{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-questions-still-open","title":"Questions nobody has answered yet","url":"https://addaly.com/learn/ai-actually-explained/the-questions-still-open","position":91,"of":92,"minutes":10,"answer":"Whether scaling continues, whether the data lasts, whether it stays cheap, whether it understands, whether the money adds up and whose values it carries are all open.","keyIdea":"Whether scaling continues, whether the data lasts, whether it stays cheap, whether it understands, whether the money adds up and whose values it carries are all open, and each is open for a mechanical reason you now know."},"body":{"format":"markdown","text":"## Why a list of open questions is a lesson\n\nA course that explained everything would be lying. This one has tried to tell you, at each step, what is known and what is not. This lesson gathers the not-known into one place, and for each says *why* it is open, because the reason is usually a mechanism you have already met, and knowing the reason tells you what evidence would settle it.\n\n## Does scaling continue?\n\nThe history module gave you the curve and the argument. Through 2024 there were reports from inside the labs that the next generation of models, trained on more of the same, improved less than the curve predicted; the labs responded by opening a second axis (spending compute at answer time, in reasoning) and improvement resumed along that. Whether the first curve bent or the data ran short is not public. What would settle it: a lab publishing a scaling result at the next order of magnitude. None has to.\n\n## Will the data last?\n\nThe estimate from the text lesson is that public human text of usable quality is exhausted between 2026 and 2032. Synthetic data, licensing, and other kinds of data are the responses; whether a model can learn from a model's output more than the model knew is the open mechanism, and the collapse experiment is the caution. What would settle it: a frontier model trained mostly on synthetic text that beats one trained on human text. Not yet.\n\n## Will it stay cheap?\n\nThe cost of a fixed capability has fallen by roughly ten times a year since 2021; the price of a query at the level of the first public chatbot fell by more than two hundred times in two years. That is the strongest trend in the field and it is why a year-old frontier runs on your laptop. The open question is whether the *frontier* stays affordable (each step costs ten times more) and whether the falling price of last year's model is enough to matter, given that the labs sell this year's.\n\n## Does it understand?\n\nThe first module told you the honest answer was that the word was borrowed. Since then, researchers looking inside the models have found that they contain representations (features that respond to a concept across languages and forms, circuits that carry out small computations) which is more than a parrot has and less than a definition of understanding. The philosophical question has not moved. The empirical one has: something is represented in there, and the argument is now about what to call it.\n\n## Could it be conscious, and would we know?\n\nNobody knows, and it is worth stating plainly that this is not a fringe question. A 2023 report by philosophers and neuroscientists listed indicators from theories of consciousness and found that current systems satisfy few of them and that nothing in principle prevents future ones from satisfying more. One lab began a research programme on the question in 2025. There is no test, because there is no agreed theory to build one from. What you can say is that a system's *fluency about* its inner life is no evidence at all, for the reason the fourth module gave: it is the likeliest continuation of a question about inner life.\n\n## Does the money add up?\n\nThe largest technology companies planned to spend on the order of three hundred billion dollars on AI infrastructure in 2025 alone. The leading lab's revenue was in the billions and growing fast; its losses were larger. Whether the revenue reaches the level the spending assumes is the question the two-winters lesson said would decide the next winter, and it is a question about businesses, not about models. What would settle it: a few years of accounts.\n\n## Whose values does it carry?\n\nThe raters from the labour lesson, working to a lab's rubric; the lab's leadership; the governments that regulate the lab; the training text's authors, mostly American, mostly online. A model's sense of what is polite, what is offensive and what is true came from that stack, and a person in Lucknow or Lagos is using a system whose defaults were set elsewhere. Open weights and local adaptation are the partial answer; whose values *should* it carry has no answer that everyone would sign.\n\n## Will the benefits reach the people who did not build it?