{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"what-a-model-actually-is","title":"It is a file, and here is how big","url":"https://addaly.com/learn/ai-actually-explained/what-a-model-actually-is","position":5,"of":92,"minutes":9,"answer":"A trained model is a static file of billions of numbers, containing no sentences and no lookup table, which is why it has a cut-off date, cannot check anything.","keyIdea":"A trained model is a static file of billions of numbers, containing no sentences and no lookup table, which is why it has a cut-off date, cannot check anything, and does not change when you correct it."},"body":{"format":"markdown","text":"## The thing itself\n\nAfter training finishes, what exists is a file. Not a program in the ordinary sense, not a database of answers, not a copy of the internet sitting on a server. A file, containing a very long list of numbers.\n\nThe numbers are called **weights**, or **parameters**. Each one is a single value, usually with a few decimal places, and each does nothing on its own. Arranged in the right order and run through the right arithmetic, they produce the behaviour.\n\nSizes make this concrete. A model described as \"8B\" has eight billion of these numbers. If each number is stored in sixteen bits (two bytes) the file is about sixteen gigabytes. Store each in four bits instead, a compression called **quantisation**, and the same model becomes roughly four and a half gigabytes: small enough to sit on a phone, and small enough to run on a laptop with 8GB of memory using free software such as `llama.cpp` or Ollama. That is a real thing you could do this week, on hardware you probably own.\n\nThe biggest commercial models are not published, so their sizes are guesses. The public open-weight ones run from around half a billion parameters up to several hundred billion. A 400B model needs a server, not a laptop.\n\n> Figure: What a model weighs, at four bits a number. Bars (gigabytes on disk): 1B: runs on a mid-range phone 0.6; 4B: runs on a recent flagship phone 2.4; 8B: runs on an ordinary laptop 4.7; 14B: wants a graphics card 8.2; 70B: wants a server 40. Quantisation stores each parameter in four bits rather than sixteen, so the file is a quarter of the size. An 8B model at about five gigabytes fits on a laptop with 8GB of memory; a 70B model does not.\n\n## What is not in the file\n\nOpen the file in a text editor and you will find no sentences. There is no article, no page, no lookup table, no index of facts. You cannot search it for \"the capital of Kenya\" and find a row.\n\nThis is worth sitting with, because the natural mental picture (that the model has stored the internet somewhere and fetches from it) leads to wrong predictions about behaviour. If the model held the text, it could quote exactly and would never invent. It does neither. What survives training is the *statistical shape* of the material, not the material.\n\nA rough comparison: the text used to train a large model runs into the trillions of words, which is many terabytes. The resulting file is a few hundred gigabytes at most. Whatever happened in between, it was not storage.\n\n## The honest complication\n\nThe clean version above is not quite the whole truth, and the exception is at the centre of the biggest legal fight in this field.\n\nModels do memorise some things word for word. Passages that appeared thousands of times in the training material (a famous poem, a licence header, a common code snippet, a heavily quoted paragraph) can come back verbatim. Researchers have extracted such passages deliberately. It is a small fraction of the material and the models are not designed to do it, but it happens.\n\nSo \"the model does not contain the text\" is true for almost all of the text and false for a little of it. Whether that little makes the model a copy of the works in a legal sense is exactly what courts in several countries are being asked to decide, and they have not agreed. There is a lesson later in this course on that argument. For now, the point is only that the honest statement is more careful than either slogan.\n\n## The file does not change\n\nHere is the practical consequence that surprises people most.\n\nWhile you are talking to a chatbot, the file is not being edited. It is read-only. Ten million people can use the same model at the same time, and each of them is reading the same unchanged numbers. Nothing you type alters it. Nothing you correct is retained in it.\n\nWhen a chatbot appears to remember your name from an earlier conversation, a separate piece of ordinary software stored that name as text and pasted it back in before your next message. That is a product feature built around the model, not a change inside it. You can usually find it, and delete it, in the settings.\n\nThis also explains why a model \"gets better\" only when the company trains a new one and swaps the file. Improvement arrives in lumps, on their schedule, not gradually from use.\n\n## Why this is the useful picture\n\nHold this image and several things fall into place at once. The model has a fixed cut-off, because the file was finished on a date. It cannot look anything up unless the product wraps it in something that can. It gives similar answers to everybody because everybody is reading the same numbers. It can be copied, which is why an open-weight model, once released, cannot be recalled. And running it costs money every time, because those billions of multiplications happen again for every reply.