Why AI Is Moving Beyond Parameter Counts

William Palmer

GLM 5.3 just landed at number two on Terminal Bench and number three on Legal Bench. And here is the kicker that ought to make every AI researcher rethink their budget. It uses the exact same base model architecture as GLM 5.2. Zero new parameters were added during pre training. Not a single one. The entire jump in performance came from roughly one single month of post training reinforcement learning.

William Palmer

Now, if you have been following machine learning for a while, you probably remember the classic Chinchilla scaling laws. Back then, the playbook was simple. You held a fixed ratio of about 200 to 900 tokens per parameter, you scaled up your raw parameter count, and you got a smarter model. But in this new inference inflection world, those old assumptions fall apart completely. Memorization needs parameters, sure. If you want a model to store millions of facts, you need raw parameter volume. But deep, multi step reasoning? Carrying a complex causal chain over 20 plus inference steps without losing the thread? That does not live in total parameter count once you hit a baseline knowledge threshold. That comes from effective post training depth and how you allocate compute during execution.

William Palmer

Professor Jie Tang over at Z.ai put out a post that really crystallizes this shift. He argues that parameter count on its own is basically a dead shorthand. Instead, he lays out five distinct knobs you have to tune together. First, your pre training data volume. Second, your post training reinforcement learning depth. Third, your Mixture of Experts sparsity, which people are now denoting with that XA YB notation. Fourth, who is actually running the model and under what execution runtime conditions. And fifth, how much compute you throw at inference time.

William Palmer

As someone who spends weekends working on race car engines and tuning chassis telemetry, this hits incredibly close to home. Obsessing over parameter count is like bragging about putting a massive seven liter V8 engine into a car, but completely ignoring the transmission, the gear ratios, and the suspension telemetry. If your chassis cannot put that horsepower down to the asphalt, all that raw engine volume just turns into smoke and wasted thermal energy. The post training harness and the execution environment, that is your transmission and your aerodynamics. That is what actually converts raw compute power into lap times on real software engineering tasks.

William Palmer

So how did Z.ai actually build these long horizon post training environments for GLM 5.3? They did something fascinating. They moved away from hand crafted coding puzzles and built automated pipelines that synthesize entire environments end to end. Their research agents study how real engineering work gets done in production. Then they construct fully executable environments that represent several days of work for a senior engineer. We are talking about giving the model full access to actual compute clusters, storage systems, internal documentation, code repositories, and experiment logs. The model gets tasked with diagnosing bottlenecks across an entire training stack, running experiments, and delivering a measurable speedup without breaking correctness.

William Palmer

Now, if you have worked with reinforcement learning, your immediate question is probably, how do you prevent reward hacking in a setup like that? If the model creates its own shortcuts, it will game the metric every single time. Z.ai solved this by synthesizing the verifiers without giving them access to the reference solution. Then they test those verifier agents against three rigorous checks. An oracle check, a no op check, and an unsolved state check. Only when a verifier passes all three does it generate a binary reward signal clean enough to train on directly.

William Palmer

And this shift to harness level RL is happening across the open weights landscape right now. Look at Microsoft's Agent Lightning release. They took a standard Qwen 3.5 9B model, hooked it up to an endpoint proxy for RL through the harness, and fed it just 6,000 training examples. Without touching the base pre trained weights, its score on SWE Bench Verified shot up from 41.8 percent all the way to 56.4 percent. That is a massive jump coming purely from learning how to navigate tools, manage state, and maintain causal chains inside the harness environment.

William Palmer

Standing back and looking at all of this, it feels like we are stepping into a fundamentally new phase of AI systems engineering. We have synthetic research agents collecting work patterns, synthesizing multi step environments, running synthetic judge agents to verify the setups, and then training synthetic solver models. The actual job of the human machine learning engineer is shifting away from hand crafting datasets or tuning base training runs. We are becoming environment architects, designing the playgrounds and the reward bounds where these recursive learning loops take place. And if GLM 5.3 is any indication of what happens when you turn that dial for just 30 days, the next year of post training scaling is going to be wild.