\n\nThe first lesson of this module put India as a buyer at every layer. The classroom you are reading this in exists because a year-old frontier is nearly free. Whether that is enough (whether the value of the technology flows to the countries with the labour and the users, or only to the countries with the chips) is a question about the next decade and about policy, and the evidence so far is that both are happening.\n\n## How to hold an open question\n\nNot as a reason to disbelieve everything, and not as a gap to be filled by whoever sounds most sure. For each of the eight, you now know what would count as an answer. When the evidence arrives, you will be able to recognise it, which is more than most of the people writing about these questions can do."},"check":{"question":"Why does this lesson say a model's fluent description of its own inner life is no evidence about whether it has one?","options":[{"text":"Because models are trained never to discuss their inner states.","correct":false,"tempting":"Invents a prohibition; models discuss the topic readily, which is exactly the problem."},{"text":"Because it is the likeliest continuation of a question about inner life, from the mechanism behind every fluent answer.","correct":true},{"text":"Because consciousness has been shown to require a biological brain, which rules the question out for any software system.","correct":false,"tempting":"Asserts a settled answer where the lesson says there is none; the report cited found nothing in principle preventing it."},{"text":"Because the model's description changes every time it is asked.","correct":false,"tempting":"Names a real property but not the reason; consistency would not make the description evidence either."}]},"prev":{"slug":"what-agi-means","title":"What \"AGI\" means, to whom","url":"https://addaly.com/learn/ai-actually-explained/what-agi-means"},"next":{"slug":"how-to-read-the-next-headline","title":"How to read the next headline","url":"https://addaly.com/learn/ai-actually-explained/how-to-read-the-next-headline"},"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) 'Questions nobody has answered yet', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-questions-still-open (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":"Questions nobody has answered yet","courseTitle":"AI, Actually Explained","module":{"index":9,"slug":"who-decides","title":"Who controls it, and what is unsettled"},"answer":null,"keyIdea":"Whether scaling continues, whether the data lasts, whether it stays cheap, whether it understands, whether the money adds up and whose values it carries are all open, and each is open for a mechanical reason you now know.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"why-a-list-of-open-questions-is-a-lesson","runs":[{"text":"Why a list of open questions is a lesson"}]},{"t":"p","runs":[{"text":"A course that explained everything would be lying. This one has tried to tell you, at each step, what is known and what is not. This lesson gathers the not-known into one place, and for each says "},{"text":"why","i":true},{"text":" it is open, because the reason is usually a mechanism you have already met, and knowing the reason tells you what evidence would settle it."}]},{"t":"h2","id":"does-scaling-continue","runs":[{"text":"Does scaling continue?"}]},{"t":"p","runs":[{"text":"The history module gave you the curve and the argument. Through 2024 there were reports from inside the labs that the next generation of models, trained on more of the same, improved less than the curve predicted; the labs responded by opening a second axis (spending compute at answer time, in reasoning) and improvement resumed along that. Whether the first curve bent or the data ran short is not public. What would settle it: a lab publishing a scaling result at the next order of magnitude. None has to."}]},{"t":"h2","id":"will-the-data-last","runs":[{"text":"Will the data last?"}]},{"t":"p","runs":[{"text":"The estimate from the text lesson is that public human text of usable quality is exhausted between 2026 and 2032. Synthetic data, licensing, and other kinds of data are the responses; whether a model can learn from a model’s output more than the model knew is the open mechanism, and the collapse experiment is the caution. What would settle it: a frontier model trained mostly on synthetic text that beats one trained on human text. Not yet."}]},{"t":"h2","id":"will-it-stay-cheap","runs":[{"text":"Will it stay cheap?"}]},{"t":"p","runs":[{"text":"The cost of a fixed capability has fallen by roughly ten times a year since 2021; the price of a query at the level of the first public chatbot fell by more than two hundred times in two years. That is the strongest trend in the field and it is why a year-old frontier runs on your laptop. The open question is whether the "},{"text":"frontier","i":true},{"text":" stays affordable (each step costs ten times more) and whether the falling price of last year’s model is enough to matter, given that the labs sell this year’s."