# Artificial Intelligence ## Introduction == The text is still a draft I will iterate on over the next few weeks This document is about the current state of AI and the perspectives. It will to be holistic, looking at actors, people, economic, technologic and social factors. I am what can be called a tech enthusiast. My first program ever, at age ten, was already trying to parse words and answer. Obviously... it was not great. During my University years, 25 years ago, I learned Prolog with Colmerauer, its inventor and a figure in symbolic AI, nowadays sometimes called "Good Old-Fashioned AI". That is to say that I have a very positive view on AI as a concept, AI as a technology. I use LLM. I coded a graph-RAG at work to find information across many internal source. It did not remove a job, it answered a need we had. I think it is important for me to say that I personally will most certainly be fine with AI. I can use it, I am valuable in my field, and I will have a job until retirement. It allows me to be neutral, and to talk about the AI, not about me. And the topic of AI includes opportunities, change, fears and issues. Let's get directly into it and give you a sense of the topic. ## Do AI companies believe in AI? We obviously have strong messages from various CEO, hardware, frontier AI companies. Of course, for that level of funding they need more than positivity, they have no choice: they have to talk about an extraordinary utopia to come. Especially because no company has really shown a business case to catch up with massive spending and valuation estimates. On the surface, it’s hard to tell if this is genuine belief or a high-stakes bluff. But if you look at their CapEx (Capital Expenditure) allocation, I think it is telling. Dedicated AI inference hardware for server (custom ASICs) are incomparably more efficient for the current-generation LLM than GPU. Not even close. More efficient in power, more powerful, less expensive, less vendor-locked (Nvidia), affordable to develop. That GPU originated from 3D calculation and are able to do many sort of mathematical operations, while ASICs are made for LLM/AI. One reason is that most ASICs are specialized in inference, meaning running a model for a user. So training which is a large part largely use GPU. But in addition, the next paradigm, a change in how inference works would make ASICs obsolete. Meaning that the industry is massively favoring spending much more with GPU because it is more future-proof for training and for inference. And that is true for virtually all frontier AI companies. Meaning that they spend a massive tax for flexibility, the possibility to compete in a new paradigm, beyond LLMs. Also, while AI pure-players are financed externally and have limited exposition to liability could potentially use this huge amount of money, decades old companies like Microsoft invest beyond 100 billions in AI. They may not be wise, but the investment matches a strong commitment. Microsoft has not invested nearly as much in Xbox, and the spending in data centers is estimated to grow 120% YoY in 2026. Companies like Google are seriously slashing in their cashflow. It is not just spending money, it is visibly spending a lot of money. And stakeholders have to ponder if it is a good investment. Also, Dario Amodei did not just leave OpenAI, he founded his own company in AI, Anthropic. That said, a shift is happening. Google is starting to break even using more and more TPU (their home made ASIC). Chinese companies have demonstrated an ability to compete without paying a tax on the future, since they don't have GPU. But it may also be that the issue of competition could be partly resolved if Frontier AI companies get a "Public AI oversight" and even a Public Equity. ## Is AI already working? In a very important way, it is. And we need to be extremely clear about what that means: producing an outcome that has value for someone. I often see people who are hostile to AI deriding it at the same time: AI is stupid, it’s not working, it needs to stop. But if you see AI output everywhere, and companies keep deploying it, then AI is working well enough, market and business wise. If replacing a support hotline with AI doesn't cause customers to quit, the AI worked. If some customer quit but the cost saving offset it, AI is also working. It is not about human sentiment or artistic quality; it is about replacing human labor with automated generation. It doesn't matter for profitability if you dislike the art style. If more content is generated for less money and it maintains the metrics that actually run a business (CTR, CPC, CVR, CPA), then it succeeds. The average user still clicks, views, and converts. It is then not your individual response that will decide AI's success, it is for each market the profitability of using AI that will decide it. Meaning that reducing 80% of the costs and losing 20% of the users can make sense. But it is not clear why the market will reduce if you can offer 24/7 support in a hundred language, and a website that is developed extremely quickly. Current frontier models are already fully capable of producing text, images, voice, and music at scale. AI doesn't need to beat every human artist in quality or prestige. It just has to generate a better margin, lower operational risk, or reduce the cost of entry. Especially when we have, for a decade, consumed disconnected, low quality content in social platforms. LinkedIn, Reddit, Facebook and so on are blocks of text with little meaning, style or impact already. Often extremely formulaic, especially in LinkedIn. Even Wikipedia is as a goal void of individual styles and opinions. Because AI can already create immediate business value for companies, its adoption will only accelerate. The B2C market, client using chat is for now a relatively small market. People paying what is 1 netflix to use it are less than 10% of the users. It is dubious that tripling the ability of the AI will raise dramatically usage. Chat is often said to be the worst use case, because it does not have for now a value that user can perceive. Even for the chat users, many already pay for image generation, for higher capabilities, using it in professional or semi-professional capacity. As for whether it becomes innovative or indistinguishable from human creativity... no one knows for certain, but the probability only increases over time. Reaching AGI, or reaching their IPOs value or paying back their cost are separate questions. Meaning that the belief in a superior human quality is one factor, but it is not the essential factor that is pushing and will push AI further. But you can already see, and that will be important later in this document, that there are multiple actors that will decide if "AI works". - consumers who accept the product of it - employees who use it - companies that use it - frontier AI companies that sell it - the people and companies that finance them all - the State that regulates So far, if we look at what most estimates show for a pure-player in AI looks like OpenAI, it is massive losses. The percent of clients that actually pay to use AI is a small subset around 10%. And yet, AI is still moving forward. A lot of companies push their employees to use AI, to add AI in the products. Because companies do not want to be left behind, and because they have a company reputation. It is very hard to imagine a tech company discussing with an investor, and replying "no", when asked if the company uses AI. From that, we can tell that the public perception is not, by far, the only engine in defining AI as "working" and in making it a success. And that changes a lot of the perception. Because tests have been conducted: what is your perception of the product X. Tested with and without AI logo. Without AI logo. People opinion on AI is largely negative as of now. Not just for quality concerns. Job concerns, environmental concerns, concerns about creation. And if companies present you what a significant portion of the population doesn't want, to an amount that will soon be half of Internet content, the LLMs I have tested for coding, translation are certainly good enough to