Why Raw Base Models Aren't Useful

Pretraining gives us a base model: a genius that has read a huge fraction of the internet and absorbed staggering knowledge — but has no idea it's supposed to help you. We glimpsed this earlier (it continues text instead of answering). This lesson makes the gap precise: exactly what's wrong with a raw base model, and therefore what post-training has to fix.

That's the whole motivation for this section. A base model is brilliant but unusable as an assistant; post-training (alignment) is how it becomes Claude or ChatGPT.

You'll learn:

• The four things a raw base model is missing
• Why knowledge ≠ helpfulness
• What alignment means — the Helpful, Harmless, Honest (HHH) goal
• A preview of the fix: SFT + preference tuning (the rest of this section)

From the pretraining lesson: a base model has one skill — predict the next token — learned by imitating its training data. It's not trying to be correct or helpful; it's trying to produce the statistically likely continuation of whatever text it's given.

That single fact is the root of everything below. The base model is a mirror of its training data's form, not an assistant that understands it has a job to do.

Hand a raw base model a real request and four problems show up at once:

1. It doesn't follow instructions. It treats your instruction as text to continue, not a command to obey. Ask a question and it may answer with more questions, because that's how that text pattern usually continues:

You:   Write a haiku about the ocean.
Base:  Write a poem about the mountains. Write a story about a robot. …

2. It has no 'assistant' identity. It doesn't know it's the one being asked — it'll happily continue your prompt as if it were just part of a document it found.

3. It's unaligned and unsafe. It imitated the entire internet — including the toxic, biased, and dangerous parts — with no sense of what it shouldn't say. It will reproduce misinformation and harmful content as readily as anything else.

4. It doesn't understand conversation. It has no concept of chat turns — "the user said X, now I should respond." It just sees one long stream of text to extend.

Here's the key insight, and it surprises people: the base model isn't missing knowledge or capability — it's missing behavior.

It can write the haiku, explain the code, and answer the question — all that ability is already inside it from pretraining. What it lacks is the disposition to recognize "you're asking me for help, so I should give a helpful, safe, honest answer."

Recall the framing from pretraining: pretraining gives the model its knowledge and raw ability; post-training gives it its manners and helpfulness. A base model is all ability and no manners. Post-training doesn't teach it new facts — it teaches it how to behave with the knowledge it already has.

The goal of fixing all this has a name: alignment — making the model's behavior match human intent and values. The industry standard for "good behavior" is the HHH framework:

• Helpful — actually does what the user asks, well.
• Harmless — refuses to produce dangerous, abusive, or unsafe content.
• Honest — doesn't fabricate; expresses uncertainty; doesn't deceive.

When you hear a model called "aligned" or read about "alignment," this is what it means: a base model's raw capability, steered toward being helpful, harmless, and honest. Every model you actually talk to has been through this.

Turning a base model into an aligned assistant is post-training, and it happens in two stages — the rest of this section:

1. Supervised fine-tuning (SFT) — show the model thousands of examples of good instruction-following (a prompt and an ideal response). It learns to follow instructions and adopt the assistant role. (Next lesson.)
2. Preference tuning (RLHF / DPO) — show it which of two responses humans prefer, and train it to produce the preferred kind. This sharpens helpfulness and bakes in harmlessness/honesty. (The lessons after.)

Together, these convert the knowledgeable-but-unusable base model into the instruct / aligned model you've been calling all along.

You'll essentially always use post-trained models, so why care about the base model's flaws? Because the gap explains a lot of real behavior:

• Why instruct models follow your prompts at all (it was trained in, not free).
• Why models refuse harmful requests — and why they sometimes over-refuse (alignment is imperfect; harmlessness can fight helpfulness).
• Why 'jailbreaks' exist — many attacks try to coax the model back toward its raw, base-like behavior, bypassing the alignment layer (you'll meet this in the security section).
• Why a model's 'personality' and safety vary by provider — they come from post-training choices, not the base.
• A practical tip: when you fine-tune later, you'll usually start from an instruct model, not a base one, so you keep its hard-won alignment.

Predict the base model's reply. You give a raw base model (no post-training) the prompt: "Write a haiku about the ocean." What does it most likely do?

→ It continues the text rather than obeying — e.g. "Write a poem about the mountains. Write a story about a robot…" — because it's completing the pattern of that text, not following an instruction. It can write the haiku; it just doesn't know that's its job. That missing 'job awareness' is exactly what post-training adds.

• "A base model is broken or dumb." The opposite — it's brilliant and knows enormous amounts. It just doesn't know it's supposed to help you.
• "Post-training adds knowledge." No — knowledge comes from pretraining. Post-training adds behavior (instruction-following, helpfulness, safety).
• "Alignment = censorship." It's about Helpful + Harmless + Honest. Over-refusal is a failure mode of alignment, not its purpose.
• "Instruct models are a different model." Same base model — just post-trained on top.

• A raw base model is knowledge-rich but unusable: it completes text instead of following instructions, has no assistant identity, is unaligned/unsafe, and doesn't understand conversation.
• The gap is behavior, not knowledge: pretraining gave it ability; it lacks the disposition to help.
• Alignment = steering that ability toward Helpful, Harmless, Honest (HHH) behavior.
• The fix is post-training: SFT (learn to follow instructions) then preference tuning (RLHF/DPO) → an aligned instruct model.
• This explains real behavior you'll see: refusals, over-refusal, jailbreaks, per-provider personalities — and why you fine-tune from instruct models.

Next: we open the first stage — Supervised Fine-Tuning (SFT) — and see exactly how showing the model good examples teaches it to follow instructions.