Transcription
So we have CEOs saying AI will replace all white collar jobs within 18 months. But you can't have work that has to be done and amazing abundance for all. If it's work that has to be done and and only some people can do it, then you can't have abundance. Then it's narrow. It's narrow. Exactly. We need to be very careful with AI. We need to be very careful with robotics. We don't want to find ourselves in a James Cameron movie.
I'm an opto electronics engineer and I work with precision hardware every day and I work with apply physics and I'm here to prove that that reality does not survive physics. It does not survive engineering and their displacement models do not survive basic economics and I'm going to do it with real math. So stay with me and let's get into this.
So I keep seeing the same narrative everywhere online. The narrative that AI will replace all white collar jobs. Some say within a year, some say by 2030, but the narrative still stands. And I'm not here to downplay AI. AI is obviously very powerful and it will transform the way that we work on a day-to-day basis. But I'm here to argue against the economics and the engineering behind the claim that AI will replace all of white collar jobs within 18 months or at all. And I'm going to test their timeline against real mathematical constraints and real economic constraints because the people who are pushing these narratives are not doing that. They're watching demos. They're looking on social media and they're looking at their bank accounts. And because of these things, they're pushing doomer claims saying AI will replace all positions and vocations, especially for people who are starting early in their careers.
But the reality is that there's hardware limits, there's memory limits, there's power limits, there are supply chain limits, there are logistical limits on how businesses actually adapt technology, and there are basic economic limits. And if you're a STEM student or honestly any person making career decisions right now, this video matters because when you're fearful, you can make very terrible decisions or decisions that are not based on logic or facts. We have people dropping out of engineering programs. We have people quitting careers because of things that CEOs are saying and they're spiraling. So, let's provide a different perspective based on analysis and not fear-mongering.
First, let's define what we're actually talking about when people say AI will replace all white collar jobs. What they're referring to is something called a large neuronet network and specifically a large language model. These models basically can read text, generate text, write emails, analyze images, summarize documents, and they can answer questions. The core architecture behind these models are something called a transformer, which came from Google's research facility in 2017 and a paper called attention is all you need. The reason why this paper is important is because it made AI or specifically transformers scalable when it comes to our GPUs which is the whole reason why what we know as AI exists today.
Strip away everything that you know and a neuronet network is a mathematical function that has adjustable parameters or weights. So in a neural network you take inputs you process them through a crap ton of mathematical equations and within these mathematical equations you produce an output and during the training of these neural neural networks you can adjust the parameters or the weights of every single function and produce better outputs in math form it looks like this y= f(x) parameterized by theta and what this equation means is that a neuronet network is a function that maps input x to output Y. So it's the input and it maps to the output here parameterized by weights theta and this is the weights adjusting theta to reduce error. So you're adjusting the theta to reduce the error present in this equation. So this is the neuronet network general form equation.
So the specific idea of what makes transformers powerful is something called attention. Before transformers existed, language learning models basically read information one word at a time. It's like reading a sentence. You're reading this is a sentence, one word at a time. But transformers do something a little different. When they look at every single token at the same exact time, the words themselves can decide what other words matter the most to it. So, a basic example that people use when it comes to this is the animal didn't cross the street because it was too tired. So when you read the sentence, the animal doesn't cross the street because it was too tired. Does the word it refer to animal or does the word it refers to street? Obviously it refers to animal. But a language learning model doesn't necessarily know that. You know that because you're a human. So what attention does it lets every single word take in every other word and decide for itself how relevant is that word to myself? So that mathematical equation can be represented by attention of Q K V = the soft max of Q K transpose divided by the square root of D * V. What this means and I'll repeat it again for each token for each token compute similarity scores Q transpose scale by your scaling by this to prevent saturation soft max gets weights by summing to one my handwriting sucks then mix the value vectors and the result is weighted information sharing between all the tokens. So in basic terms, this gives you a probability distribution. Meaning this equation allows the words to decide how relevant another word is to itself. That's basically it. It's a mathematical equation. It allows the words to decide, is this word relevant to me? Nah, I'll give it a lower weight. Oh, is this word relevant to me? It is. I'll give it a higher weight. It's basically relevant. How relevant are you to me?