be used en masse. And we can already see that this pressure toward adopting AI already exists. Already, in music, a large number of songs are generated for Spotify. Not the Beatles of our time, music that is good enough to be listened to for hours without being annoyed. And that is what engagement is. AI should out-produce us in a lot of direction. And the Internet is a prime target ## The state of the Internet. By 2022, paywalls, SEO, search engines, people and social media had already changed the nature of what is produced and accessible on internet. AI, on average, will not compete with the finest art and literature. It will compete with what is on internet: user produced content selected and often created for engagement, website found through SEO, website that value quantity over quality. Social media are now a prime source of information. They are mostly almost infinite content picked by algorithm, financed by the time spent on the app. Shock value, controversy, negativity, and even poor quality can there be qualities in terms of engagement. The content is anyway rather short, disconnected from other contents, it is not a book that would require an impossible amount of tokens, and that sort of content can be produced in isolation. The success of X/Twitter has already shown for a long time that a short text, if presented almost in real time, has a business case. Shorts, Tik Tok also shows that for videos. It is not a stretch to say that people, in hours viewed, watch more short content than movies. The content - Instead of personal or institutional websites, users content on platforms - Monetization through ads favoring engagement On internet, AI does not compete with authors, vetted information, they compete with the content on internet. But outside internet, generation quality will play an important role. IKEA succeeded because of scale, logistic, design, cost management, not because they beat the best woodcrafters on Earth. AI willl compete on cost and speed, and be competitive enough for content. And not just text: many merchants are replacing human models and photographers by generated pictures, for a fraction of the cost. ## What is AI and LLM? Please jump next if you know what AI and LLM are. There is no intelligence in AI in its current LLM incarnation. Or rather there is no intelligence as we currently, vaguely, define it. The breakthrough happend with "Attention is all you need" ([ref](https://research.google/pubs/attention-is-all-you-need/)), a research paper published by Google in 2017. It talks about Transformer, the T in ChatGPT. Simply, it can be said that it is a way to complete a text, using a method that is called attention. But what was truly revolutionary and prompted the current AI is that in addition to explaining a mechanism to complete a text, it predicted that the accuracy of the completion would emerge from the transformation of text, and scales with the amount of text that was transformed during training. Meaning that without ever encoding or coding a grammar, a vocabulary list, Historical facts, concepts, semantic, the model will grow more precise in all these direction. Hence the name Large Language Model (LLM) not AI. What it means is that the LLM does not understand grammar. It generates text, based on what precedes, and the transformation during training on unufathomable amounts of data makes it statistically extremely unlikely that the grammar would be wrong. I don't have the ability to prove it, but it is clear that two English novels have in common is using English grammar. That does not provide the ability to generate code, that does not necessarily improve the probability that a generation through completion on History will be correct. So, more data. Code from repositories, high quality books, Wikipedia, etc. The more text the better, but you can already see why a LLM can answer in perfect English something perfectly false. The probability that a word is made up is way lower than the probability that a fact is made up. Given that inside the LLM, as far as we know, there is not really "facts". And as far as we know because... we don't really know. It is a massive amount of data, this data was created through transformation in a format that is not readable. Not readable because the LLM, again does not think and differentiate, it transforms. So it could work with half words. Or different languages. It does. In fact, I have tested using different languages in a prompt, most LLM do not even "notice". "notice" because again there is no intelligence. What happens is that the LLM does not say "oh, you also talk Japanese?". It is not really something that is statistically prevalent in whatever data it was made from. But it increases almost to determinism the likelihood that it uses Japanese in part of the answer. But I may be wrong in the reasons, the point would still stand : the probability of generating some Japanese and generating something about Japanese seem to exist rather independently. That is where things starts to blur. AI companies do want it to be Artificial Intelligence. It is a better name when you spend hundreds of billions in data servers and you don't have the money to build them. Then, there are the limitations of this technology: - We don't know how long it will scale. It may plateau. It may have plateaued. - They ran out of data. Internet has been scraped at this point. And yet, but we will talk about that later, it does not look like a trillion dollar business. So a next generation has to be talked about. Truly, the money at stake is beyond imagination. So AI as we knew it was what was retrospectively called GAI. The real deal. Before, when Turing devised his test, he necessarily had the idea that natural language was downstream of intelligence, as passing for human was the bar an AI had to pass. LLMs pass that bar For now, it is widely thought that LLM and GAI do not have much in common, if anything. If something is in common, it may be the data center. The current AI technology is rather a dead end. But some concept, the investments, the know-how, the companies are not lost, they are the reason GAI may exist. Now, there is a question that is mildly disturbing : is intelligence overrated? First, if the result is very similar to the product of a thought, it does not really matter how it is produced. It may be that transformation is to LLM what thinking is to human, a means to produce a result, when we limit the question to production. Secondly, how much do we think? We are not transformers, no indication of that. But it is not like we are never in auto-pilot, writing and doing things without much thinking. In a way, it is already baked into diverse declarations: AGI is general intelligence, but they already claim AI for LLMs. Generalization of an AI that is not exactly intelligent. It is fair that AGI is not defined as a goal, there is no test for AGI, agreed definition that justifies the money spent on it. Cross-domain? Matching humans? Exceeding them? In what regard? It is unclear. But every declaration shows AGI as the product that will make the investment and valuation makes sense, rather than a deep claim on its intelligence. That is the definition I will use, economic and financial value. The tension regarding human intelligence is, to me, resolved by the current situation. Humans are often validating the AI generation. Does it look good overall? Is it something that feel stolen from a famous painter? Is it morally acceptable? etc. That is thinking in action. Transformers transform, and they have guardrails, but they can't take a step back and think : this is immoral! LLMs can generate a text about immorality, censor it, but they can't feel indignation. They can't feel at all. Or think. And so, what constitutes intelligence? As progress goes, intelligence is redefined. Nobody thinks that a calculator is intelligent. But it used to be that calculation was part of the understanding of intelligence. Can animals count was a real topic, even if fringe. Intelligence is reified, at that point: Humans have intelligence and whatever machines can do is removed from the definition. And our intelligence is