unicornAll work

How generative AI works

What an LLM really is, and where its “character” comes from — pretraining, RLHF and Constitutional AI.

In a nutshell When people say "AI" today they usually mean large language models (LLMs) — what sits behind Claude or ChatGPT. Not an all-knowing mind and not a simple autocomplete, but a system that has learned, on a huge amount of text, to build a plausible continuation. And its manner and "character" aren't an accident — they're a result of how it was trained.

What we call "AI"

The word is fuzzy: it's used for a chess engine and for a feed's recommendations. Here it means the narrow but most visible thing today — the LLM:

An immediate consequence: the model works with language and patterns, not "truth". It's very good at guessing how text usually continues. Sometimes plausible matches true, sometimes it doesn't.

Analogy. Someone who has read impossibly much and remembers how fluent writing on any topic sounds — but with no internet at hand and no habit of saying "I don't know". Ask, and they'll confidently continue in tone, because that's how it's usually written. Often right. Sometimes elegant but off.

Where the model's "character" comes from

Claude has a recognisable manner: polite, careful, honest. That's not a "personality" but a result of training in several stages.

Pretrainingreads all the text Fine-tuning · RLHFlearns to converse Constitutionchecks against principles
Raw text-predictor → conversationalist → principled conversationalist.

1. Pretraining. The model "reads" a huge corpus and tunes billions of weights to predict the next piece well. It absorbs language, facts, styles, code — but it's still a "raw" predictor with no manners.

2. Fine-tuning with people (RLHF). Next it's taught to be a conversationalist: people show good and bad answers, rate options, and the model is tuned to produce what's useful and appropriate. This is where "character" forms — politeness, admitting uncertainty, refusing harm.

3. Constitutional approach (Claude). Anthropic gives the model a set of principles (a "constitution" — be honest, safe, respectful) and it learns to check its answers against them, largely correcting itself. Hence Claude's careful, principled tone.

On character: the manner is a choice by its makers, baked in by training. Different models are "raised" differently, so they answer the same question in different tones and with different limits.

Why it matters

Source

Anthropic Academy course AI Capabilities and Limitations, lessons "What We Mean by AI" and "How AI Gets Its Character".

All theory