\n\nAll of that follows from the file. Not from intelligence, not from intent: from the fact that the finished product of training is a large, static, unreadable list of numbers that somebody has to run."},"check":{"question":"A user corrects a chatbot's mistake, the chatbot accepts the correction, and the user assumes the mistake is now fixed for everybody. Why is that assumption wrong?","options":[{"text":"The correction is queued for review and applied to the model only if enough other users report the same error.","correct":false,"tempting":"Imagines a voting mechanism that does not exist in the model; feedback may be collected, but it does not edit weights."},{"text":"The correction is applied to the user's own copy of the model, but other users have separate copies that were not updated.","correct":false,"tempting":"Assumes each user has a private copy; everyone is served by the same read-only file."},{"text":"The weights are read-only during use, so the correction lives only in that conversation's text and disappears when it ends.","correct":true},{"text":"Corrections are only retained when the user is on a paid plan, where conversation history is stored.","correct":false,"tempting":"Confuses stored conversation text, which is a product feature, with a change to the trained file."}]},"prev":{"slug":"what-training-means","title":"What training on data actually means","url":"https://addaly.com/learn/ai-actually-explained/what-training-means"},"next":{"slug":"prediction-is-the-engine","title":"Nearly all of it is prediction","url":"https://addaly.com/learn/ai-actually-explained/prediction-is-the-engine"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'It is a file, and here is how big', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/what-a-model-actually-is (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"It is a file, and here is how big","courseTitle":"AI, Actually Explained","module":{"index":1,"slug":"what-the-word-names","title":"What this word actually names"},"answer":null,"keyIdea":"A trained model is a static file of billions of numbers, containing no sentences and no lookup table, which is why it has a cut-off date, cannot check anything, and does not change when you correct it.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"gigabytes on disk","title":"What a model weighs, at four bits a number","series":[{"label":"1B: runs on a mid-range phone","value":0.6},{"label":"4B: runs on a recent flagship phone","value":2.4},{"label":"8B: runs on an ordinary laptop","value":4.7,"highlight":true},{"label":"14B: wants a graphics card","value":8.2},{"label":"70B: wants a server","value":40}],"caption":"Quantisation stores each parameter in four bits rather than sixteen, so the file is a quarter of the size. An 8B model at about five gigabytes fits on a laptop with 8GB of memory; a 70B model does not."},"alt":"Figure: What a model weighs, at four bits a number. Bars (gigabytes on disk): 1B: runs on a mid-range phone 0.6; 4B: runs on a recent flagship phone 2.4; 8B: runs on an ordinary laptop 4.7; 14B: wants a graphics card 8.2; 70B: wants a server 40. Quantisation stores each parameter in four bits rather than sixteen, so the file is a quarter of the size. An 8B model at about five gigabytes fits on a laptop with 8GB of memory; a 70B model does not."}},"blocks":[{"t":"h2","id":"the-thing-itself","runs":[{"text":"The thing itself"}]},{"t":"p","runs":[{"text":"After training finishes, what exists is a file. Not a program in the ordinary sense, not a database of answers, not a copy of the internet sitting on a server. A file, containing a very long list of numbers."}]},{"t":"p","runs":[{"text":"The numbers are called "},{"text":"weights","b":true},{"text":", or "},{"text":"parameters","b":true},{"text":". Each one is a single value, usually with a few decimal places, and each does nothing on its own. Arranged in the right order and run through the right arithmetic, they produce the behaviour."}]},{"t":"p","runs":[{"text":"Sizes make this concrete. A model described as “8B” has eight billion of these numbers. If each number is stored in sixteen bits (two bytes) the file is about sixteen gigabytes. Store each in four bits instead, a compression called "},{"text":"quantisation","b":true},{"text":", and the same model becomes roughly four and a half gigabytes: small enough to sit on a phone, and small enough to run on a laptop with 8GB of memory using free software such as "},{"text":"llama.cpp","code":true},{"text":" or Ollama. That is a real thing you could do this week, on hardware you probably own."}]},{"t":"p","runs":[{"text":"The biggest commercial models are not published, so their sizes are guesses. The public open-weight ones run from around half a billion parameters up to several hundred billion. A 400B model needs a server, not a laptop."