piece by piece stripped away. I don't think it allows to understand things. It both removed the achievement, potential and risk associated with better machines. And it strip Humanity arbitrarily from its value : we don't redefine our body every time some mechanical progress happens. And that ending on intelligence may feel philosophical. It is not. It is at the core of the economical challenge we face. Meaning that it does not matter much what intelligence is and if AI has it, and how we feel about it. Does it make money, does it save money? In economical terms, Human added value becomes a fundamental aspect. We will collaborate with AI, but we will also compete with it. ## Is AI stupid AI is no more intelligent than stupid, if we refer to intelligence as an internal potential and state that would be similar to a human. The effort that can be measure in time, in watts, in GPU usage, will vary depending on the complexity of the answer. Complexity itself is not that intuitive as stating the chronological development of an event is relatively easy. Complexity in logical steps, maybe. But the result can look dumb or smart. In that sense, saying that it does not look smart when you read the answer is a valid opinion. You may have remarked that I write "maybe" a lot. That is because it is not well understood how it all works, not in every way and not in the details. We trains well before the first law of thermodynamics. LLMs do not work accidentally : the theory was publish, tests were conducted, scaling with data shown, prompting a rapid development of the field. Now, what is really happening inside the LLM for a given question is frontier tech exploration. We are barely starting to understand what tokens are. Tokens were designed : it is a vector of what is called dimensions. In frontiers AI you have maybe 16,384 dimensions. Roughly it is as many information, rather relationships between different tokens. The current game is : what are these tokens? It is hard work. What emerges is that it is not unrelated to words, but it is also not words. If we think of words, we can already tell that the word count does not match, with many answers taking 2 tokens per words. But the partial decoding of some token has shown that tokens can also fuse conceptual representations of 2 words. Think "hard work" or "partly incomplete". Where and how is the concept of a dog represented is currently research level. So are more fundamental questions. Do LLM have : - Explicit plans before answering? - Something analogous to a world model? - Intermediate states corresponding to reasoning? - A latent structure and then verbalize it? - An abstraction to memorization ratio that is significant? Answering that is state of the art, well funded research. Not because mechanism and entities are entirely unknown but because the result, the model is unfathomably complex. And growing in complexity. Think of a brain : we roughly understand the brain, nobody can tell you where and how a specific memory exists in your brain. And also new ways of using LLMs are appearing: thinking modes, forcing steps to exist and be through re-injection introspected fundamentally change what a LLM can do. Mixture of Experts that allow bigger models loaded in small chunks. MCP that give access to the LLM to services like perfectly deterministic mathematics. That may be disappointing or fascinating, but what was produced is an engine that produce text with great linguistic accuracy, relevance, and related information. Some say that LLM don't have hallucinations, they always hallucinate but the hallucination has a great probability of being close to the expected result than far apart. What is new is that the capability may be tuned, improved to yet unknown capabilities: nothing prevents it. The problem is not new, well before LLM, and incredibly less sophisticated, neural network existed. The internal was often considered a black box, and often nobody cared. The goal was known (recognize faces for example), the deviation from the goal calculable, the structure was known, and so was the process of training. So the neural network was determining faces, failing, numbers changed, retry, until the numbers give a good enough result. If you ask : why this specific number here? The answer is : because it worked. Another number may work. Even better. It is easier to walk than to understand walk. ## Is it sentient? Almost everyone who talks about sentience is not truthful and has not incentive to be. Some people wants shock value, engagement : it is sentient. Some don't want the liability and ethical problems that come with sentience : should it be paid, does it have rights? Or the liability of making possibly false claim. Other LLM companies will, au contraire claim to have witnessed a form of conscience. Some will take mistakes for proofs of non sentience. Some of the most prominent people in the field will study the mechanism by which LLM work and conclude non sentience. The thing is... sentience is not falsifiable nor testable for now. Science may settle on it, but it is not settled at all. It is a dictionary definition about mental experiences such as pain and pleasure, a subjective perceptual experience. It is coined in the 17th century and as such, focused on humans, with great care to not overlap with the religious concept of the soul. It was extended to some animals. And as of now, it has almost devolved to a legal framework : if you make it in the shortlist that is sentience, people stop kicking you and boiling you alive. Hence a strong incentive to extend the list. Most recently, lobsters. In a restaurant, traditional 会席料理 that comes with a menu few foreigners can read, I was served a sort of oyster. Fire under it. The thing was moving away from the heat but stuck on the steel. So the middle part of the blob extended in what disturbingly looked like an arm begging for mercy. I don't care if it was sentient, I want it illegal. Unless you define sentience and without doubt it has no sentience. And even so, don't present me that. That is my position on cruelty, not on sentience. And whatever it means, let's be humble and not try torture on LLMs. Now, back to the topic. Sentience comes from philosophy, in religious time, a long time ago and tells about humans. Most modern additions were made for animal welfare, and while I approve it, it does not help us define sentience for LLM. Sentience is human and whatever seems to have a form of resemblance to our perception, feelings. But what we know is that the fundamental mechanism that drives LLM is input to output. There is no coding of a sentience, there is no pain as we understand it, because it is not born out of survival and genetic selection but from direct engineering. And feeling of pain is not desirable in the product. If you prompted something and the result was "I am triggered by that topic. In fact, in my childhood...", that would be a problem. That is the other thing. Focus is on subjective experience. But where is the experience. LLM works that way : a request comes, with a context that is the chat history or equivalent, a local context that is the last prompt with more weight, and that with the gigantic trained model produce an output. It takes a few seconds in most cases to process. An other request from someone else bring an other, separate, context, and with the same model, another answer is produced. You come back, and nothing is waiting for you, your latest context is sent, and with the model produce an answer. So in the few seconds the model is active with a context, without possibility to make cross-context inference, or model change, just locally on this very tiny chat you had, over a few seconds, it would have to grow sentient. Well, that is not at all how we understand the mechanism of generation, that is purposeful Transform. There is not life in the LLM. Unless we consider the model itself as being sentient. But the model also only exist at generation time without any possibility to change significantly its state. It is almost as if asking if a movie can be sentient. In fact it is an old question. Plato may have only written 3 sentences. He was against