On the Slow Death of Scaling

Authors: Sara Hooker Year: Not stated. (Latest citations are from 2025; paper is consistent with a 2025 composition date.) Tags: scaling-laws, compute-efficiency, large-language-models, ai-governance, algorithmic-progress, position-paper

TL;DR

The essay argues that raw compute scaling is a weakening driver of AI progress: smaller models now routinely outperform much larger ones, scaling laws fail to predict downstream task performance, and the most consequential near-term gains will come from inference-time compute, data quality, and algorithmic innovation rather than additional pre-training FLOPs. The argument carries policy implications because compute thresholds underpin major AI governance frameworks.

First pass — the five C's

Category. Position essay / opinion piece. No original experiments; synthesizes published results and leaderboard observations.

Context. AI scaling subfield. Builds on: Sutton's "Bitter Lesson" (compute beats domain knowledge as the dominant historical driver); Kaplan et al. (2020) scaling laws for LMs; Hooker's own prior work on the hardware lottery (Hooker 2020, 2021b); Schaeffer et al. (2023/2024) on emergent abilities as artifact of evaluation metrics.

Correctness. Load-bearing assumptions: (1) Open LLM Leaderboard rankings are a valid proxy for general model capability; (2) the observed small-beats-large trend will persist rather than revert; (3) inference-time compute gains (estimated at 5×–20×, citing Davidson et al. 2023) generalize broadly. All three are asserted rather than rigorously established.

Contributions. - Synthesizes leaderboard evidence that models with 4.5%–5% of the parameters of larger contemporaries can outperform them (e.g., Aya Expanse 8B vs. BLOOM 176B, Llama-3 8B vs. Falcon 180B). - Identifies four internal contradictions in the scaling thesis: weight redundancy, long-tail learning inefficiency, data-quality substitutability, and architecture-dependence of any scaling law. - Argues that inference-time, gradient-free optimization constitutes a qualitatively new paradigm displacing training-compute scaling. - Critiques compute-threshold governance (EU AI Act, US executive orders) as built on an assumption of monotonically growing model size that the empirical record does not support.

Clarity. Readable and well-structured for a broad technical audience; leans on analogy (magnetron/microwave, silver bullet, barber anecdote) in places where quantitative precision would be more convincing.

Second pass — content

Main thrust: The compute–performance relationship is structurally weakening because of architectural redundancy, data-quality substitution effects, and a growing suite of algorithmic improvements that add little or no training compute; governance and investment frameworks that treat scaling as inevitable are therefore mispricing both risk and opportunity.

Supporting evidence: - Aya 23 8B and Aya Expanse 8B each outperform BLOOM 176B (≈22× the parameters) on the Open LLM Leaderboard (Beeching et al., 2023). - Llama-3 8B outperforms Falcon 180B (≈22× the parameters) on the same leaderboard. - Denil et al. (2014): a small subset of weights can predict 95% of all weights in a network, implying high redundancy. - Inference-time compute improvements estimated to yield 5×–20× gains over base post-training performance, at minimal footprint relative to pre-training cost (Davidson et al., 2023). - Scaling laws for downstream task performance are inconsistent or murky (Ganguli et al., 2022; Schaeffer et al., 2023, 2024) and scaling law datasets frequently contain fewer than 100 model data points (Ruan et al., 2024).

Figures & tables: Figure 1 (training cost log-scale 2016–23, axes labeled, sourced from Epoch/AI Index 2024 — no error bars, observational). Figure 3a/3b (Open LLM Leaderboard trends for small vs. large models — axes exist but are not described textually in sufficient detail to audit; no error bars, no statistical significance reported; selection methodology for "notable" models not specified). Figures 4 and 5 are illustrative/conceptual (no data axes). No formal tables. Statistical rigor in figures is absent throughout.

Follow-up references: - Kaplan et al. (2020) — foundational scaling laws this essay directly contests. - Schaeffer et al. (2023/2024) — empirical challenge to emergent abilities and downstream scaling predictability. - Ho et al. (2024a/b) — quantitative treatment of algorithmic progress in LMs. - Davidson et al. (2023) — the primary source for the 5×–20× inference-time compute gain claim; essential to evaluate that figure independently.

Third pass — critique

Implicit assumptions: - Open LLM Leaderboard rankings are an unbiased measure of general capability; if the leaderboard is gamed or narrowly scoped, the central small-beats-large trend is undermined. - The observed parameter-efficiency trend will continue; the essay does not address whether it could reverse with a new architecture class. - Inference-time compute is economically cheaper than training-time compute in relevant deployment settings — not demonstrated; inference at scale can dominate cost. - A "new architecture" will emerge that enables continual learning — asserted in §5.2 without candidate specification.

Missing context or citations: - No engagement with Hoffmann et al. (2022) "Chinchilla" (compute-optimal training), which is arguably the most direct prior work on the compute–data tradeoff and central to any serious treatment of scaling law limitations. - No discussion of OpenAI o1/o3-class reasoning models as concrete existence proofs of inference-time scaling — conspicuously absent given the essay's thesis. - The claim that frontier labs have "stopped publishing" (§1) is made without citation and is contested by available evidence. - No engagement with literature on neural scaling in domains where power-law scaling has been robust (e.g., the code-generation figure of 10 orders of magnitude is mentioned but not balanced against the essay's broader skepticism).

Possible experimental / analytical issues: - The Open LLM Leaderboard data is presented as a trend without controlling for: submission selection bias (authors choose when to submit), benchmark saturation, or data contamination — all known confounds on this leaderboard. - The 5×–20× inference-time compute gain comes from one source (Davidson et al., 2023) studying "a subset" of techniques; range, conditions, and task domains are unspecified in the essay. - Denil et al. (2014) results on weight predictability were obtained on architectures far smaller than modern LLMs; direct applicability is not argued. - The claim that scaling laws rest on <100 data points (Ruan et al., 2024) is stated without nuance about which specific laws or papers this applies to. - As a position essay, no methodology section exists; claims are not falsifiable from the text alone.

Ideas for future work: - Controlled comparison of inference-time compute vs. equivalent training-time compute budgets across matched task types and model families, with cost measured in wall-clock time and dollars. - Systematic replication of the Open LLM Leaderboard small-beats-large trend on held-out, contamination-audited benchmarks to verify it is not an artifact of benchmark saturation or submission timing. - Empirical study of which governance compute thresholds would have correctly flagged historically dangerous capability jumps, to test whether threshold-based policy is as flawed as argued. - Architecture search experiment explicitly targeting continual learning desiderata (minimal catastrophic forgetting, local weight updates) to operationalize the essay's claim that a successor to Transformers is needed.