}]},{"t":"h2","id":"what-is-not-in-the-file","runs":[{"text":"What is not in the file"}]},{"t":"p","runs":[{"text":"Open the file in a text editor and you will find no sentences. There is no article, no page, no lookup table, no index of facts. You cannot search it for “the capital of Kenya” and find a row."}]},{"t":"p","runs":[{"text":"This is worth sitting with, because the natural mental picture (that the model has stored the internet somewhere and fetches from it) leads to wrong predictions about behaviour. If the model held the text, it could quote exactly and would never invent. It does neither. What survives training is the "},{"text":"statistical shape","i":true},{"text":" of the material, not the material."}]},{"t":"p","runs":[{"text":"A rough comparison: the text used to train a large model runs into the trillions of words, which is many terabytes. The resulting file is a few hundred gigabytes at most. Whatever happened in between, it was not storage."}]},{"t":"h2","id":"the-honest-complication","runs":[{"text":"The honest complication"}]},{"t":"p","runs":[{"text":"The clean version above is not quite the whole truth, and the exception is at the centre of the biggest legal fight in this field."}]},{"t":"p","runs":[{"text":"Models do memorise some things word for word. Passages that appeared thousands of times in the training material (a famous poem, a licence header, a common code snippet, a heavily quoted paragraph) can come back verbatim. Researchers have extracted such passages deliberately. It is a small fraction of the material and the models are not designed to do it, but it happens."}]},{"t":"p","runs":[{"text":"So “the model does not contain the text” is true for almost all of the text and false for a little of it. Whether that little makes the model a copy of the works in a legal sense is exactly what courts in several countries are being asked to decide, and they have not agreed. There is a lesson later in this course on that argument. For now, the point is only that the honest statement is more careful than either slogan."}]},{"t":"h2","id":"the-file-does-not-change","runs":[{"text":"The file does not change"}]},{"t":"p","runs":[{"text":"Here is the practical consequence that surprises people most."}]},{"t":"p","runs":[{"text":"While you are talking to a chatbot, the file is not being edited. It is read-only. Ten million people can use the same model at the same time, and each of them is reading the same unchanged numbers. Nothing you type alters it. Nothing you correct is retained in it."}]},{"t":"p","runs":[{"text":"When a chatbot appears to remember your name from an earlier conversation, a separate piece of ordinary software stored that name as text and pasted it back in before your next message. That is a product feature built around the model, not a change inside it. You can usually find it, and delete it, in the settings."}]},{"t":"p","runs":[{"text":"This also explains why a model “gets better” only when the company trains a new one and swaps the file. Improvement arrives in lumps, on their schedule, not gradually from use."}]},{"t":"h2","id":"why-this-is-the-useful-picture","runs":[{"text":"Why this is the useful picture"}]},{"t":"p","runs":[{"text":"Hold this image and several things fall into place at once. The model has a fixed cut-off, because the file was finished on a date. It cannot look anything up unless the product wraps it in something that can. It gives similar answers to everybody because everybody is reading the same numbers. It can be copied, which is why an open-weight model, once released, cannot be recalled. And running it costs money every time, because those billions of multiplications happen again for every reply."}]},{"t":"p","runs":[{"text":"All of that follows from the file. Not from intelligence, not from intent: from the fact that the finished product of training is a large, static, unreadable list of numbers that somebody has to run."}]}],"check":{"question":[{"text":"A user corrects a chatbot’s mistake, the chatbot accepts the correction, and the user assumes the mistake is now fixed for everybody. Why is that assumption wrong?"}],"options":[{"text":[{"text":"The correction is queued for review and applied to the model only if enough other users report the same error."}],"correct":false,"tempting":[{"text":"Imagines a voting mechanism that does not exist in the model; feedback may be collected, but it does not edit weights."}]},{"text":[{"text":"The correction is applied to the user’s own copy of the model, but other users have separate copies that were not updated."}],"correct":false,"tempting":[{"text":"Assumes each user has a private copy; everyone is served by the same read-only file."}]},{"text":[{"text":"The weights are read-only during use, so the correction lives only in that conversation’s text and disappears when it ends."}],"correct":true,"tempting":null},{"text":[{"text":"Corrections are only retained when the user is on a paid plan, where conversation history is stored."}],"correct":false,"tempting":[{"text":"Confuses stored conversation text, which is a product feature, with a change to the trained file."}]}],"order":[3,1,2,0],"why":null},"prev":{"slug":"what-training-means","title":"What training on data actually means","minutes":7},"next":{"slug":"prediction-is-the-engine","title":"Nearly all of it is prediction","minutes":8}}}