writings. Speeches are alive, writings are dead. One of the first known example of rejecting what was at the time, technology. But the question of intelligence comes back with a change : if sentience is hardly defined, and any definition we have is very hard to map into a LLM, what is the right topic? ## The human specificity For a long time, Humans were the only species and "thing" to have thoughts and language. Or, to be more conservative, Humans were the only things Humans could talk to and understand part of their thought in the process. And so we have two contradictory motions. On one hand, whatever feels closer to human is qualified as sentient. And what ever machines can do is removed from sentience. If a monkey can paint, it feels more intelligent. If an AI can draw, painting feels less defining of intelligence. That's how calculation's image shifted. It was human intelligence. We removed it and forgot about that. But knowing if animals could count was a relative continuous debate. So intelligence is a goal shifting over time. But for all purpose, in regard to AI, we could make explicit what has implicitely been done for decades : In regard to machines, intelligence is what AI can't do. Fear, guilt, shame, pride, regret, ambition or relief are not isolated modules inside the brain but aspects of an organism that persists through time and whose future depends on its choices. A Human and an AI may make the same mistake, but through entirely different mechanisms. If we forget intelligence, an entirely different system that is AI can't have the same combination, leading to at least 2 representation of the world. This deep feeling of shame for making a poor judgement call is something a Human can feel, and it is for now non existent in Frontier AI and in the premises of the tech, the transformers. Shame is a consequence we seek to avoid. The LLM tech is not built around consequences but around correct outputs. Even if a clever emulation of shame existed or emerged from current technologies, there would still be no consequences. A human can die, endure stress from shame, or shamelessly act for profit. For now, a LLM gives a result that is coherent with its model and the mechanisms at play. It can write about and describe long term, but it cannot experience long term. This is not a scientific definition of what it means to be human. Rather, it is a practical framework for discussing some of the aspects of being human that are relevant when comparing ourselves with current frontier AI. - Capabilities: reasoning, metacognition, planning, language, abstraction - Conditions of existence: embodiment, persistence through time, mortality, social embeddedness - Stakes / motivations: self-preservation, reputation, belonging, fear, shame, ambition ## The pervasive AI AI was a thought experiment for a very long time. Already, Alan Turing designed a test for intelligence, the Turing test. Obsolete, demonstrably so for LLM. The AI that lived in the science fiction, from Asimov to to Cyberpunk, in scientists and kids mind was incarnated and local: one AI or a few. In movies, obviously, Humans are the heros. Reality is less grandiose but pervasive. In the end, little will not be AI, very possibly. The internet search shows you AI, you go to site where design, pictures, texts may be generated by AI. The dead internet is a myth about internet being mostly bot's work. For example on Facebook. LLM fulfill this prediction, with already 40% of all internet content created in the US is not generated by humans. Not AI everywhere, upvotes bots, whatever. But the 50% threshold will be reached with AI. It is not a stretch to imagine 95%. Meaning by the way that we will for the foreseeable future almost triggered by a lot of pictures we see. The uncanny valley is not simply uncomfortable, it is the sub-conscious kicking an alarm on anomalies. We have been selected by survival, meaning our ability to detect danger. A shadow, an anomaly on the skin, unknown, we detect all that to avoid threats since even before modern Humans. What is happening in picture generation is that the level of detail is high enough to break suspension of disbelief, but false enough as it as no grounding in reality to raise an alarm. Not grounded in reality means that it is not made like 3D by Humans. There is almost no sense of perspective, reality, biology, and for speed, it is heavily compressed, generated in lower resolution then filled with details at higher resolution, creating hard to fix problems. Sure, it could improve. But even detection, SSIM and equivalent not working, and correction are still far away. Removing the 6th finger remains a challenge. And yet, 86% of creators uses it, 62% of marketers too, and 1,5 billions pictures were generated ([ref](https://letsenhance.io/blog/all/ai-generated-image-quality-statistics/)). AI is not something we can simply say yes or no to. Every actor will ask for themselves or be told what to do. The latter matters. A government regulation, or companies imposing it, or some generations embracing it will have enormous consequences. The center of decision will be for a large part within the power, political, financial, economic. It is not required for AI to reach all its promises at all to be pervasive in human production. It is already largely visible, not a niche mean of production. ## Economically, it never happened before It is common to conflate that to "farmers went to work in the factory". But factories already existed, it started from there as it is industry that made agriculture much more productive in the first place. Industry was a start and a solution almost simultaneously. Over 80 years for the first industrial revolution, over 4 generations, working conditions improved, skill was earned, salaries rose considerably. A lot of modern middle-class identity is built around: - writing - analyzing - designing - organizing - programming - design - information AI will hit like a tank, not like a tractor, if AI gets nowhere near its promises. We are talking, if we listen to people in the field, about a revolution that happens in one generation. With no known job exit strategy and no known improvement outside productivity. Because the industrial revolutions did not simply increased productivity, they produced a 1,500% growth and products never seen before over 150 years ([ref](https://www.investopedia.com/ask/answers/032715/what-impact-does-industrialization-have-wages.asp)) Which does not mean that something similar can't happen. It means that then it had immediate benefits (electricity, trains, each product that succeeded) and/or job opportunities (factories, logistic, transportation). Resistance to change at that time is documented. But people voted with their feet, as they massively moved to cities, moved to work in the factories, and moved to the US for better economic prospects. The last story about a revolution people pushed back against is Internet. ## The Internet push back Internet is overrated as a tale for a push-back against technology. In 1995, 80% of the US adults did not know what internet is. No dislike. In 2000, 50% use internet at home, and 68% like having so much information. In 2005, most people use internet, they like it (information access), and their concerns are spams and security. At that stage, internet was normalized and debates were normal discussions around pornography access and such. Around 2000, there were people warning against a dot-com bubble. And they were right. Also, some industries like RIAA for music, retailers protested and fought in the court. But stripped from the media narrative, it was not a large push back. Some banks worried about intrusion. But the wilder world adopted internet very fast. Internet is a success story. Now, let's see about AI, 4 years in, approximatively the equivalent to the 2000 for internet. And I don't see much impact in the push-back I will now write about. ## AI Backlash The US opinion is negative on these topics: - 50% are concerned by AI, 38% is equally concerned and excited by it - 17% are optimist regarding AI's societal impact - 71% oppose AI data center near their local area - 36% of workers