After this explainer you canDescribe the two stages of training, say where training data comes from and why it is disputed, and name what laws ask makers to publish.
The short answerA model learns by reading huge amounts of text and practising guessing the next token, a word or part of a word. Then people teach it to be helpful and safe.
In simple words
- First it practises on huge amounts of text, pictures or code.
- Then people show it good answers and rate its replies.
- Training runs on thousands of computer chips and costs a lot of money.
- Some writers and publishers say their work was used without permission.
The short answerDevelopers show a model huge amounts of data and let it adjust itself until its predictions improve. Then they fine-tune it with examples, people’s ratings and rewards for correct answers, so it follows instructions.
In simple words
- Pre-training: the model learns from huge amounts of text, images or code.
- Fine-tuning: examples, ratings and rewards teach it to follow instructions and refuse harmful requests.
- Laws use the amount of training computing as a rough sign of how capable a model is.
- Where the training data comes from is a growing legal fight.
The short answerPre-training fits the parameters to minimise next-token prediction loss by gradient descent; post-training then aligns behaviour with demonstrations, preference learning and reinforcement learning. Laws measure total training compute; for a dense transformer, it is roughly six times parameters times training tokens.
Key points
- Pre-training minimises cross-entropy loss on the next token, using gradient descent.
- Compute for a dense transformer is roughly 6 × parameters × training tokens.
- Post-training adds supervised examples, preference learning and reinforcement learning on checkable tasks.
- Laws use compute thresholds: 10²⁵ in the EU, 10²⁶ in California, New York and Connecticut.
- 1DataWeb pages, books, code, licensed archives
- 2Pre-trainingGuess the next token, trillions of times, for weeks on thousands of chips
- 3Fine-tuningGood examples, people’s ratings, and rewards for checkable answers
- 4ReleaseTested, then offered in an app, through an API or as a download
Training compute, in operations · compute ≈ 6 × parameters × training tokens (dense models)
- EU, 1025 Above this, a general-purpose model is presumed to have “systemic risk”, the EU’s term for the most powerful models.In force
- California, Connecticut, New York, 1026 Above this, a model counts as a frontier model, one of the most powerful (explainer 5).CA in forceCT in forceNY from Jan 2027
- Llama 3 405B, 3.8 × 1025 Meta’s largest Llama 3 model, for scale. Above 1026, both lines apply.
The lines come from the laws in our AI rules checker. Each step on the scale is ten times more computing.
Words to know
- Training data
- The text, pictures and code a model learns from.
- Pre-training
- The first, long stage: the model practises guessing the next token.
- Fine-tuning
- The second stage: people teach the model to be helpful and safe.
Step 1: practice (pre-training)
During training, the model guesses the next token a huge number of times and checks itself against the real text. Each wrong guess changes the model a little, so it does better next time.
This first stage is called pre-training. It takes weeks or months on thousands of powerful computer chips.
Step 2: learning to help (fine-tuning)
The second stage is called fine-tuning. People show the model examples of good answers and rate its replies.
This teaches it to follow instructions and to say no to harmful requests. It does not always work, so a model can still make mistakes or say harmful things.
Whose data?
The data comes from the web, books, code and other sources. Some newspapers and artists have gone to court over this, and some AI companies now pay for data.
New laws also ask AI makers to publish a summary of the data they used.
Try it yourself
Give a chat app a clear instruction, such as “Answer in exactly three bullet points.” Following instructions like this is what fine-tuning taught it.
Check yourself3 quick questions
During training, what happens after the model makes a wrong guess?
A model has finished its first stage of training. It knows a lot, but it often ignores instructions. What comes next?
A photographer learns that their photos were used to train a model without permission. What does the explainer say about cases like this?
Words to know
- Training data
- The text, pictures and code a model learns from.
- Pre-training
- The first, long stage: the model practises guessing the next token.
- Fine-tuning
- The second stage: people teach the model to be helpful and safe.
Step 1: pre-training
In pre-training, the model reads a vast collection of text, code or images. It keeps guessing the next or missing piece, checks the real one, and changes its parameters a little to do better next time.
This takes weeks or months on thousands of chips, and it has long been the most expensive step. The amount of computing is counted in operations: simple calculations, such as adding or multiplying two numbers. Epoch AI tracks this number for well-known models.
Step 2: fine-tuning
A pre-trained model knows a lot but is not yet a helpful assistant. Developers train it further on examples of good answers, and on people’s ratings of its answers.
This step is meant to teach it to follow instructions, refuse harmful requests and say when it is unsure. It does not always succeed: OpenAI found that GPT-4 became worse at judging its own certainty after this step.
Newer reasoning models also practise on problems whose answers a computer can check, such as maths and code. They are rewarded when they get them right, as DeepSeek described for its R1 model.
Where the data comes from
Training data includes public web pages, books, code, licensed archives and data the developer creates. Publishers and artists have sued over the use of their work, and some now sell licences instead.
