{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"what-changed-in-2022","title":"What changed around 2022","url":"https://addaly.com/learn/ai-actually-explained/what-changed-in-2022","position":49,"of":92,"minutes":7,"answer":"The 2022 breakthrough everyone noticed was access; the breakthrough underneath it was scale.","keyIdea":"The 2022 breakthrough everyone noticed was access; the breakthrough underneath it was scale."},"body":{"format":"markdown","text":"## It felt like an overnight arrival\n\nIn November 2022 a chatbot was put on a public website. Within two months it had a hundred million users, which was faster than anything before it. To almost everyone it looked as though AI had been invented that winter.\n\nIt had not. Three separate things had been building, and one of them finally reached the public.\n\n## One: a design from 2017 that could be scaled\n\nIn 2017 a group of researchers published a design for handling sequences of words called the transformer. Its important property was not that it was cleverer in some abstract way. It was that the work could be split across thousands of chips running at the same time, instead of having to grind through a sentence word by word.\n\nThat sounds like an engineering detail. It was the whole thing. It meant that throwing more money and more chips at the problem now translated into better models, which had not reliably been true before.\n\n## Two: scale did not run out of steam\n\nThrough 2018 to 2022 the models got bigger, the training text got bigger, and the electricity bills got enormous. The expectation was that this would plateau. Mostly it did not. Quality kept climbing, and abilities appeared that nobody had specifically built in, like writing usable code or handling a language that barely featured in training.\n\nThis was a surprise to the field, and it is worth sitting with. The recipe was not a new idea about the mind. It was the same idea, made very large. That is uncomfortable for people who expected intelligence to require a deeper insight, and it is a real reason the last few years have been hard to predict.\n\n## Three: a text box with no manual\n\nHere is the part usually missed. A comparably capable model, GPT-3, had existed since 2020. It was available to developers with an account and some code. Almost nobody outside that world ever touched it.\n\nWhat arrived in November 2022 was not a more powerful model. It was a doorway: a free web page, a box, no instructions, no setup, no jargon. Your aunt could use it, and did.\n\nSo the thing everyone noticed was **access**. The thing underneath it was **scale**. Those are different events, and collapsing them is how people end up believing a scientific miracle occurred in one particular quarter.\n\n## What follows from getting this right\n\n**Progress here is partly an industrial story.** Chips, data centres, electricity, water, capital. Whether the next jump happens depends as much on supply chains and power grids as on ideas. That also means the number of organisations that can build a frontier model is small, and it is worth noticing who they are.\n\n**Sudden visibility is not sudden capability.** The same gap exists right now. Things that will feel like an overnight arrival in three years are already working in labs and behind developer accounts today.\n\n**The pace is disputed.** Some serious researchers think scaling continues to deliver for years. Others think the useful text has largely been used and the returns are already bending. Both camps are made of people who understand this far better than any confident article you will read. If someone tells you which way it goes, they are guessing with more vocabulary than you."},"check":{"question":"Why did AI seem to arrive all at once in late 2022?","options":[{"text":"Comparably capable models already existed but were reachable only by developers. A free box on a web page let millions of ordinary people meet one within a few weeks.","correct":true},{"text":"A scientific breakthrough that year made language models work for the first time.","correct":false,"tempting":"Sudden visibility strongly implies sudden invention, and it is hard to believe something this striking was quietly sitting around for two years."},{"text":"Computers finally became fast enough in 2022 to run these models.","correct":false,"tempting":"Hardware progress is real and is genuinely part of the story, but it is gradual, and no threshold was crossed in that particular year."},{"text":"That was the year companies started training on the whole internet instead of small curated datasets.","correct":false,"tempting":"Trained on the internet is the popular one-line explanation, but it was already roughly true of models released in 2020, so it cannot explain 2022."}]},"prev":{"slug":"bigger-kept-working","title":"Bigger kept working, and the argument about why","url":"https://addaly.com/learn/ai-actually-explained/bigger-kept-working"},"next":{"slug":"what-it-cost-to-build","title":"What it cost to build one","url":"https://addaly.com/learn/ai-actually-explained/what-it-cost-to-build"},"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) 'What changed around 2022', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/what-changed-in-2022 (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What changed around 2022","courseTitle":"AI, Actually Explained","module":{"index":5,"slug":"how-it-got-here","title":"How it got here, and what changed in 2022"},"answer":null,"keyIdea":"The 2022 breakthrough everyone noticed was access; the breakthrough underneath it was scale.