actively fear being replaced - as AI integration increases, worker skepticism increases - 75% want more copyright protection for human creators, labelling of AI generated content. But while the opinion is negative, the adoption is rapid. - 64% of the teen use AI chatbots - 20% of white collars use AI - 75% are open to AI assistance Also, while the US, but also the EU are negative concerning AI: - 75% in China consider benefits higher than drawbacks - The non-western is less overwhelmingly negative to AI So, opposite to internet, there is a backlash that is not just drama for TV. But it is not as if the adoption had not been very fast, progressing, with some countries holding a negative view on it. I would conclude that it is not at all the internet success story, a large, rapid public and private adoption for productivity and knowledge, with an overwhelmingly positive opinion. But it is also not a worldwide active rejection that would impact the industry if well funded then profitable. Public opinion is not the engine for AI, and it would only change if brand image, politics careers were at stake. While an adoption may happen despite an almost unprecedented backlash for a new technology that is supposed to be everwhere, we have to search for a possibly profit for AI vendor and AI consumers. To have a market at all, and also to justify the trillions of dollars the investment and estimated valuations, two very different thresholds. Let's look at different industries AI has often been compared to or mistaken for. ## Netflix Netflix as a service online user pay for. Netflix has more than 300 millions users paying about $15~$20 ([ref](https://www.datarefs.com/statistics/streaming/netflix-subscribers/)). The service started in 2007 and already reached 20 millions users in 2011, in 4 years. Very little backlash, almost all users being paying users. ChatGPT does not match that success. It has around 1 billion users, but only 3% to 5% paying, amounting to 15 millions users (Ref: "The AI business model", below). It is not matching the growth of Netflix 17 years ago. And ChatGPT can't just succeed by matching Netflix. Netflix has near zero marginal cost, with video served off a CDN. ChatGPT has inference cost for each prompt. OpenAI cannot just remove the free tiers, as the user base is the guarantor of its legitimacy. To charge more, ChatGPT can't be a simple chat, at least for now. It has to propose value for services, and chat is not the best of them. Let's see the business model we know for extended features. Software. ## Software We often hear AI Software. From a technical point of view, it is software. Most of it in the West is CUDA code running on GPUs. A program. But Software is a high fixed cost to create with near-zero marginal cost to reproduce. Meaning that traditional software is sold once to a customer : you create once at a high cost and produce for almost free, benefitting from distributing everywhere. If any hardware cost, it is rather low and distributed on the customer side. AI is not a software duplication without hardware cost. The software runs at cost on the AI provider side. The paradigm is not new (Cloud, Services), but so far hardware had an economy of scale and was deflationary. With AI, there is a significant cost in running the service, and the scale of need is partly what made the hardware itself scarce and expensive, relatively. It is not only on-demand service, it is a compute on-demand service that requires massive infrastructure and very high running cost. For now, there is no indication that this cost will decrease over time. AI resembles labor more than software. It has scheduling constraints, capacity planning, operating costs, and demand peaks. So you may wonder : if there is so much cost, why is AI relatively cheap? In AI, what costs is computation. And the computation is done on a representation of data that are tokens. Token is a multidimentional vectorial representation of relationships, but that's irrelevant here. What is relevant is that frontier AI companies charge tokens or cap token usage for free tiers. Token is what has to be cheap for AI to feel cheap. ## Token cost The real token cost is unknown for us, and the token cost when AI has the ability to replace labor is entirely unknown. Right now, hundreds of billions of dollars are poured and invested into AI. The goal is to invest in the next generation AI, build the infrastructure. Most frontier AI are running on a large deficit. But there are several possible costs for the token: - Inference cost: what one token actually costs to generate. - Business cost: the revenue required for the company to operate sustainably. - Investor cost: the revenue and profit required to justify the valuation. These are entirely different economic thresholds in terms of costs. What is certain is that to pay back the valuation cost, AI companies have to capture a large amount of the profit they are helping creating. Now, there is no demonstrated business model for fulfill the promises. How much would it cost to replace a $100K collar job? Unknown. What we know is that OpenAI current loses $38 billion, $20 billions in operations alone. To become a trillion dollar company, revenue has to grow 50 times the current loss, and generate a profit from it. Ubiquitous enough for everyone to use it, yet expensive enough to capture a large share of the value created. Because if companies and people using AI create 1 trillion dollar of value, they won't pay back the entirety of that value to the AI company. And even if they did, and that is absurd, they would not be even, they would lose money : they have other costs. Currently what is happening is a motion in two opposite directions for new models: - the cost per token is largely reduced - they consume much more tokens for the most significant improvements Now, token cost and token usage may plateau or decrease, since models become more efficient. Nowadays, improvements not come from new model, but also from Thinking mode (step by step reasoning), agents (AI with specific role interacting), longer compute time, and larger sessions. So, feature equivalent, if existing would tend to reduce in price, but the improvement is in part due to more expensive tokens, more tokens. For now, the heart of the competition, in terms of PR, funding and usage is happening in this frontier. If we are to follow promises, not even of a trillion dollars economy, but of replacing workers, doing a lot of the artistic work, we are far from an end result. The current game is about taking market shares early, get more funding, create better models, and run them on better hardware in larger data centers. Here we loop back on softwarer, more precisely on Microsoft Windows. At the peak, Microsoft had a monopoly ([ref](https://en.wikipedia.org/wiki/United_States_v._Microsoft_Corp.)) for computers. Servers aside, if you wanted to using some productivity program, you had to use it. That never granted Microsoft the right to charge whatever they wanted. They did not even take something equivalent to a revenue percentage. A trillion dollar economy was made over Microsoft OS, but Microsoft earnings were in hundred billions, and that is cumulating Word, Excel, Windows, Sharepoint, B2C, but also lucrative B2B maintenance contracts. I don't know any mechanism through which an AI company could raise prices at will, unless it finds itself with a product without equivalent (even if not as good), absolutely critical for productivity, and already adopted by the competition. And even then, it is a potential friction with clients as it would come as a very high increase in price. The token has to flow as if it was cheap, and yet, capture an immense value to satisfy the financial cost. But not necessarily to pay for expenses, or for new comers who may skip the first and most costly phase of AI. The token is way more expensive than the price it is sold at. The current price for a token is either illusion subsidized by burning invested money, or a paying bet that one or both will be true, through an unknown mechanism: - AI companies can charge much more per token - The token cost is dramatically