acceptodds
Under review as a conference paper at ICLR 2027

How You Build Superintelligence: From Scaling Laws to Construction Laws

Abstract

A capability target does not specify the system that should be built to attain it. Scaling laws predict performance along a chosen development path, yet even exact knowledge of a scaling trajectory can leave the optimal resource allocation unidentified. We formulate construction laws: target-conditioned maps from identified mechanisms and physical resource constraints to a complete executable recipe, a forecast, and an auditable qualification test. We formalize superintelligence as an expert-relative capability specification that fixes the task population, within-task reliability, across-task breadth, and common information conditions. Our main inverse theorem converts uncertainty in forward predictions, resource costs, and task breadth into a guarantee on a qualifying construction; the controlling quantity is the price of target tightening, not prediction error alone. We then derive explicit construction laws for resource allocation, numerical precision and training data, reusable information, and conditional specialization, including an exact rule for jointly choosing sharing strength and specialist data. Prospectively frozen neural constructions reduce packed-weight storage by 1.33–1.59× while satisfying a predeclared 0.010-nat noninferiority criterion in all six target–seed comparisons. In a controlled hidden-support neural task, reusable structure reaches the same 80% reliability target with 32 rather than 256 task queries, while task-private structure removes the advantage. In external multi-hop retrieval, inherited transition structure consistently improves low-budget recovery, but a preregistered 2× budget-crossing forecast misses validation by 0.66 percentage points, illustrating the instability of discrete construction decisions near a target boundary. Construction laws turn capability scaling from a forward forecast into an inverse question: which system reaches the target, at what resource cost, and with what evidence?

Primary area:
learning theory
Keywords:
scaling laws, construction laws, learning theory, resource allocation, inverse design, model scaling, quantization, conditional computation, computational inheritance, superintelligence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_wMg3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces construction laws that feature a performance target, resource budget, and an identified mechanism and maps that to a recipe, forecast, and qualification test. Unlike scaling laws that predict performance along a chosen path, construction laws actually decide which system needs to be built for a desired capability. These laws are developed for resource allocation, numerical precision, reusable information, and conditional specialization. Experiments show storage savings from numerical-precision training recipes and reduced task-local queries through reusable structure. However, the prediction of 2x query reduction in the HotpotQA setting does not hold on validation.

Critical strengths
  1. The distinction between scaling laws and construction laws is interesting. In particular, the two-resource example clearly shows that even exact knowledge of performance along one scaling path may not identify the optimal resource allocation.
  2. The identify/bound/select/freeze/qualify steps from Table 1 is a useful research methodology. I also appreciate that the paper honestly reports a negative result with the 2x budget crossing forecast.
Critical weaknesses
  1. The superintelligence framing is much broader than the experiments. In the paper, it is defined as “reliable performance beyond a declared expert reference over a declared breadth of tasks,” but the experiments study numerical precision on TinyStories, reuse in a synthetic setting, and conditional shard retrieval. These experiments show specific examples of construction laws but do not validate the broader superintelligence framing.
  2. It is unclear whether the construction-law procedure selects better recipes than alternative model-selection approaches. The numerical-precision experiment does not compare against a matched-budget human-tuned or Bayesian search, does not test the neighboring five-bit recipe for E2, and does not establish minimum-storage optimality. Therefore, it is unclear how much benefit comes from the construction-law procedure rather than the fitted response model itself.
Questions for authors

Clarification Requests:

  1. For Theorem 3.1, Appendix K says that the mathematical tools are established and the contribution is in combining them into target-to-recipe statements. It is still unclear to me which part of the guarantee is technically new rather than an application of existing robust optimization results.

  2. How sensitive are the selected numerical-precision recipes to uncertainty in the fitted response model? The precision coefficient QQ reaches its imposed upper bound, and the mechanism checks have a maximum residual of 0.04010.0401 nats/byte. Do alternative reasonable fits select the same models and bit widths?

Suggestions for Future Work:

  1. Can the authors evaluate the neighboring five-bit recipe for E2 or compare against a matched-budget model-selection baseline? This would help clarify whether the construction-law procedure selects a better recipe or only a viable one.
3 comments

Thank you for the constructive review. The requests to test E2 at five bits and to assess fit sensitivity go directly to the strength of our empirical conclusions.

