Pre-Training Gives LLMs Their Capability. Post-Training Gives Them Their Behavior.

The base model provides the capability. Post-training gives it a character.

(This image was generated using AI technology.)

Not a member? Read this story for free here.

The model you use has two biographies — and one of them you can help shape

When a language model refuses a request, it has not necessarily reached the edge of what it can do. When it adopts a helpful tone, follows a specific format, or avoids certain topics, those behaviors often come from a separate process entirely — one that many practitioners collapse into a single story about “how AI learns.”

That distinction matters if you evaluate models, fine-tune open-source systems, design AI products, or simply try to understand why two models with similar benchmark scores can feel so different in practice.

The more useful framing is this:

Modern language models are not shaped by one learning process, but by two very different phases.

Phase One: The Capability Engine

Pre-training is the part you’ve probably heard about: a model trained on vast amounts of text, learning to predict the next token. That description is accurate but undersells what actually happens.

The model doesn’t just learn language patterns. It develops implicit representations of causality, factual relationships, and the internal logic of fields like law, medicine, and code. Not because it was explicitly taught these things, but because predicting text well enough requires them. The next token isn’t arbitrary; it follows from everything before it, including meaning.

This phase is expensive. Compute costs for frontier models run into hundreds of millions of dollars. It’s the domain of large labs: OpenAI, Anthropic, Google, Meta, Mistral. The result is a base model — enormous latent knowledge, no particular disposition to help or harm, no consistent persona.

Raw capability.

Phase Two: Where Character Is Formed

The base model isn’t what you interact with. What gets deployed — every commercial API, every chat interface — has been shaped by a second process: post-training.

Post-training is a family of techniques, not a single step.

Supervised fine-tuning, or SFT, is usually the first layer. Human annotators write examples of good responses: clear, helpful, appropriately formatted. The model trains on these examples and learns the basic habits of an assistant: follow instructions, stay on topic, answer in a conversational register, and respect the format requested by the user. SFT does not add much new knowledge. It teaches the model how to use its existing capabilities in a more useful way.

RLHF — reinforcement learning from human feedback — goes further. Rather than showing the model only examples of good outputs, human raters compare pairs of responses and indicate which one they prefer. These preferences train a reward model: a separate model that predicts which outputs humans are likely to judge as better. The language model is then updated to produce responses that score well according to that reward model. In simple terms, SFT teaches the model by demonstration; RLHF teaches it by preference. (I’ve written separately about why reinforcement learning became such a powerful idea in AI; the important point here is how that logic gets applied to language model behavior.)

OpenAI’s 2022 InstructGPT paper made the stakes concrete: a 1.3B parameter RLHF-trained model was consistently preferred over a 175B base model in direct comparison. The behavioral shaping mattered more than raw scale.

RLHF works, but it is complex. Human preference data is expensive. Reward models can be unstable. And optimizing too aggressively for a learned reward signal can produce strange behavior. DPO — direct preference optimization — addresses part of this complexity. It still uses preference pairs, but it removes the separate reward-model training step. Instead, it updates the language model directly toward preferred responses and away from rejected ones. That makes the pipeline simpler and easier to use, which is one reason DPO became popular in open-source fine-tuning.

Anthropic’s Constitutional AI takes a different route. Instead of relying primarily on human raters, it gives the model a set of written principles and uses those principles to critique and revise responses. The important shift is not only technical, but philosophical: the desired behavior is made more explicit. You are no longer just asking, “Which response do humans prefer?” You are asking, “Which response better follows these stated principles?”

These methods are not cleanly separated in practice. Modern post-training pipelines often combine supervised examples, preference data, synthetic data, model-generated critiques, safety evaluations, tool-use training, and domain-specific fine-tuning. But they share the same broad goal: turning a raw capability engine into a model that behaves in a particular way.

What Jailbreaks Reveal

The distinction between pre-training and post-training shows up in practice in a specific and revealing way.

Jailbreaks work by trying to circumvent post-training constraints.