So, the reason why transformers or AI as we know it today became so aggressively popular is because they're extremely fast. They're they're speed the speed is insane. Think of it like this. Before transformers, we use something called RNNs, which is recurrent neuronet networks. And an RNN reads every single word one by one. Like I said earlier, like you're reading a sentence. In basic claims, let's say we have a word that says a cat is chasing a mouse and I'm writing. In order for a recurrent neural network to understand this sentence, it would have to read every single one of these words in order. So, in order for it to understand mouse, it would have to go through all of these words first. And you can't really speed it up. Every step is a step-by-step process. It doesn't matter how many processors you have, how strong your GPU is. None of that matters. The model is going to get stuck at every word and go down one by one in a straight line. Transformers basically broke that during training and when you're processing your input, transformers can look at all the words at the same time. So, going back to this sentence, an RNN is 1 2 3 4 5 6. A transformer is all of them at once transformer. And that makes it extremely efficient when you're training these models on input data. And the reason why this is important is because these models run on GPUs or the chips that AI runs on is GPUs. And GPUs are designed exactly for stuff like this that doesn't necessarily run in order, but rather runs in parallel to each other. So transformers relative to recurrent neuronet networks are extremely efficient.
Now this probably confuses you because when you're reading a sentence from chat GBT or Perplexity or Gemini, it still writes one word at a time to you. But that has nothing to do with how the model is being trained because when it's being trained and that's the part that costs hundreds of millions of dollars where you see all these data centers and stuff, that's where being in parallel really matters the most. And that's what Transformers basically unlocked in society. That's why you're seeing such fast growth when it comes to AI where AI just a few years ago can only take in millions of parameters. Now it's taking in billions of parameters which gives the perspective that we're experiencing neverending exponential growth.
So so far you're probably like well it sounds like AI is improving fast and it's just basic mathematics on why it will continue to improve. But this concept is probably the single most important concept of this entire video and why that statement is not necessarily accurate. If you understand this concept, you will never fall for the doomer oh AI will take everyone's job and we'll all be homeless perspective again.
So the S-curve has four phases. The first phase is early the early phase. Second is inflection. Then you have the steep climb and the upper bend. So the first phase early progress is very slow. It's inefficient and most people ignore it. This phase happened around 2017 to 2021 when it came to AI for AI and this is when the technology is very new progress is slow it's inefficient and people ignore it and around phase two inflection is when someone discovers a key idea that unlocks scaling and the key idea that was discovered was actually in 2017 however AI or chat GBT and stuff like that didn't start taking off until about 2021ish. So this is about starting about 2021 where people started really putting money into AI and and saying finding out that it's scalable with GPUs and all these things and that's when people would say oh wow AI is really taking off we have to really pay attention to this and then you have the steep climb where everything feels explosive everything is scaling extremely efficiently and every investment is producing dramatic results. So this is where AI was into the phase like oh it's taking jobs like people had the perspective that during this phase all jobs would be taken that's what people are saying they're like oh AI is taking jobs and the next phase is the upper bend which constraints start dominating you're still improving but the cost to improvement ratio is not as crazy as it was in the inflection phase where it's just shooting straight up and the steep climb where it's shooting even further up. And the reason why I'm bringing this up is because this is not a theory that I just made up. I didn't just make up a random chart. This is an observed pattern that happens all across engineering when it comes to semiconductors, networking, battery technology, imaging sensors, everything. This is not something that just randomly happens for AI. This is a very normal thing that happens throughout the entirety of technology. And this is where the doomer narrative completely just falls apart. The people saying AI is going to be exponential forever and they just assumed that it was going to be an infinite trend that it's going to last forever. And that's just not the reality of any technological advancement in human history.
So the real question is where exactly are we on the curve? And that's relatively debatable. People at Frontier Labs might argue we're still mid-steep, but the constraints I'm about to show you are increasingly the binding factor. And these constraints are going to basically show why I think we're so far up and we're nearing the upper bend. Either way though, we're we're obviously in phase three of the AI, but we're either mid-steep or we're entering the bend. The argument still holds. the curve will flatten. Infinite exponential growth is improbable and it's just not how technologies behave.
And guys, this is not just my opinion because I'm not doing another think piece on what I think AI will happen with. There's literal research on this, published research on this. In 2020, researchers at OpenAI, which is Sam Altman's own company, published a paper called scaling laws for neural language models. The key finding in this paper was that model performance improves predictably as you add compute data and parameters. But of course, it follows a power law. So the paper itself says almost word for word that performance improves predictably as long as we scale up N and D in tandem, but enters a regime of diminishing returns if either N or D is held fixed while the other increases. So what exactly does this mean? This statement means by OpenAI themselves that scaling works. Scaling works pretty good and is determined by mathematical equations which shows why GPT-2 to GPT-3 felt like a massive leap in terms of production and usability from AI. However, this means also that the jump from GPT-4 to GPT-5 costs vastly more but delivers a much smaller perceptive leap in performance from AI. As technology progresses, it has massive leaps throughout the middle. And I'm writing yellow note, massive leaps. Um, which shows why early models of GPT seemed insane. But scaling although it works gets more expensive per unit of improvement and you notice it less as it progresses. It basically says the there's not necessarily a linear relationship between GPU power or GPUs scaling and AI.
