AI scaling has more than one dial
For years, scaling in AI meant one knob: add parameters, add data, get a better model. Z.ai CEO Jie Tang, promoting the release of GLM 5.3, used the week's clearest single argument for why that's no longer the whole story — a parameter count only tells you something "alongside three others: how much data you have, where you intend to spend your compute, and who will run the model, under what conditions." Scaling, on this account, isn't a volume knob anymore. It's a control panel: model size, pre-training data, compute per pass, mid-training, and post-training-plus-RL are five separate dials, and which one is worth turning next depends on where a model already sits.
GLM 5.3 is Tang's own demonstration. It uses the identical architecture and size as its predecessor GLM 5.2 — same base model, nothing added — and its real capability gains came from about a month of reinforcement learning inside long-horizon synthetic environments: simulated multi-day engineering tasks, complete with compute clusters, storage, real codebases, and experiment results, that the model has to work through end to end, diagnosing bottlenecks and delivering a measurable fix. Non-marginal gains, from the post-training-plus-RL dial alone, with the model-size dial left untouched.
The reason that dial matters more now traces back through how the field has actually scaled models over time. Kaplan's 2020 scaling laws pushed toward sheer size — roughly 2.7 parameters for every training token — and models got huge. Chinchilla, in 2022, corrected the ratio the other way, toward roughly 20 tokens per parameter, better balancing size against data. Today's correction is a third one, driven by a fact neither of those eras optimized for: a model gets trained once but run billions of times at inference, so a smaller model can easily justify far more training per parameter than either law assumed — Llama-2-7B trained on roughly 290 tokens per parameter, Gemma-2-9B on roughly 889. The MoE (Mixture-of-Experts) architecture most frontier models now use makes the same point structurally: total parameters are the kitchen's full staff — knowledge capacity, facts, long-tail memory — while activated parameters and reasoning depth are who actually works a given problem, walking it through a chain of steps to a plated answer. Tang's underlying claim is exactly this distinction: advanced skills, like finding a real security vulnerability, depend on holding a causal chain together across twenty-plus reasoning steps without losing the thread, and knowing more facts isn't the same as thinking longer about a hard one.
Two more items this week back the shift independently of Tang's own framing. A SemiAnalysis benchmark comparison spanning three eras of LLM progress — early scaling, reasoning, agentic — finds the time for open models to close each new capability gap has roughly halved with every era, evidence that whatever a frontier lab does in private eventually gets reverse-engineered by everyone else, sharpening the pressure to find the next lever once size and data alone stop being a moat. And this week's top-papers roundup led with "Skaling: Chinchilla's Exponents Meet Kaplan's Coupling", a generalized neural scaling law coupling model size and data through a single interaction exponent, aimed specifically at fixing systematic errors in the Kaplan/Chinchilla-era formulas at the data-scarce, heavily-overtrained regime GLM 5.3's own approach lives in — the academic literature catching up to what practice had already started doing.
The industry isn't done scaling — it's learning which dial to turn next. That's a genuinely different, more interesting story than "bigger model wins," and GLM 5.3 is this week's cleanest evidence a lab can post a real capability jump from post-training and RL alone, without touching model size at all.
