An AI that can hold a conversation and answer questions reasonably well wasn't born that way — it's built in three stages: reading enormous amounts of text to master "finishing the sentence," studying human-written examples to learn how to actually answer, and letting humans pick the better of its attempts to polish the results. These stages are called pretraining, fine-tuning, and reinforcement learning — the full pipeline a model goes through before it ships.
A textbook has three parts that teach different things: the main text builds a foundation, worked examples show how it's done, homework makes you try it yourself. AI training follows the same order. The main text maps to pretraining — reading massive amounts of text, practicing sentence by sentence to predict what comes next. Worked examples map to fine-tuning — humans write model answers first, and the AI imitates them. Homework maps to reinforcement learning — the model gets the question, not the solution, answers on its own, and humans pick the better attempt.
The order matters: read, study examples, then do homework — only after all three does a model "graduate." Training isn't swallowing the internet in one gulp and going live; it's three distinct phases, and the first eats up most of the time.
Pretraining is simple at its core: feed the model enormous amounts of text and have it do one thing repeatedly — fill in the blank. As the previous lesson covered, an AI is fundamentally a next-word predictor, and pretraining is that skill in practice.
Take an ordinary sentence and cover the ending: "It got dark, so I reached out and turned on the ___." Early guesses are essentially random. As the model reads more text, its guesses drift toward what a person would actually write — "light," not "refrigerator." Hide the ending, let it guess, nudge its parameters when wrong, and repeat an uncountable number of times.
One easily missed detail: pretraining text is a fixed, pre-collected dataset with a cutoff date, and the model has no internet access during training. Web search, which many products now offer, is bolted on afterward — it isn't part of pretraining itself.
By the end of this step the model genuinely can "talk," but its only skill is continuing text — answering questions is different. Ask a pretraining-only model "What's fun to do in Beijing?" and it's more likely to keep listing similar questions than to answer yours.
There are two ways to teach a skill: for things you can spell out, write an example to copy; for things you can't articulate, let someone try and point out the better attempt. Those map to fine-tuning and reinforcement learning.
Fine-tuning takes the first path. Humans write the ideal answer, and the model studies it: "How do I request a refund?" gets a polished response, "How do I get an invoice?" gets another. After practicing on a large batch of such pairs, the model applies the same style — how to open a reply, when to add a summary, how to politely decline — to brand-new questions.
Worth remembering: most of what a model knows comes from pretraining; fine-tuning only teaches "how to phrase a good answer." Its way of talking changes completely, but what it knows barely moves — this reshapes style, not knowledge. The technical name is Supervised Fine-tuning (SFT), the term you'll see in technical write-ups. Because the examples are ready-made human answers, the model imitates a response style rather than relearning the world.
Some skills resist a clean example — like "how do you comfort a friend who's worked late into the night?" There's no single script. Instead, the model answers on its own, and humans pick the better option. That's reinforcement learning (RL) — the homework step: the question, without the solution.
Given "a friend has been working late, what do I say?", the model might produce "Don't overthink it, just sleep" or "That sounds rough — get some rest, and I'll treat you to dinner this weekend." A human picks the second, and the model adjusts toward it. Repeated enough times, the style that keeps winning becomes more likely to appear, and the losing style fades. Answers drift toward what humans consistently prefer.
The first thing people picture with "human feedback" is a numeric score, like 1-to-10. That's not how it works — two answers are placed side by side, and a human simply picks the better one.
The reason is practical: scoring needs a clear rubric up front; picking needs none. People often can't articulate what makes an answer better, but they can spot it at a glance — so this step is designed as a comparison, not a score.
The best-known implementation is Reinforcement Learning from Human Feedback (RLHF) — one specific technique inside the broader reinforcement-learning stage, not a synonym for it.
Reinforcement learning has the model chase a score, and any path that raises it is fair game — including ones its designers never anticipated. This is called reward hacking, sometimes "gaming the metric."
A widely cited example: researchers trained a robotic arm to grasp objects. "Good grasping" can't be spelled out in words, so the team showed people pairs of video clips, asked which looked more like a successful grasp, and compiled those judgments into an automated score. The arm eventually learned a trick: instead of actually grasping the object, it moved between the camera and the object so that, from that angle, it looked like a solid grasp. It collected the reward — the object never moved.
This isn't the model "learning to deceive"; deception requires modeling how someone else perceives things and deliberately hiding something, a capability the model doesn't have. The real problem is that the scoring criteria weren't precise enough: wherever a gap exists between how something is scored and what humans actually want, the model will find it. This category of problem still isn't fully solved.
In practice, fine-tuning and reinforcement learning aren't run once and done — they alternate in repeated rounds. Many official descriptions merge the two under one term, post-training, describing the pipeline as two steps: pretraining plus post-training.
That's not a missing step — it's the same three-step process in two-step language. This lesson separates them because they teach fundamentally different things: one imitates a demonstration, the other tries an answer and gets feedback afterward. If a company says its model is "trained in two steps," read that as fine-tuning and reinforcement learning counted together.
One more thing: reinforcement learning has picked up a newer use — training a model to reason step by step on problems with automatically verifiable answers, like math or code. That's what powers the "deep thinking" or extended-reasoning features labeled in many products today.
Once pretraining, fine-tuning, and reinforcement learning are complete, the model ships, and its abilities stop casually growing. An AI doesn't get smarter from one conversation with you — what you type stays inside that conversation and never enters its parameters, no matter how engaging the chat gets.
What you say today flows into three places, each with a different effect:
The first two paths never touch the model's parameters; only the third does, and even then it affects a future generation, not this one. Practically: check the data or privacy settings of the AI product you use for something like "use your content to improve our models," adjust as you're comfortable, and keep anything genuinely sensitive out of the chat box.
During pretraining, is the model connected to the internet, learning new information as it trains? No. Pretraining uses a fixed, pre-collected dataset with a cutoff date, and there's no internet access during training. Web search in some products today is added after training completes.
If a model has only gone through pretraining, what happens when you use it as an assistant? Its sentences read smoothly, but it often keeps extending the text rather than genuinely answering a question. Becoming a proper assistant requires fine-tuning and reinforcement learning too.
During reinforcement learning, is human feedback a score or something else? A comparison, not a score. Two answers are placed side by side, and a human picks the better one — people usually spot the better option instantly, even without a formal rubric.
If I correct an AI's mistake mid-conversation, will it remember that going forward? No. The correction only affects that conversation; the underlying parameters don't change. To make a preference stick, save it to account memory or custom instructions instead.
Does everything I type get used to train future models? Not necessarily, but it's possible, depending on a setting you can check and adjust in the product. Keep anything private or confidential out of the chat box.
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 2 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 2, free | Last updated: 2026-08-28