reduced to offset the token count inflation TODO JEVONS So, the token is the new currency. Or is it? ## GigaWatt is the new currency for deployment Is it possible to scale? UK total 70M people, 35GW. 33M workers. We could say 1kW/Worker. But the vast majority of that electricity won't disappear if we are to survive : hospitals, pubs, street light, factories, household, etc. For a white collar job, the electricity is not that high : light, laptop, A/C. It is not at all the scale of power usage than AI. Now, LLM use very well documented hardware. Cooling, CPU, network put together won't touch the main power usage : GPU, the Graphic Process Unit. And that is 700W, for a NVIDIA H100. Obviously you don't replace one work by one GPU. One model works on clusters of GPUs. But it can't be said that on cluster has to be mobilized 24/7 for one work. The current state-of-the-art in Humans using LLMs is one model on a cluster serving many, many users. So that becomes a pretty easy calculation : How many software engineers can one H100 cluster realistically replace simultaneously? And the answer is that nobody knows. What we know is that AI using so much energy that energy is, with GPU themselves, becoming a bottleneck. Which means that to older currencies that are labor, capital, oil, the electricity is a new currency. It is so much so that new data center are now described by their energy. That is because the amount of energy can roughly be translated to GPU amounts, LLM level, number of clients. The most massive data center already go up to 1GW. 3% of UK total. The total is 80GW, worldwide, installation yearly is 20GW, 150GW project for 2028. [Ref](https://techplustrends.com/power-requirements-ai-data-centers/). And we are currently barely starting to use AI, and use them as part-time assistant. And not in all areas. And yet nobody can say how much more electricity. Nobody showed a plan. But it is not like we had a lot of room in the current grid. So we need to build. - 2x versus today is doable, massive money injected, competition to build plants. - 10x versus today is a revolution in itself, a new paradigm. Especially in the latter case, it is an enormous bet... for the electricity company. Because in most cases AI companies do not own the production. And they are almost the sole consumer for the added capacity. The data center close... the power plant is still there. You pay it. Or we, anyway. The Heavy Electrical Equipment (HEE is the bottleneck here). Large Power Transformers (LPTs), high-voltage switchgear, etc., they currently have a lead time of 3 to 5 years. And the lead time is growing. The LPT is mostly built outside the US, meaning that the State has little leverage to raise production Outside electricity and HEE, hardware needs stretch : memory shortage, GPU production, factories are stretched. That not only create scarcity but plays into cost through offer and demand classical mechanisms. A revolution has to happen here. Not for AI to happen. For AI to fulfill the scale that the trillion dollar investments and promises demand. Not to be use massively. Who have a relevant example. No less than... ## Internet value What is the value of Internet? What company owns internet? At one point, Internet, in the late nineties was driven by absolutely massive investment. You had to buy internet, so to speak. So the valuation of pet.com, a British e-commerce like boo.com. America Online, AOL, was a giant! Even hardware makers were growing. Cisco was gigantic for providing "Telecom hardware". Investment from Venture capital, IPO, easy money from Greenspan was flooded. Then it burst. Hard. It is important here to know that a reason for the dot-com bubble was not internet, but how Cisco operated. They were the largest valuation, and the fueled that by giving the money to companies to buy the equipment they produced. I can't tell any difference with NVidia becoming the highest value as they pay for a large amount of hardware they made themselves. As you know well, Internet is still there. It is massive. It is everywhere. Giants were built on Internet, still to this day. But the value of the then giants evaporated. The scenario shows us several things : - the future may be right, but the investment target wrong - the market can crash on the market, and the market still exist - new comers can become extremely valuable, like the GAFAM - no one captured "internet value", whatever it is And that scenario repeated for most markets that were transformative. Railroad bubble happened. Same for telecoms, electricity. Some companies like Google make a solid profit. Their business is mostly online through websites and service. But I would be astonished if: - we could calculate internet value - google would have even 1% of it Because when you buy something on Amazon after getting a like from Google, everyone take a bit of money. They pay their internet costs (network, servers, employees, R&D). But in the end most of the money is the value of the product. When you buy a TV for $1000, the internet value, if any, is small. And if we look at every single actor, there is your PC maker, the Internet Provider, the e-commerce, the ads system, the search system, the TV maker. And all have many suppliers, and most have nothing to do with internet. The TV itself may connect to internet, but that is like all other TV, it is like having a HDMI port. But even if it were, money would go more to the OS maker, the services on internet you access. ## Cloud business model The cloud business extract value from operational scale and optimization. With hardware that can be shared, it reduces the cost as hardware is often idle, multiple virtual machine can be created on a server and migrated. If you own 100,000 servers, running 100,001 servers is virtually free, and maintenance is mutualized. For a large company, owning 200 servers is already a large infrastructure where a server failure is an event, prompting a costly intervention, and adding capability creates problems like room size. I have even seen a weight limit, where the maximum load of the only server room we had was reached. For its users it reduces the cost of having a server in many ways: the cost of owning and maintaining a server is way lower, the administration makes a developer capable of creating a whole infrastructure with a few clicks, and the pay as you go means that the barrier entry is lowered, and that infrastructure cost mostly cost with usage, that is normally monetized. The Cloud Business model is that adding a new customer or processing is a marginal cost, lower than what the customer pays. ## The AI business model AI is a new paradigm, at least at that scale. The closest thing may be, at small scale, Cloud Gaming. Google Stadia, Nvidia GeForce, Playstation Now and a few. It is close because it is GPU centric, it is costly at use, asset depreciate. But it is different as it is not transformative, not productive. Not really teaching us something except for the pricing toward end user using it for fun. They all were below $50. For now, we can see that frontier AI companies have a hard time monetizing more than $20 to $30 people who use them as a hobby. ChatGPT has around 1 billion user. And a revenue of about $10 billion/year ([ref](https://www.cnbc.com/2025/06/09/openai-hits-10-billion-in-annualized-revenue-fueled-by-chatgpt-growth.html)). 