirteenth International Conference on Learning Representations , October 2024

    2024

  101. [109]

    Interpretability in the wild: a circuit for indirect object identification in GPT-2 small, November 2022

    Wang, K., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small, November 2022. Preprint arXiv:2211.00593 https://arxiv.org/abs/2211.00593 [cs.LG]

    2022 arXiv

  102. [110]

    Algebraic Geometry and Statistical Learning Theory

    Watanabe, S. Algebraic Geometry and Statistical Learning Theory. Cambridge University Press, 2009

    2009

  103. [111]

    Mathematical Theory of Bayesian Statistics

    Watanabe, S. Mathematical Theory of Bayesian Statistics. CRC Press, Taylor & Francis Group, USA, April 2018. doi:10.1201/9781315373010

    2018 doi

  104. [112]

    Recent advances in algebraic geometry and Bayesian statistics

    Watanabe, S. Recent advances in algebraic geometry and Bayesian statistics. Information Geometry, 7 0 (Suppl 1): 0 187--209, November 2022. doi:10.48550/arXiv.2211.10049

    2022 doi arXiv

  105. [113]

    Review and prospect of algebraic research in equivalent framework between statistical mechanics and machine learning theory, May 2024

    Watanabe, S. Review and prospect of algebraic research in equivalent framework between statistical mechanics and machine learning theory, May 2024. Preprint arXiv:2406.10234 https://arxiv.org/abs/2406.10234 [cond-mat.stat-mech]

    2024 arXiv

  106. [114]

    Y., Guu, K., Yu, A

    Wei, J., Bosma, M., Zhao, V. Y., Guu, K., Yu, A. W., Lester, B., Du, N., Dai, A. M., and Le, Q. V. Finetuned language models are zero-shot learners. In International Conference on Learning Representations, 2021

    2021

  107. [115]

    Deep learning is singular, and that’s good

    Wei, S., Murfet, D., Gong, M., Li, H., Gell-Redman, J., and Quella, T. Deep learning is singular, and that’s good. IEEE Transactions on Neural Networks and Learning Systems, 34 0 (12): 0 10473--10486, December 2023. doi:10.1109/TNNLS.2022.3167409

    2023

  108. [116]

    Some moral and technical consequences of automation

    Wiener, N. Some moral and technical consequences of automation. Science, 131 0 (3410): 0 1355--1358, 1960

    1960

  109. [117]

    Yampolskiy, R. V. Unexplainability and incomprehensibility of AI . Journal of Artificial Intelligence and Consciousness, 7 0 (02): 0 277--291, 2020

    2020

  110. [118]

    Yang, R., Mao, J., and Chaudhari, P. Does the data induce capacity control in deep learning? In Proceedings of the 39th International Conference on Machine Learning, volume abs/2110.14163 of Proceedings of Machine Learning Research, pp.\ 25166--25197. PMLR, October 2021

    arXiv

  111. [119]

    Coherent extrapolated volition

    Yudkowsky, E. Coherent extrapolated volition. Technical report, Singularity Institute for Artificial Intelligence, 2004. URL https://intelligence.org/files/CEV.pdf

    2004

  112. [120]

    The hidden complexity of wishes

    Yudkowsky, E. The hidden complexity of wishes. LessWrong, 2007. URL https://www.lesswrong.com/posts/4ARaTpNX62uaL86j6

    2007

  113. [121]

    and McDaniel, P

    Zohuri, B. and McDaniel, P. Thermodynamics in Nuclear Power Plant Systems. Springer, 2015

    2015

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.