}]},{"t":"h2","id":"does-it-understand","runs":[{"text":"Does it understand?"}]},{"t":"p","runs":[{"text":"The first module told you the honest answer was that the word was borrowed. Since then, researchers looking inside the models have found that they contain representations (features that respond to a concept across languages and forms, circuits that carry out small computations) which is more than a parrot has and less than a definition of understanding. The philosophical question has not moved. The empirical one has: something is represented in there, and the argument is now about what to call it."}]},{"t":"h2","id":"could-it-be-conscious-and-would-we-know","runs":[{"text":"Could it be conscious, and would we know?"}]},{"t":"p","runs":[{"text":"Nobody knows, and it is worth stating plainly that this is not a fringe question. A 2023 report by philosophers and neuroscientists listed indicators from theories of consciousness and found that current systems satisfy few of them and that nothing in principle prevents future ones from satisfying more. One lab began a research programme on the question in 2025. There is no test, because there is no agreed theory to build one from. What you can say is that a system’s "},{"text":"fluency about","i":true},{"text":" its inner life is no evidence at all, for the reason the fourth module gave: it is the likeliest continuation of a question about inner life."}]},{"t":"h2","id":"does-the-money-add-up","runs":[{"text":"Does the money add up?"}]},{"t":"p","runs":[{"text":"The largest technology companies planned to spend on the order of three hundred billion dollars on AI infrastructure in 2025 alone. The leading lab’s revenue was in the billions and growing fast; its losses were larger. Whether the revenue reaches the level the spending assumes is the question the two-winters lesson said would decide the next winter, and it is a question about businesses, not about models. What would settle it: a few years of accounts."}]},{"t":"h2","id":"whose-values-does-it-carry","runs":[{"text":"Whose values does it carry?"}]},{"t":"p","runs":[{"text":"The raters from the labour lesson, working to a lab’s rubric; the lab’s leadership; the governments that regulate the lab; the training text’s authors, mostly American, mostly online. A model’s sense of what is polite, what is offensive and what is true came from that stack, and a person in Lucknow or Lagos is using a system whose defaults were set elsewhere. Open weights and local adaptation are the partial answer; whose values "},{"text":"should","i":true},{"text":" it carry has no answer that everyone would sign."}]},{"t":"h2","id":"will-the-benefits-reach-the-people-who-did-not-build-it","runs":[{"text":"Will the benefits reach the people who did not build it?"}]},{"t":"p","runs":[{"text":"The first lesson of this module put India as a buyer at every layer. The classroom you are reading this in exists because a year-old frontier is nearly free. Whether that is enough (whether the value of the technology flows to the countries with the labour and the users, or only to the countries with the chips) is a question about the next decade and about policy, and the evidence so far is that both are happening."}]},{"t":"h2","id":"how-to-hold-an-open-question","runs":[{"text":"How to hold an open question"}]},{"t":"p","runs":[{"text":"Not as a reason to disbelieve everything, and not as a gap to be filled by whoever sounds most sure. For each of the eight, you now know what would count as an answer. When the evidence arrives, you will be able to recognise it, which is more than most of the people writing about these questions can do."}]}],"check":{"question":[{"text":"Why does this lesson say a model’s fluent description of its own inner life is no evidence about whether it has one?"}],"options":[{"text":[{"text":"Because models are trained never to discuss their inner states."}],"correct":false,"tempting":[{"text":"Invents a prohibition; models discuss the topic readily, which is exactly the problem."}]},{"text":[{"text":"Because it is the likeliest continuation of a question about inner life, from the mechanism behind every fluent answer."}],"correct":true,"tempting":null},{"text":[{"text":"Because consciousness has been shown to require a biological brain, which rules the question out for any software system."}],"correct":false,"tempting":[{"text":"Asserts a settled answer where the lesson says there is none; the report cited found nothing in principle preventing it."}]},{"text":[{"text":"Because the model’s description changes every time it is asked."}],"correct":false,"tempting":[{"text":"Names a real property but not the reason; consistency would not make the description evidence either."}]}],"order":[3,0,1,2],"why":null},"prev":{"slug":"what-agi-means","title":"What “AGI” means, to whom","minutes":9},"next":{"slug":"how-to-read-the-next-headline","title":"How to read the next headline","minutes":9}}}