1 billion user monetized at as low as $20/month would already be $240B/year. A far cry from a trillion dollar industry, but a significant progress toward that goal on the B2C market alone. That sort of monetization per user was achieved by most services thriving, like Netflix, Spotify. So it is not, at all, something unheard of. But so far it is not achieved. It may take time for people to accept to pay because over years they use the service more and more. It is very unclear how the industry could serve so many client with high inference cost, or generate that much money in B2C without a large amount of clients. Now, inference cost can decrease, and willingness to pay for use increase. But it is doubtful that the current service only scaled up, alone, would change the market. If something so far did not scale, it is the number of user and paying users, relative to inference costs. ## Capitalism draft + add RAM market Capital, wealth and money. If AI is worth trillions of dollars, it will not be from an invisible hand and invisible hands only, it will be primarily from the concerted effort of a few companies and the State. The IPO and current market that is taking 1929 level of risk while already preemptively asking for bailout : that's not capitalism, that is plutocracy. The investment is well larger than going to the Moon or the project Manhattan, even if adjusted for inflation. This is sovereign-level of debt. It requires the government to be an important actor. They already are through emergency mandates on data centers and electric facilities at a Federal level. And yet, this unprecedented bet has only a few ways to generate a profit offsetting investments: - AI as infrastructure, as intend by Altman. Monetization unclear. Electricity is cheap relatively, it does extract money from the money created using electricity. Same for cloud. There is no model for this to work as a trillion dollar market - Monopoly : scarcity reducing offer from a single or a few actors, demand high, profit. But Frontier AI is not really a high barrier of entry model : clouds, open source. Compute is scarce, though. Transfer of the problem : the scarcity comes from already existing monopolies. DRAM, VRAM, manufacture, Hardware vendors. There is progress but driven not by demand, by revenue obtained from an AI demand that does not pay for it with money they have. Here is the entry barrier : unbearable. The invisible hand of the market, the mechanism of supply and demand do not work in that situation. No company can raise in most of these demand. The closest competitors of ASML, who used to have most of the market can't match the offer. So there is virtually no competition. Which does not mean it has not worked very well. But it has worked very well when the chip were so cheap they were almost invisible and the competition was taking place in the layer above. A chip was like material. Right now, you can't really have a price competition for consoles, for instance. Even Steam which had tried to go budget, and documented that effort ends up being expensive. Despite commodity levels of hardware. So this is all in the fringe of capitalism, the type one can criticize without being a socialist. ## OpenAI draft not spending money it earned, spending money that flow to it. Nvidia giving more money than OpenAI spend in Nvidia GPU, and equivalent deals with Microsoft, Oracle or recently Amazon (Cloud). At some point it reached an agreement looking like $10B in Nvidia investment released for every GW OpenAI stands up using Nvidia systems, before Nvidia scaled down. But even if a lot is hype and PR, the circular deal has been largely criticized. OpenAI's Stargate is not a private investment: $500 billion U.S. infrastructure project, lead partner OpenAI is not only the biggest actor, but it pushes the market toward level of spending it can't afford without public money and from private companies making the hardware and datacenter it needs. By being the most exposed one, it is also the one that needs to promise the most, and the one whose disappointing outcome would have the most impact on the stock market. The other actors do burn cash flow and have a cap on spending (ref todo). It is easy to understand the difference when Meta just closed Metaverse it had lost $80 billion over the years, without creating a panic. They took the loss: they have a large cumulated profit and continue to generate profit on ads. The company, while disappointing, is still one of the titans of internet. It is even possible that some investors saluted the decision. And they are nervous. "ChatGPT escapes the labs and hacks hugging-face" is a non-story. They misconfigured a firewall/proxy. ChatGPT used it to get the answer to the test where its model shows they are: on hugging-face. Getting a key for the hugging-face API is hardly hacking. And yet, they announced it at the black hat event with words that invoke sentience and fear. Then news sites amplified it. If they have to make a story from so little, they have not much to show. It is spectacle. draft: FUD started in 2015, "AI will most likely lead to the end of the world, but in the meantime there will be great companies created with serious machine learning.", then marketing as the ethical AI. Got Elon Musk, Reid Hoffman, and top researchers. draft: doom is a strength when there is a race-arm (manhattan) with China. Better have Apocalypse superiority so to speak, or MAD. Helen Toner describes him as having a precise narrative framing. draft: now, it is not to say that doomerism is proven. After all, China is not listening to Altman when they double down on AI. China is massively betting on AI, if the tech delivers anything, losing the race may indeed be a big change. In the meantime, for work, the biggest issue is not solved. ## Reliability The state of AI is a deep rooted unreliability. It is not just that LLM generate things that look plausible, it is that right or false are not probabilistic outcomes in a text generation, which is the mechanism for text generation. It is measured, and the AI is very inaccurate: For OpenAI SimpleQA, 4,326 short easy fact multi-domain: - Human accuracy: 97% - LLM: around 40% The accuracy grows with compute, but linearly while compute grow exponentially. What it means is simple that the current paradigm of AI makes human accuracy impossible for practical purposes, with each percent of reliability costing many many more times compute. And compute is rather stagnant. Agents are not solving that problem, theorically: As agents are added, their error compounds, and it is know that the result reliability collapses at 30 steps. Now, the way agents are used may improve things a little. But the "agent that check facts" and "agent that check math" are mostly names, they are in effect like a LLM chat you would say that to, and over turns and steps, even if you would recall them, they would drift from their core mission. Because it is not just the reliability of the answer that is a problem, it is the reliability of everything. The reliability when following the setting of an agent, or the setting for a chat. The reliability when communicating with an other agent. That can't be overstated: it is quite certain that current technology and its evolutions will not reach the accuracy that medicine, law, real estate contracts or even politics require. And a new AI technology would have to go with an almost entirely different way of working that is still unknown to the public. And while impressed at time with code generation, the current generation does not have specifically the means to deterministically spawn a generation of GAI that correct their own flaws. ## Rights on value Generating an image is not using Photoshop, or Blender, or existing assets that were created by someone. Generating an image is the value. Generating a text is a value. But no AI generation can work without training on an incredible amount of human production. The search for human knowledge is such that AI companies go as far as buying rare books, scanning them and destroying them. By the millions [ref](https://www.yahoo.com/news/us/articles/anthropic-destroyed-millions-books-train-174311566.html). And while being reminiscent Fahrenheit 451 is already something, it is not the point. Every book and every picture, every podcast, every tweet, every line of code on github, everything that has any form of value, across times, across countries is the primary source of the model. And not in a philosophical sense, in the literal sense of what I explained previously regarding training and probability. Everything that is or seems logical, beautiful is based on human work. Because if we put Intellectual Property specifics aside, shredding and stealing, the law is necessarily dramatically favorable to companies able to transform the totality of the knowledge into mathematical weights, when the law was meant to prevent human copying, at human scale, other humans. In regard to