The precision experiment selected recipes using a fitted absolute-loss target and then qualified them using a predeclared 0.010-nat comparative noninferiority rule. All six tested target–seed comparisons satisfy that rule, with packed-weight storage reductions of 1.33 to 1.59 times relative to their references. This establishes prospective qualification for those constructions. Minimum-storage optimality requires additional comparisons.

In particular, the untested E2 five-bit neighbor has a calculated packed-weight size of 4.501 MB, compared with 5.391 MB for the selected six-bit recipe. The original response model predicts a loss of 0.820668 for this five-bit configuration, above the nominal absolute target of 0.818168. Its performance under the separate comparative criterion remains an empirical question.

We have prepared a controlled two-seed five-bit E2 experiment with the original architecture, token count, training-work cap, data, and optimizer. Only precision changes. We will report the observed result, including a failure if the candidate does not qualify.

The fitted precision coefficient also reaches its imposed upper bound, and the largest held-out mechanism residual is approximately 0.0401 nats per byte. We are examining how reasonable alternative fits affect the selected recipes. We will keep the general uncertainty-qualified theorem separate from experimental claims supported by this particular fitted response.

Your suggestions identify exactly which evidence can establish the reliability and limits of the selected recipes. We appreciate them.

Thank you again for suggesting the direct test of the E2 five-bit neighbor. We have now completed the experiment, and the result is informative for both the empirical claim and the construction procedure.

We trained architecture C (7.123 million parameters) at five bits, preserving the original 70,189,056 processed training tokens per seed, the training-work cap, and both seeds.

The five-bit model satisfies the 0.010-nat comparative noninferiority criterion against the eight-bit reference in both seeds.

On the test stream, the paired five-bit minus eight-bit loss differences are +0.004390 and +0.003054 nats/byte. The Bonferroni-adjusted simultaneous 95% upper limits are 0.005124 and 0.003738, respectively, both below the declared 0.010 threshold.

The packed-weight storage measurements are:

  • Eight bits: 7.169 MB
  • Originally selected six bits: 5.391 MB
  • Newly tested five bits: 4.501 MB

Thus, the five-bit construction achieves a 1.593× storage ratio relative to the eight-bit reference, compared with 1.330× for the originally selected six-bit construction. It uses 16.5% less storage than the six-bit recipe while satisfying the same empirical qualification criterion.

This finding also reveals a limitation in the original selection. The inverse selected six bits under a fitted absolute-loss target, while the empirical qualification rule permits a 0.010-nat degradation relative to the eight-bit comparator. The six-bit recipe qualified, but the newly tested five-bit neighbor also qualifies at lower storage cost.

We will revise the paper to report this result and align the discussion of optimality with the actual qualification criterion.

The new models were trained after your review using a fixed protocol. Their evaluation reuses the test stream from the original submission, so we report this as post-review evidence with evaluation-batch uncertainty conditional on the two seeds. An independent holdout, broader fit-sensitivity analysis, and matched-budget model-selection comparison would strengthen the conclusion further.

We appreciate the suggestion. It has produced a useful additional measurement and identified a concrete improvement to the construction methodology.

Reviewer_wMg3OpenReview ↗

Official Response From Reviewer

Thank you for the response and for running the five-bit E2 experiment. The result confirms that the original procedure did not select the lower-storage qualifying recipe. Additionally, the fit-sensitivity and matched-budget comparison still remain unresolved. I therefore maintain my rating.

Reviewer_caiUOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper formulates model design as an inverse problem, where given a desired performance level and compute constraints, it uses fitted scaling laws to choose the cheapest training configuration, with bounds on how the forecast error impacts this choice.

Critical strengths

I find the transparency of the experimental setup to be a helpful.

Critical weaknesses
  1. I think the authors do not clearly explain what is new as compared to Hoffmann et al., 2022. Specifically, the paper positions itself as extending compute-optimal scaling by taking the target as input. However, this formulation appears to be essentially the dual (i.e., a rewriting) of the original objective. I think the paper would strongly benefit from adding a rigorous discussion of what this adds in practice.

  2. Furthermore, I wonder if the authors can provide more discussion as to how does the theory serves to advance our understanding of the proposed approach, i.e., what is the purpose of Theorem 3.1, for instance?