When a model refuses a request, that refusal does not always mean the underlying capability is absent. In many cases, post-training has shaped the model not to surface certain kinds of information under normal conditions. A jailbreak is an attempt to get around that shaping, to push the model toward behavior its alignment layer was trained to avoid.

This tells you something specific: the model you interact with is not just a bundle of capabilities, but a shaped interface around those capabilities. Two separate layers, and the gap between them is, in a sense, where alignment work lives.

Pre-Training as Commodity

For several years, the dominant assumption was straightforward: more pre-training compute means better models. Scale the data, scale the parameters, scale the compute. The scaling laws seemed to hold without limit.

They still largely do. But the situation has changed in one specific way.

Open-source base models, such as Llama, Mistral, Qwen, have reached a level where the pre-training quality gap between open and closed models has narrowed significantly for a wide range of tasks.

You can download a well-pre-trained base model today and get results that weren’t possible two years ago.

Pre-training has become closer to the floor than the whole game. Increasingly, the visible differentiation happens above it.

Frontier labs are investing heavily in post-training refinement: multiple rounds of RLHF and DPO, careful data curation, domain-specific fine-tuning. Claude feels different from GPT-4o, which feels different from Gemini: the differences trace partly to divergent post-training choices, not just to pre-training scale or benchmark performance.

The personality, the refusals, the particular way each model constructs an answer: all of that is post-training.

Test-time compute — the approach behind OpenAI’s o1 and DeepSeek-R1 — adds a third dimension. These models allocate additional computation at inference time, working through internal reasoning chains before producing an answer. Whether this is post-training, inference-time behavior, or something genuinely new is an open question. What’s clear is that it follows the same underlying logic: capability is the starting point, shaping it is where the work happens.

Why This Matters in Practice

Benchmarks often measure capability. Users experience behavior.

That difference matters when choosing or adapting a model. A high score on MMLU, HumanEval, or a reasoning benchmark tells you something about what a model can do under test conditions. It tells you much less about how the model will behave inside a product: whether it asks useful clarifying questions, follows a house style, refuses too aggressively, formats outputs reliably, uses tools sensibly, or stays consistent across repeated interactions.

For many real-world applications, that second layer determines whether the system is usable.

A customer-support assistant that knows the right answer but sounds evasive is a bad assistant. A research assistant that has strong latent knowledge but cannot cite uncertainty clearly is risky. A sales assistant that understands a company profile but cannot write in the tone your team expects is not ready for deployment.

This is also why fine-tuning is often misunderstood. In most practical settings, you are not trying to “make the model smarter.” You are trying to make its behavior more consistent: in a domain, in a workflow, in a format, or in a voice. Post-training is where that consistency is created.

And that brings us to the part most practitioners can actually influence.

The Part That’s In Your Hands

You can’t change what a model learned during pre-training. That’s fixed in the weights.

But post-training? Parts of it are accessible.

The most immediate lever is the system prompt. You’re not retraining anything, but you’re operating directly in the space post-training created — shaping tone, persona, focus, constraints. It’s behavioral influence without touching weights. I wrote about my own ChatGPT system prompt in a separate piece; the broader point here is that prompting is not just instruction-giving. It is a lightweight form of behavioral design.

The more technical lever is fine-tuning. LoRA and QLoRA — low-rank adaptation methods — make it possible to fine-tune open-source models on a single GPU, training a small set of adapter weights rather than the full model. This is SFT at small scale, and it produces real results: models that consistently follow domain-specific conventions, adopt particular styles, or stay within defined boundaries. Tools like Hugging Face TRL and Unsloth have made this accessible to practitioners without dedicated ML infrastructure.

The process is the same as what large labs use, applied at a different scope and with different data.

The base model provides the capability. Post-training gives it a character.

And in small but meaningful ways, you can shape what that character becomes.

I write about AI for leaders and professionals who want to understand what’s actually changing. Follow along or explore my full catalog in the **AI Writing Compass**.

This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.

Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!