Lilian Weng’s Deep Dive Overturns Three Years of Large‑Model Scaling Law Assumptions

After a 13‑month hiatus, former OpenAI safety research VP Lilian Weng published a comprehensive technical post titled “Scaling Laws, Carefully,” in which she revisits, critiques, and reconstructs the scaling‑law narrative that has guided billions of dollars of large‑model investment.

1. Origin: The GPT‑3 Scaling Law

In 2020, OpenAI researcher Jared Kaplan introduced the seminal Scaling Laws paper, establishing that model loss on a log‑log plot follows a stable power‑law decline with respect to parameter count (N), data volume (D), and compute (C). Kaplan’s resource‑allocation rule states that a ten‑fold increase in compute should be matched by a 5.5‑fold increase in parameters and only a 1.8‑fold increase in training tokens, effectively advocating “big models, modest data.” GPT‑3 (175 B parameters, 3000 B tokens) exemplified this “heavy‑parameter, light‑data” regime, and the industry largely adopted it.

2. Reversal: The Chinchilla Model

In 2022 DeepMind’s Hoffmann team replicated the scaling‑law experiments at a much larger scale, employing three complementary fitting algorithms and covering models up to 160 B parameters. Their comparison of the Gopher model (2800 B parameters, 3000 B tokens) with the newly introduced Chinchilla model (≈¼ Gopher’s parameters, >4× the data) showed that Chinchilla uniformly outperformed Gopher across all evaluation metrics. This led to a revised optimal parameter‑to‑data ratio of roughly 1 : 20, a finding that explains the strong performance of later open‑source models such as Llama and DeepSeek, which follow the balanced growth rule rather than the GPT‑3‑style imbalance.

3. Where Kaplan’s Theory Faltered

The original Scaling Laws suffered two major flaws. First, the experiments only reached 1.5 B parameters, yet the derived power‑law was extrapolated to the trillion‑parameter regime, amplifying any small fitting error across several orders of magnitude. Second, the study counted only non‑embedding parameters, deliberately excluding the large embedding matrices that dominate small‑model parameter budgets. Subsequent 2024 studies that incorporated embedding counts showed the power‑law exponents converging toward Chinchilla’s values, indicating that Kaplan’s conclusions are valid only for a narrow, low‑parameter range.

4. Bugs in the Chinchilla Paper

Epoch AI’s 2024 replication of the Chinchilla fitting code uncovered two critical bugs. Bug 1: the loss was averaged instead of summed, causing the optimizer to believe the model had converged early and truncating the training process. Bug 2: the core exponents α and β were rounded to two decimal places; this tiny rounding error, when exponentiated, produced an illusion of high statistical significance. After correcting these issues, the refined exponents are α≈0.3478 and β≈0.3658, reaffirming the necessity of synchronous, proportional growth of parameters and data.

5. The Imminent Data Wall (2026‑2028)

All scaling‑law derivations assume an unlimited supply of high‑quality, non‑redundant training data. Forecasts, however, predict that the total amount of unique, high‑quality text producible by humans will be exhausted between 2026 and 2028. Beyond that point, models must rely on repeated use of existing data, whose marginal utility decays exponentially. A 2023 “effective data” formula quantified this decay, and a 2026 follow‑up study showed that strong weight‑decay regularization can mitigate over‑fitting caused by data repetition. Consequently, the era of merely stacking compute and parameters is reaching its practical limit.

6. Core Insight: Scaling Laws as Engineering Heuristics

Weng embeds an interactive simulator in her blog that lets readers adjust fitting precision, noise level, and fitting interval. Experiments with the tool reveal that seemingly minor engineering details—such as loss‑value decimal precision, sub‑percent noise fluctuations, and the chosen fitting range—can dramatically alter extrapolated predictions. The overarching conclusion is that scaling laws are not immutable physical truths but highly sensitive empirical guidelines that must be applied with careful attention to methodological nuances.

Weng’s three‑year effort thus reframes the community’s understanding of model scaling, urging practitioners to treat scaling laws as flexible, data‑aware engineering tools rather than universal laws.