Training a model on another model’s answers is called distillation. Some AI companies’ terms, such as Anthropic’s, forbid using their services to train a rival model. Labs accuse each other of distillation, and it is one of the stories we follow.
What the law asks
In the EU, makers of general-purpose models must publish a summary of their training content and respect publishers’ opt-outs. California asks developers to post a summary of their training data.
Laws also use training compute to find the most powerful models. The EU line is 10²⁵ operations, which means 1 followed by 25 zeros. California, New York and Connecticut use 10²⁶, ten times more.
Try it yourself
Give a chat app a clear instruction, such as “Answer in exactly three bullet points.” Following instructions like this is what fine-tuning taught it.
Check yourself3 quick questions
Which step teaches a model most of what it knows?
Company A trains its small model on thousands of answers from Company B’s model. What is this called?
An AI company releases a general-purpose model in the EU. What must it publish about training?
The objective
Pre-training minimises cross-entropy loss: the model is penalised by how unlikely it found the true next token. Backpropagation computes how each parameter should change, and an optimiser such as Adam applies small updates.
Repeated over trillions of tokens, this turns random weights into a model of language, code and the world described in its data.
Compute, data and scaling
A widely used rule of thumb puts training compute for a dense transformer at about 6 × parameters × training tokens operations. Scaling-law research (OpenAI, 2020) found loss falls smoothly as compute grows.
Worked example: DeepMind’s Chinchilla had 70 billion parameters and 1.4 trillion training tokens. The rule gives 6 × 7 × 10¹⁰ × 1.4 × 10¹² ≈ 5.9 × 10²³ operations, far below the EU’s 10²⁵ line.
The Chinchilla study (2022) found that, for a fixed budget, parameters and data should grow together, at roughly 20 tokens per parameter. That optimum ignores the cost of running the model afterwards.
Smaller models are cheaper to run, so it can pay to train them far longer. Meta trained Llama 3 8B on over 15 trillion tokens, against a Chinchilla optimum of about 200 billion.
Post-training
Supervised fine-tuning on demonstrations comes first. Preference learning follows: reinforcement learning from human feedback, as in OpenAI’s 2022 InstructGPT paper, or simpler methods such as direct preference optimisation (Rafailov and others, 2023).
Reasoning models are also trained with reinforcement learning on tasks whose answers can be checked automatically, such as maths and code. DeepSeek reported in 2025 that its R1-Zero learned long step-by-step reasoning this way, without supervised examples first.
Data, distillation and the law
Distillation trains a student model on a teacher model’s outputs: its full probabilities when the teacher is in-house, or usually just its text through an API. Labs try to stop rivals from distilling their models: OpenAI says it stopped Moonshot-linked users from extracting its models’ hidden reasoning, though it gave no evidence for the link.
EU copyright law lets rights holders opt out of text and data mining (Directive 2019/790, Article 4(3)), and the AI Act asks for a public training-content summary. California’s AB 2013 requires dataset documentation, and Connecticut privacy notices must say whether personal data trains language models.
Thresholds: the EU presumes systemic risk above 10²⁵ operations of cumulative training compute (Article 51). California’s SB 53, New York’s RAISE Act and Connecticut use 10²⁶, and California’s count includes later fine-tuning and reinforcement learning.
Try it yourself
Give a chat app a clear instruction, such as “Answer in exactly three bullet points.” Following instructions like this is what fine-tuning taught it.
Check yourself3 quick questions
A dense model with 10 billion parameters is trained on 1 trillion tokens. Roughly how much training compute is that?
By the Chinchilla finding, about how many training tokens suit a 10-billion-parameter model on a fixed budget?
A lab trains a model with 8 × 10²⁵ operations, then adds reinforcement learning worth 3 × 10²⁵. Does it cross California’s 10²⁶ line?
Sources
- AI Act, Article 53 (EU AI Act Service Desk)
- AI Act, Article 51 (EU AI Act Service Desk)
- Data on notable AI models and their training compute (Epoch AI)
- Civil Code section 3111 (California Legislature)
- Scaling Laws for Neural Language Models (Kaplan and others, 2020, arXiv)
- Training Compute-Optimal Large Language Models (Hoffmann and others, 2022, arXiv)
- Introducing Meta Llama 3 (Meta)
- The Llama 3 Herd of Models: 3.8 × 10²⁵ operations for the largest model (Meta, 2024, arXiv)
- Training language models to follow instructions with human feedback (Ouyang and others, 2022, arXiv)
- Direct Preference Optimization (Rafailov and others, 2023, arXiv)
- DeepSeek-R1: reasoning through reinforcement learning (DeepSeek-AI, 2025, arXiv)
- GPT-4 Technical Report (OpenAI, 2023, arXiv)
- Distilling the Knowledge in a Neural Network (Hinton and others, 2015, arXiv)
- Commercial terms of service, section D.4 (Anthropic)
- SB 53 (California Legislature)
This explainer was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it lists. The live parts come from our checked stories, trackers, model comparison and rules checker.