Training only happen once from everything that I have read. That does mean an AI can not reference new information from the internet.
Another interesting fact I read was no data from the learning is stored. The LLM is a huge matrix of weights (numbers) that is loaded into ram. Only number from the learning, but no data. Here I mean data as in a file of information. The matrix with the weighted numbers is the only data.
Might be of interest to read.
Ah, maybe that is what it did.
I was amused by it feeding my own sentences back to me.
Yes, the tokenizing reduces it to numbers … every token has a reference number.
It can regenerate a sentence from the numbers so it must be able to convert tokens back into strings of letters.
Yes, that is correct. The tokenizer takes the text and converts it into numbers.
The LLM can only predict the most probable token ID and hands that ID back to the tokenizer. The tokenizer then converts that ID back into words — the reverse of text‑to‑token.
Keep it going. We are all learning with you.
What does it do if the question contains a new token that is not in the LLM?
Good question Neville.
Lets look at how an AI is built.
So with a 50k tokens vocabulary the tokenizer can break any word into tokens. Even a made up word.
How many pages of data are tokenized? How about 125 million worth!
(Estimated if data was stored on pages.)
But to answer you question directly. No new tokens can be added after the AI is built. No new tokens, no new learning.
might not fit here-- probably past this- but I liked how Claude AI explained it- goes hand in hand with article @easyt50 linked to.
How I learned (training) — this already happened, it’s over
1. Pretraining
Anthropic fed me an enormous amount of text from the internet, books, code, and other sources. My core task during this phase was simple: predict the next word (technically, “token”) in a sequence, over and over, billions of times. To get good at that prediction, I had to internalize a lot — grammar, facts, reasoning patterns, coding conventions, how arguments work, etc. This is done by adjusting billions of internal numerical parameters (weights) using an algorithm called gradient descent, which nudges the parameters slightly after each prediction to reduce error.2. Fine-tuning / alignment
After pretraining, I went through further training to shape how I behave — not just what I know. This includes:Training on examples of good responses (helpful, honest, harmless)
Reinforcement learning from human (and AI) feedback, where responses are rated and I’m adjusted to produce more of the highly-rated kind
Specific work to make me less likely to produce harmful, deceptive, or low-quality output
The key point: none of this happens while I’m talking to you. My weights are frozen. I don’t update them based on our conversation, and I don’t remember this conversation once it ends (unless you’ve enabled a memory feature that summarizes past chats for future reference — that’s a separate system, not me “learning”).
How I generate a response — this happens right now, live
Your message (plus everything else in this conversation) gets converted into tokens — chunks of text.
Those tokens flow through my neural network — layer after layer of matrix multiplications using those frozen weights from training.
At the end, the network produces a probability distribution over “what token is likely to come next.”
One token gets picked (with some randomness — that’s why I might phrase things differently each time), it gets appended to the sequence, and the whole process repeats to generate the next token. This continues until the response is complete.
So in a sense, I’m not “thinking” the way a person does — I’m doing a very sophisticated form of pattern completion, shaped by everything absorbed during training, applied fresh to your specific input each time.
One honest caveat: even people who build systems like me don’t have complete visibility into why a given set of weights produces a given output — the internal representations are complex enough that “which pattern led to this specific word choice” isn’t always traceable. That’s an active area of research at Anthropic called interpretability.
How does it know what complete is?
So it cant deal with a completely new word, because it can not tokenize it to a known token, and it can not create a new token … they are frozen after the initial model build step.
We might consider relaxing the AI guidelines for posts in this topic only.
What do you think?
It is hard to bring in facts about AI itself , without using AI responses.
The guidelines are here
I have stopped reading this post as there is so much AI in it already and quite often loose the plot between one real contribution and one of AI generation.
Stick with it. It is a great way to learn about how AI works. Understanding it will remove some of the resistance.
It is almost impossible to discuss AI without presenting some AI outputs …
so, I am relaxed about the guidelines within this topic … but we dont want to see any pure AI replies without some human content … this is a discussion, not a robotic school.
Trying but when you have to look up what is LLM to start to follow… outside of my scope. I did sign up for a course on line, but it did not fulfil my requirements, perhaps it’s just me and lack of interest.
I was in the same boat … but I followed @easyt50 a step at a time and asked questions. Now I understand a tiny bit of this huge topic.
No, that is not the case. Think of it more like we have the alphabet. We can use it for any word in our English language. The tokenizer has thousands of tokens, so a new word is no problem.
A new token is not needed for a new word. The tokenizer simply breaks the new word into known tokens it already has.
At least that how I understand it.
So in a totally new case, where it could not decompose the word, it would break it into a sequence of letters?
That is not intuitive, but it may well be how it works. …eg when dealing with acronyms.
So, when you can’t get whatever something about AI is saying, for any reason, reply by asking for clarification. Responses to such requests will be of help to others as well as you. When I’m in need of such clarification, I ask ‘Can you explain that like I’m two years old?’ which means ‘Pleas explain that in the simplest terms possible, and define any technical terms’. Doing so has always worked for me, and I even use that approach when formulating an AI prompt (the means by which I define my question).
It predicts a special “end” token. All based on the patterns the AI has learned. It also has a hard limit as a backup- It is only allowed to use so many tokens.
I’ve tried to read the majority of the post- explaining the token a little more- if this was mentioned, I missed it.
Like @easyt50 said
A new token is not needed for a new word. The tokenizer simply breaks the new word into known tokens it already has.
A token is a chunk of text- maybe a short word or part of a word. for example “cat” = 1 token or “un”+“believ”+“able”= 3 tokens. By breaking down words it can process words technically it has never seen before- typos, slang, other words that contain the parts. etc. By keeping a smaller set of word pieces it can reconstruct almost any word. A very rough estimate would be 1 token per 3/4 of a word.
Just adding this- Tokens also can be/are pieces of images. That is why AI can recognize images that are uploaded.
- Most current AI s are still using convert speech to text (for audio). However they are developing audio using tokens.
How is it that it does not make mistakes and put together tokens into meaningless nonexistant words?
No statistical prediction process is perfect … there should be errors.
Maybe the quantity of text used to train it is so huge that the errors of prediction are attenuated to near zero.?
Have you ever seen it make that sort of error?
Fluency/grammar level errors are rare supposedly. Overwhelming information that is consistent and repetitive. Grammar rules are universal across nearly all text.
Factual accuracy is where more mistakes happen- “What year did X happen” or “what did person Y actually say” because they have a lot fewer training examples & consistency. That is evidently where hallucinations start to occur. So grammar - consistent -easy to predict. But Facts can be contradicted across sources- not as consistent thus more errors