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"it-felt-like-an-overnight-arrival","runs":[{"text":"It felt like an overnight arrival"}]},{"t":"p","runs":[{"text":"In November 2022 a chatbot was put on a public website. Within two months it had a hundred million users, which was faster than anything before it. To almost everyone it looked as though AI had been invented that winter."}]},{"t":"p","runs":[{"text":"It had not. Three separate things had been building, and one of them finally reached the public."}]},{"t":"h2","id":"one-a-design-from-2017-that-could-be-scaled","runs":[{"text":"One: a design from 2017 that could be scaled"}]},{"t":"p","runs":[{"text":"In 2017 a group of researchers published a design for handling sequences of words called the transformer. Its important property was not that it was cleverer in some abstract way. It was that the work could be split across thousands of chips running at the same time, instead of having to grind through a sentence word by word."}]},{"t":"p","runs":[{"text":"That sounds like an engineering detail. It was the whole thing. It meant that throwing more money and more chips at the problem now translated into better models, which had not reliably been true before."}]},{"t":"h2","id":"two-scale-did-not-run-out-of-steam","runs":[{"text":"Two: scale did not run out of steam"}]},{"t":"p","runs":[{"text":"Through 2018 to 2022 the models got bigger, the training text got bigger, and the electricity bills got enormous. The expectation was that this would plateau. Mostly it did not. Quality kept climbing, and abilities appeared that nobody had specifically built in, like writing usable code or handling a language that barely featured in training."}]},{"t":"p","runs":[{"text":"This was a surprise to the field, and it is worth sitting with. The recipe was not a new idea about the mind. It was the same idea, made very large. That is uncomfortable for people who expected intelligence to require a deeper insight, and it is a real reason the last few years have been hard to predict."}]},{"t":"h2","id":"three-a-text-box-with-no-manual","runs":[{"text":"Three: a text box with no manual"}]},{"t":"p","runs":[{"text":"Here is the part usually missed. A comparably capable model, GPT-3, had existed since 2020. It was available to developers with an account and some code. Almost nobody outside that world ever touched it."}]},{"t":"p","runs":[{"text":"What arrived in November 2022 was not a more powerful model. It was a doorway: a free web page, a box, no instructions, no setup, no jargon. Your aunt could use it, and did."}]},{"t":"p","runs":[{"text":"So the thing everyone noticed was "},{"text":"access","b":true},{"text":". The thing underneath it was "},{"text":"scale","b":true},{"text":". Those are different events, and collapsing them is how people end up believing a scientific miracle occurred in one particular quarter."}]},{"t":"h2","id":"what-follows-from-getting-this-right","runs":[{"text":"What follows from getting this right"}]},{"t":"p","runs":[{"text":"Progress here is partly an industrial story.","b":true},{"text":" Chips, data centres, electricity, water, capital. Whether the next jump happens depends as much on supply chains and power grids as on ideas. That also means the number of organisations that can build a frontier model is small, and it is worth noticing who they are."}]},{"t":"p","runs":[{"text":"Sudden visibility is not sudden capability.","b":true},{"text":" The same gap exists right now. Things that will feel like an overnight arrival in three years are already working in labs and behind developer accounts today."}]},{"t":"p","runs":[{"text":"The pace is disputed.","b":true},{"text":" Some serious researchers think scaling continues to deliver for years. Others think the useful text has largely been used and the returns are already bending. Both camps are made of people who understand this far better than any confident article you will read. If someone tells you which way it goes, they are guessing with more vocabulary than you."}]}],"check":{"question":[{"text":"Why did AI seem to arrive all at once in late 2022?"}],"options":[{"text":[{"text":"Comparably capable models already existed but were reachable only by developers. A free box on a web page let millions of ordinary people meet one within a few weeks."}],"correct":true,"tempting":null},{"text":[{"text":"A scientific breakthrough that year made language models work for the first time."}],"correct":false,"tempting":[{"text":"Sudden visibility strongly implies sudden invention, and it is hard to believe something this striking was quietly sitting around for two years."}]},{"text":[{"text":"Computers finally became fast enough in 2022 to run these models."}],"correct":false,"tempting":[{"text":"Hardware progress is real and is genuinely part of the story, but it is gradual, and no threshold was crossed in that particular year."}]},{"text":[{"text":"That was the year companies started training on the whole internet instead of small curated datasets."}],"correct":false,"tempting":[{"text":"Trained on the internet is the popular one-line explanation, but it was already roughly true of models released in 2020, so it cannot explain 2022."}]}],"order":[2,1,3,0],"why":null},"prev":{"slug":"bigger-kept-working","title":"Bigger kept working, and the argument about why","minutes":10},"next":{"slug":"what-it-cost-to-build","title":"What it cost to build one","minutes":9}}}