Questions for authors

See above.

3 comments

Thank you for raising the comparison with Hoffmann et al. It gives us an opportunity to sharpen the central mathematical question.

Compute-optimal scaling determines the best achievable performance at a given budget. When that frontier is fully known, its inverse also determines the minimum budget for a specified target. We agree that this follows directly from existing optimization principles.

The remaining question is whether that optimal frontier determines the system that achieves it. Surprisingly, even complete knowledge of the globally optimal scaling curve can leave the construction unidentified.

Consider two resources with total cost and performance burden

\qquad L_\theta(x_1,x_2)=\frac{1-\theta}{x_1}+\frac{\theta}{x_2}.$$ The globally optimal scaling frontier is $$L_\theta^*(C)=\frac{\left(\sqrt{1-\theta}+\sqrt{\theta}\right)^2}{C}.$$ For $\theta=0.1$ and $\theta=0.9$, these frontiers are identical at every budget. Yet the optimal resource ratios are $3:1$ and $1:3$, respectively. Transferring the first allocation to the second system requires $7/3$ times the minimum cost to attain the same target. Thus, even the complete optimal value curve can determine the price of a capability while leaving the construction attaining it unidentified. Additional measurements of how individual resources affect performance are needed to recover the recipe. The paper studies this identification problem together with uncertainty-qualified construction. Theorem 3.1 characterizes the resource overhead required when performance, cost, and task coverage are uncertain. Its concentration and robust-optimization ingredients have established precedents. The resulting guarantee identifies the price of target tightening as the quantity governing the stability of the construction decision. We will incorporate the stronger counterexample, clarify the relationship to compute-optimal scaling, and explain more directly what Theorem 3.1 contributes. We appreciate the opportunity to sharpen this point and would welcome your thoughts on the strengthened formulation.

Thanks for the reply! I think it would be very helpful for the authors to truly clarify what they mean by "whether that optimal frontier determines the system that achieves it," and how this connects to the points I have made above.

Thank you for following up. By the "optimal frontier," we meant the scalar function giving the minimum achievable loss at each resource budget, with the allocations attaining that minimum left unspecified. By the "system that achieves it," we meant the corresponding optimizer: the architecture and resource allocation that actually attain the predicted performance.

These contain different amounts of information. In our example, the two worlds have identical minimum-loss-versus-budget curves, but their optimal resource allocations differ. Knowledge of the optimal value curve alone cannot recover those allocations.

Your observation about Hoffmann et al. is correct. Their multivariate scaling model characterizes how performance depends on model size and training data. With that response model and the compute constraint identified, one can optimize the allocation directly. Given this information, target-conditioned inversion follows from standard constrained optimization.

Our example illustrates what happens when the available measurements identify only the optimal values, or a restricted scaling trajectory, without identifying the relevant resource tradeoffs. We agree that this observation alone is insufficient to establish a new methodology.

The additional problem studied in the paper arises when the response model and resource costs are uncertain, and qualification must hold over a specified breadth of tasks. Theorem 3.1 provides upper and lower bounds on the minimum qualifying cost and relates the overhead of conservative selection to the price of tightening the target. Its concentration and robust-optimization ingredients have established precedents, which we will discuss more clearly.

We will revise the introduction to acknowledge the constructive role of multivariate scaling laws, state precisely which information is assumed available, and explain the additional guarantees developed under uncertainty.

Thank you for pressing on this point. It has helped us sharpen both the formulation and the scope of our contribution.

Reviewer_bMmTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper shifts the focus from "scaling laws" (forward predictions of performance) to "construction laws" (inverse mappings from a capability target to an optimal resource recipe). The authors provide a theoretical framework that translates forward prediction error into a guaranteed, conservative resource allocation overhead (the "price of target tightening"). They derive exact inverse rules for numerical precision, task sharing, and inheritance, and evaluate them across three experiments: a numerical precision allocation on TinyStories, a synthetic neural-world inheritance task, and a multi-hop retrieval task on HotpotQA.

Critical strengths
  1. I think treating resource allocation as a formal inverse problem is a very useful conceptual shift. Theorem 3.1 and the derivation of the "price of target tightening" (Equation 8) provide a rigorous way to understand how much extra compute or storage we must pay to guarantee a target when our scaling laws are noisy.

