A small model wins, and the paper shows exactly which rung the win comes from

InstructGPT outperforms GPT-3 at 100x smaller scale, and the paper is explicit that this rests on a scale asymmetry repeated throughout the pipeline: the reward model supervises PPO training from below, a judge nearly 30x smaller than the policy it steers, and the constitution later replaces human harm labels in Constitutional AI with roughly ten principles standing in for tens of thousands of comparisons, while the API prompt distribution supplies label-free scale to the cheapest stage of the pipeline. The paper stages its headline claim as a ladder rather than a single number: T0 is scaled to match FLAN so neither baseline wins on data volume alone; FLAN sets the comparison floor for instruction following because its fixed task compilation underrepresents real usage; and the GPT-3-prompted baseline shows the limits of prompting alone, trailing trained instruction-following substantially even when optimized against the very reward model used for RLHF itself. Each rung isolates a different specific reason the previous one falls short. The 2017 paper stages an analogous ladder rung five years earlier, an algorithm-class asymmetry rather than a parameter-count one: A2C closes its Enduro exploration gap with DQN, matching a value-based method's previously reported score not by outperforming it on the merits of policy-gradient learning but because human-shaped preference feedback supplies exactly the dense progress signal random policy-gradient exploration could never find on its own. As with T0-versus-FLAN or the GPT-3-prompted baseline, the comparison isolates one specific bottleneck, here exploration under sparse reward, rather than asserting a general win, and DQN itself was never retrained for the comparison; its number is simply cited from prior work, the same citation-not-rerun structure the paper's other baseline rungs share. CAA adds a rung of its own, six years later and over a technique choice rather than a model or algorithm: few-shot prompting loses out to system-prompting when the paper tests which prompting method should serve as CAA's baseline, a single instructional system prompt beating few-shot examples at steering any of the seven behaviors tested. As with T0-versus-FLAN, the comparison isolates one specific variable, here which prompting technique to standardize on, rather than asserting a general verdict on few-shot prompting; the paper is explicit that it never exhaustively searched either configuration, so the gap it reports is a property of this comparison, not a settled fact about the technique. Every later baseline number the paper reports simply as "prompting" means this winning configuration, system-prompting, because that is the rung the ladder settled on before the paper's real comparisons, against activation steering, begin. The newly-added sparse-autoencoder paper builds its own ladder the same way, rung by rung, addressed at a single question: is the dictionary's interpretability edge real or just an artifact of which decomposition method got which advantages. Default-basis-baseline loses to sparse-autoencoders, the floor rung, no more interpretable than the random-directions baseline it sits beside once outlier dimensions are checked and found not to rescue it. PCA trails sparse-autoencoders on the causal-localization test that matters most, needing more patched components and larger edit magnitudes to reach a given KL divergence on the IOI task, and the paper isolates why: an unsparsified version of its own architecture (alpha=0) loses the advantage over PCA entirely, so the gap is specifically the sparsity penalty's doing, not dictionary learning's greater flexibility. ICA converges with sparse-autoencoders where PCA cannot: the dictionary's autointerpretability edge shrinks through the model's depth until it is merely comparable to ICA by layer 4 and shows minimal improvement by the final layer, the strongest non-dictionary baseline closing most of the gap the weaker ones never do. But the ladder's rungs are not run on equal footing: PCA gets more training data than ICA, fit on the full multi-million-vector activation dataset used to train the autoencoders themselves while ICA, converging far more slowly, is deliberately given only 2GB, a data-budget asymmetry the paper states outright as a reason to read ICA's late-layer catch-up as an even stronger result than the raw numbers show, since it is achieved from a visibly smaller sample.