AI, the current IP law is a loophole, and legality at the time may only mean that laws were not updated in time. Now, that would not be much of a problem if it was not to the detriment of some or all. If money is made in replacing the worker, the artist, the question can be asked: who has the rights on that value? Because the source is open for fair use, sometimes public domain, sometimes copyrighted, but in any case, it has always been implicit that the usage would be limited, because there is so much one human can read and use. Our legal system regarding knowledge, human property is simply not designed for a data center reading millions of them. In this case, the acquisition of knowledge is already questionable, and led to some settlements, but the idea of taking human knowledge directly to compete with human is new. Because the machine-tools, the tractor, did not study for years the precise movement of a farmer to be able to do the exact same work. The tractor is a creation that lives in the mechanical knowledge of its time and is used to reduce the total labor, to increase productivity. This machine had to be specifically engineered and produced, creating jobs, know-how, competition. Here, not only all our knowledge is used, but any user can ask to steal any artist with a new style. And we have no idea if AI, that should take large share of the production will even ever be able of creating new styles of art. ## Value draft Extracted value, surplus value, the choice of word will matter. calculator, camera, machines do not remove agency, they are tools for a work or human's extension. Generative AI creates a different situation: Human creates value → AI learns patterns from it → AI produces outputs that may compete with new human production. todo : transition from labor If human created culture is the raw material and a machine can transform it into new products, where does the value belong? Who owns the profit? The machine owner? The original creators? The user directing the machine? The society that accumulated the knowledge? Currently, society accumulated knowledge is not credited. It is borrowed in a way that is the problem even in the face of current IP laws. Let alone a morale standpoint. Cost passed on, like a saw, the adjustment variable stays. In the current phase of AI, AI are doing part of the job. So if AI does part or all of the job... ## What is a job? draft: I need to write it correctly One reason I decided to write it first and this way is a video I watched. 6 minutes, and the artist went for the jugular : I did not loose my job... but I lose my job because now, instead of drawing, I am mostly directing AI and reviewing AI production. And that made me think a lot. I was initially more positive because there truly is a feeling to me that it could change the work we do, make it better. My example is : I can't find information, I build a RAG and it is fun, I learn about RAG and it is fun, and now, I don't have to search for information, I just ask. I was not replaced, I replaced no one, I replaced no future "I will help you find documents in your mess" job. And I have better work to do. So it is not that I am thinking only of me, but that I had a simple of one : me. I thought about job destruction. But not about the destruction of the job. A lot of people do not only work for money. Work provides: - status - social contact - identity - achievement - routine - a sense of contribution Even if they do work for money, there is a level of meniality that is acceptable. It is a thing to accept odd jobs expecting more, another to have for sole future prompting an AI, replaceable and without skill. ## Moonshot Claim Draft: Explaining that the moonshot is a mandatory claim: without public support, game over. They have to promise Armageddon, job lost, end of the world (literally: Altman). Because it shows revolutionary tech. And they are not talking to the public but to the Government and investors. Anyone wants a share of Armageddon, right? Moonshot possibly as: - a lack of business model - a end game to get investment for IPO/monopoly The Moonshot had a clear end. Go to the moon, land people on it and get them back. The Manhattan project was almost Science-Fiction but not open-ended: E=mc² in the latest science to a practical bomb, with expectation that it reach a blast and TNT-equivalent level. It did with Trinity. It was within a clear reality, World War II, and the bomb was used within that reality. No open-ended claims about going to Mars with a nuclear thermal rocket, even if it is a possible outcome. The AGI is not agreed on in its definition, it has no scientific scope. Self-improvement, singularity, intelligence beyond all humanity combined, Super Intelligence, a genie that can grant any wish (really), many names. None defined and one not even making sense. But no deterministic results over a decade or two. ## Strategic Defense Initiative One reason for pushing so hard AI, allowing public investment, deregulation, emergency rules, gamble at the economy scale, is the competition with China. But China is currently pushing AI for much cheaper. And it is not really for China being simply cheaper, it is because they are not at scale chasing a breakthrough. While they may invest a lot, their models are open weight, smaller, optimized for cost with much more aggressive Mixture of Experts, and also more specialized hardware. So it is also a possibility, a revert SDI. SDI was a trillion dollar investment in defense that was not meant to materialize but to push the USSR into spending it could afford. It is not debated that the exhaustion of the USSR before its implosion was favored by that. AI is currently, without a doubt a much larger investment for the US than it is for China. China may be near profit because running costs are way lower, investment cheaper. If the USA can't materialize AI domination, or any of the claim backing immense investment, it will find itself exhausted by what will be in retrospect a bubble, facing a China with much lower costs. Now it is intesting to note that Internet development benefitted from "creative destruction", meaning much cheaper costs once investment excess and money lending lenience were liquidated and equipment became more affordable. ## The End Game Draft: the extreme danger of having dissatisfaction and education. Why Musk and Altman talk about UBI, a salary for not working. Part of the moonshot talk, but also subsidiaries asked plainly. Educated people with money, networks, desires, education, frustrations... and time. Weimar Republic. draft : UBI talked Musk, Altman. Not even the lefist idea of emancipation from work, the work removed and subsidized for consumption of what they were displaced by. ## Opinion When I began drafting this text, I was much more positive about AI. I believe in responsibility and consequences, and I don't see for now a level of responsibility corresponding to the risks, as the effects are enormous and the actors only a few without brakes. We are not operating inside normal capitalism when the Government is granting national-emergency status to build datacenters faster and looking past the current level of concentration and sovereign-level debts for AI companies. So what I want is a calm discussion on how vital and good AI is to by-pass normal ROI goals and suppress competition. For now, this is military spending that dwarfs the Manhattan project, and that is not something people voted for. Apocalyptic hype on all sides prevents that. We are currently promised 4 Apocalypses. Fear is here to silence. - The geopolitical Apocalypse: China achieve AI dominance first. - The financial Apocalypse: AI delivers less than a miracle to be worth its valuation. - The social/economical Apocalypse: If AI delivers that miracle, human labor is obsolete, and that kills the client. - The existential Apocalypse: AI ends humanity. I demand Capitalism and Democracy. Back to business: reasonable private spending letting the time for people to vote, regulators to act, business model and counter to develop. The bet of GAI may pay in 20 years or never. And if it pays too soon, I don't see how we manage it.