  2. The authors tested a preregistered 2x budget-crossing forecast for conditional retrieval, which misses its validation target by just 0.66 percentage points. To me, this negative result is a strength, as it cleanly illustrates the authors' theoretical point that discrete construction decisions are highly unstable near target boundaries.

Critical weaknesses
  1. The paper uses the concept of "superintelligence" as its framing device, but the empirical validations are conducted at a very small scale. For example, the numerical precision inverse uses the TinyStories corpus with models containing fewer than 10 million parameters. I think there is a significant gap between the smooth convexity assumed by the construction laws and the unpredictable, sometimes discontinuous scaling behaviors observed in modern 7B+ parameter language models.

  2. The explicit derived target-to-resource allocation relies on perfectly separable burdens and strictly convex optimization landscapes. I am concerned that these assumptions might not hold in the real-world training of deep neural nets, where ingredients like data quality, precision, and parameter count often exhibit complex, non-separable interaction effects.

Questions for authors

How well do you expect the exact target-to-resource allocations to hold up in larger regimes (e.g., 7B to 70B parameters), where scaling laws sometimes exhibit discontinuities or emergent behaviors that violate smooth convexity?

1 comment

Thank you for the constructive assessment and for recognizing the significance of the failed retrieval crossing. Your question about 7B to 70B regimes helps clarify which claims depend on particular response models.

Theorem 3.1 concerns a finite family of executable recipes under simultaneous error bounds for performance, resource costs, and task breadth. Its statement permits discrete architecture and precision choices and makes no convexity or differentiability assumption. The closed-form allocation and specialization results in Sections 4 and 5 have additional assumptions, including separable or quadratic response mechanisms. Applying those exact formulas to substantially larger models requires new empirical mechanism identification and reliable error bounds in that regime.

The retrieval study illustrates the challenge. Inherited transition structure improves low-budget recovery, yet a prospectively specified twofold budget-crossing forecast missed validation by 0.66 percentage points. A small error near a decision boundary can change the selected construction. The theorem characterizes when uncertainty-qualified decisions remain valid; the fitted examples show both successes and limits.

We retain "superintelligence" as a scientifically specified capability destination. The definition fixes an expert-relative reference, task breadth, within-task reliability, and common information conditions. Construction laws ask which executable system could meet such a target, at what resource cost, and with what evidence. The present experiments establish findings about individual mechanisms in controlled settings, and we will state the limits of extrapolation prominently.

We appreciate your question and will sharpen this separation in the revised paper.

Reviewer_k9HzOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Scaling laws traditionally map a resource constraint into a best possible performance. This paper investigates "construction laws" which map a desired performance into a recipe + forecast, accounting for resource constraints as well as uncertainty in the forecast. The core technique involves taking a user-specified CDF describing what risk they are willing to tolerate, and finding a recipe that can tolerate this uncertainty.

Critical strengths
  1. The paper targets the important problem of accounting for uncertainty when utilizing scaling laws.
  2. The paper has a breadth of empirical/theoretical results concerning this problem statement.
Critical weaknesses
  1. I found the paper quite difficult to read. Specifically, the organization in abstract/introduction did not reflect a coherent story, the explanation of the method relatively overcomplicated, and a lot of the technical decisions in the main body are presented without clear motivation (why is it important to have an exponenitated Renyi divergence? Why is it important for the CDF to satisfy DKW?). Forgive me if I'm off base but a lot of the writing feels AI-generated (which is bad only because I feel that it is impeding my ability to understand the results).
  2. Though there are a lot of mathematical expressions, the core result is quite simple: take an existing scaling law and invert a specified risk CDF to define the recipe.
  3. The experiments feel quite arbitrary, I'm having a hard time seeing (1) what hypothesis was set out by the authors (2) why the experiments test this hypothesis (3) whether these conclusions are upheld.
  4. The current proposed methods require users specifying CDF's, but its unclear to me how this would be done in practice. It would be helpful to have stronger recommendations surrounding this.
Questions for authors
  1. Where is the scaling law for the TinyStories experiments visualized? 11 experiments varing model size configurations feels hard to confirm a scaling law off of unless it can be audited more visually.
1 comment

Thank you for the careful reading. The opening should present the construction decision and experimental hypotheses before introducing the tools used to analyze them. We will reorganize the paper around that sequence.

On your CDF question, the user specifies a task population, a capability threshold, the required breadth of tasks, a confidence level, and resource limits. For each candidate recipe, this specification induces a distribution of task burdens over the population. That distribution can be estimated using sampled evaluation tasks. DKW supplies a concentration bound for independent task draws and a fixed finite recipe family; adaptive recipe selection requires an appropriate extension or held-out qualification. The CDF arises from measured recipe performance on the specified task population.

The exponentiated Rényi divergence has a narrower role. In our separable power-law allocation model, it gives an exact expression for the resource penalty associated with a fixed spending path. We will explain the cost penalty first and move the optional divergence interpretation away from the main argument.

The TinyStories experiment fits a local response model using 11 calibration configurations, checks four separate mechanism configurations, and evaluates six prospectively selected target–seed constructions. We agree that the fitted behavior needs a direct visual audit. We are recovering the calibration measurements to show observations, predictions, and residuals together; the scope of the fitted law will be stated carefully.

The three empirical questions will also be stated at the start of the experimental section: whether a selected precision recipe qualifies after training, when reusable structure reduces task-local acquisition, and whether a forecasted discrete retrieval-budget crossing survives external validation. The third forecast missed its target, and that failure remains part of the evidence.

Thank you for identifying changes that will help readers assess the mathematical and experimental results more directly.

Reviewer_tzf6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes “construction laws” as the inverse of scaling laws: “A forward model evaluates a recipe. A construction law returns one.” This is set up as a path to “How You Build Superintelligence”.

The core relationship studies is the construction of “recipe” to reach a target. A target, in turn, is defined over a population of tasks, with a tolerance ε for each task, and a breadth β: the fraction of tasks that must meet ε. Cost is the largest ratio of resource use to budget across resources. “Superintelligence” is defined in these terms as performance above an expert reference over a high breadth of tasks. It is not measured. The paper then goes on to present theory over a setting that bounds errors in per-task forecasts (η), in costs (ξ), and in the sampled-task distribution (u).

Critical strengths

The only salvageable value this paper may offer is that it is remotely adjacent to a real open problem in the field — which is that scaling are incredibly useful (constructive) tools given an architecture or system and its associated degrees of freedom, but it does not give insight as to the design of that system.

Critical weaknesses

It reads as loosely cobbled, barely readable AI slop — either not read by the authors, or passed with little attention to readability by them. It subsequently fails to clearly elucidate meaningful contributions to the field. Style and readability: It is rife with coined terms defined in passing or not at all: recipe, burden, breadth, dilation, qualification, inheritance, tariff, “Blind Foundry”, cold controller, canary, “price of target tightening”

Sentences are opaque, dense and needlessly hard to understand.

Content and rigor: The paper is not rigorous or coherent — with experiments and theory thinly connecting and even blatant errors. Path insufficiency claim and example given actually prove the opposite. A simple resource wise scaling clearly is sufficient in the setting provided.

Impact and novelty: The paper posits “how to build superintelligence". It does further and contrasts the “constructive laws” to the scaling laws which have been a foundational underpinning of the field. Most critically, scaling laws have offered constructive predictive power — by predicting expected performance as a function of degrees of freedom scaled — while this paper offers no predictive power.

It does not equip us with an experimental setting where the core question “what recipe” is answered in any applicative manner that can be carried or generalized towards real settings.

Questions for authors

None.

1 comment

Thank you for the detailed assessment. Your review raises three important questions: whether the resource-allocation example establishes the claimed identification problem, whether the proposed procedure has predictive value, and whether the presentation allows readers to evaluate these contributions.

On the resource-allocation example: We agree that resource-wise scaling measurements can be sufficient to identify an optimal allocation. Indeed, this is one of the central achievements of compute-optimal scaling, and our paper builds on that work.

The example considers a narrower observation setting. Its burden function is

Lθ(x₁,x₂) = (1−θ)/x₁ + θ/x₂.

Along the observed scaling path x₁ = x₂ = t, every value of θ produces exactly the same curve, Lθ(t,t) = 1/t. Yet the optimal resource ratio is x₁/x₂ = √((1−θ)/θ), which depends on θ.

The observation along this path therefore leaves the optimal allocation unidentified. Independent resource-wise measurements can identify θ and resolve the ambiguity, exactly as you suggest. This is the motivation for studying which measurements determine the construction decision. We will state the observation restriction more prominently and clarify the connection to established multivariate scaling methods.

On predictive value: We agree that a construction procedure should earn its usefulness through prospective predictions. Section 6 contains three tests with different outcomes.

In TinyStories, the selected architecture, precision, and training allocation were frozen before the corresponding models were trained. All six target–seed comparisons satisfied the declared 0.010-nat noninferiority criterion, with packed-weight storage reductions of 1.33–1.59 times. These results establish prospective qualification within the tested family; comparisons with neighboring recipes are needed to assess optimality.

In the controlled inheritance experiment, retained task structure reduced the queries needed to reach the common 80% reliability threshold from 256 to 32. A task-private intervention removed the reuse advantage, providing a direct test of the proposed mechanism.

In external retrieval, inherited transitions improved low-budget shard coverage, but the prospectively predicted twofold budget crossing failed validation by 0.66 percentage points. We report that failure because it reveals a genuine limitation of selecting constructions near a qualification boundary.

These results establish specific, testable resource conversions. Their generalization to substantially larger systems remains an open empirical question.

On presentation: Your concerns are well taken. We will simplify the opening, define each necessary quantity where it first enters the argument, replace experiment-specific names with descriptive language, and organize the results around the hypotheses and predictions actually tested.

Finally, we use superintelligence as an operational capability target defined relative to expert performance, task breadth, reliability, and common information conditions. The current experiments examine mechanisms relevant to constructing systems under such specifications. They do not establish attainment of that target.

We appreciate the opportunity to clarify these points and will use the revision to give readers a more direct account of the assumptions, predictions, and evidence.

Comments on the paper

1 comment

We thank all five reviewers for their careful assessments and constructive discussion. We have uploaded a revised manuscript and supplementary package incorporating the theoretical, experimental, and presentation improvements prompted by their comments.

Theory. The introduction now recognizes the constructive role of multivariate scaling laws and develops the additional questions of resource identification, predictive uncertainty, and target qualification. Sections 2–3 sharpen the mathematical formulation and explain how uncertainty near a capability threshold affects construction costs. Section 5 develops the sharing and specialization results from their underlying prediction problem. We have also strengthened the assumptions and proofs, including the resource-budget condition in Appendix A.3.

Calibration and numerical precision. We recovered all 11 original TinyStories calibration measurements and four held-out mechanism checks, added their plots, and evaluated four alternative response-model fits. The analysis quantifies the sensitivity of construction decisions to model estimation and identifies where changes in the fitted response alter the selected resource allocation.

Following the reviewers' suggestion, we trained and evaluated the neighboring five-bit E2 configuration. Both seeds satisfy the declared 0.010-nat noninferiority criterion, with 16.5% less packed storage than the original six-bit construction. This experiment motivates a refined selection rule that aligns resource optimization with the qualification criterion. A retrospective equal-budget grid comparison over the recorded candidates identifies the same five-bit construction.

We also strengthened the statistical assessment of the six original prospective experiments. All six satisfy the noninferiority criterion after simultaneous confidence adjustment. The complete measurements and statistical analysis are provided in the paper and supplement.

Information reuse and external validation. Section 6.2 demonstrates that reusable task structure reduces the queries needed to reach the common 80% reliability target from 256 to 32. A task-private intervention establishes the dependence of this advantage on transferable information. Section 6.3 quantifies the benefits of inherited transitions in external multi-hop retrieval and examines how a 0.66-percentage-point validation shortfall changes the predicted twofold budget crossing. Supplement J.2–J.3 provides the experimental protocols, controls, and complete numerical results.

Presentation. We have revised the writing throughout, simplifying terminology, defining quantities when introduced, motivating the mathematical results before presenting their notation, and connecting each experimental hypothesis to its measured outcome. The abstract and introduction now give a more direct account of the scientific question and contributions.

We retain superintelligence as an operational capability target defined through expert-relative performance, task breadth, reliability, and common information conditions. The mathematical framework addresses construction decisions toward such targets, with the scope of the experimental findings carefully specified.

We appreciate the reviewers' contributions to strengthening this work and welcome further questions during the remaining discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.