So there are counterarguments and there are ways that companies are trying to navigate this because this paper is not it didn't release a few months ago. they they've known this and there's a newer scaling axis called test time compute. This is the approach behind models like OpenAI's 01 and 03 basically their newer models which says rather than just making the model bigger you spend more compute during inference. So you're letting the model think longer on harder problems. That's why when you open AI sometimes you think you see a thinking option and in this thinking option you don't get the answer immediately because they decided that wow we can't just make models bigger and the the models get better and it works it produces real capability gains especially on math coding and reasoning tests but here's the thing behind these models test time compute and that's what it's called when you have the thinking and all that stuff makes my infrastructure argument stronger It doesn't make my argument weaker. If every query now requires 10 times or 100 times more inference compute because the model is thinking you need way more GPUs, more power, more memory bandwidth per user. The deployment cost of replacing a white collar worker goes up. It doesn't go down. So with the initial math behind it, they were just allowed to make GPT. They were just allowed to have more GPUs because earlier more GPUs did lead to bigger models and bigger models led to smarter. But now you have to have thinking and when it comes to thinking you need more infrastructure, not less and it makes the replace white collar narrative. So it's a capability game that comes with much more infrastructure cost. And I will say this, these infrastructure costs are technically possible to absorb by companies. But I will also say this, quantization, which is when you have a larger quantity of things, reduces the precision of the weights. Meaning, yes, they could just have more infrastructure, but that will also lead to another problem. So instead of 16 bits per parameter, you use eight or four. That cuts memory, but it speeds up inference. Other architectures that companies can deploy and I'm going to read some of them out to you. Mixture of expert architectures which only activate a fraction of the model's total parameters per query. So a one trillion parameter model might only use 100 billion parameters for any given output. And distillation trains smaller models to mimic larger ones. And there's something called speculative decoding which uses a small fast model to draft tokens and a large model to verify which speeds up all of this all of these processes and it speeds up inference or running information and all of these are things that make AI better. So quantization it produces it reduces weights mixture of experts and distillation and speculative decoding and these are good engineering wins and they do meaningfully reduce cost per token if you were to run AI but these things don't eliminate constraints that these companies are facing. They just push them further out. Meaning you still hit memory ceilings. The memory ceilings don't disappear. Bandwidth limits and power costs obviously don't disappear. You hit them at a higher scale than you would have hit them without these techniques, but you still hit them eventually. So the problem is just better masked. The initial problem is not solved. So efficiency improvements are also why I'm not saying AI will stop getting better. I'm saying the rate of improvement per dollar is showing. And the absolute infrastructure required for civilization scale AI is enormous and is bigger than you think it is. So the catalyst that comes with stopping AI from just taking all jobs is not because the AI is not sophisticated enough. It's just a hardware architecture fit. It's not magic.
So to recap why AI improved so fast. It wasn't divine intelligence from Silicon Valley. Two things basically aligned at the perfect time. Transformers were able to scale in parallel and that paper came out in 2017. And during that same time, GPUs accelerate exactly the operations that transformers need such as what statistics, probability, and matrix math. And because those things just happen to happen at the same time, boom, we have an AI growth. But that alignment produced the steep part of the S-curve that I showed earlier. And that boom is what everyone felt. But alignment from the Transformers architecture to hardware GPUs is a one-time unlock. Once you've captured that, and we did meaningfully capture that, the next leap requires either a fundamentally new architecture, meaning Transformers will need something new. We need a new type of of language learning models in general, or a fundamentally new hardware paradigm. And right now, we have neither. We use transformers. That's what they use. and our hardware is actually being stumped by this the GPU processing and you don't get to just replay the same unlock two times in a row.
The S-curve explains why we can see impressive demos, real legitimate improvements in AI while still not seeing overnight replacement of entire industries. Demos basically operate at the speed of the GPU cluster that they're on. The real world which we all live in operates at the speed of hardware manufacturing such as Micron, Samsung, Nvidia. Power grid buildout which is why everyone's electric bills is going up and the speed at which enterprises can take in this AI and dish it out to all his workers. And obviously institutional trust you have to trust the companies who are running AI. These are completely different time scales than Transformers and their architectures. And the Doom narrative massively conflates these time scales constantly. We even have CEOs conflating it.
So, back to the initial claim. All white collar jobs will be gone soon. Because that's why I made this very long mathematical nerdy video. To accept that timeline, you need a way more smart model. You need civilization scale deployment of AI. That means replacing, not assisting, replacing hundreds of millions of knowledge workers across every industry simultaneously. For that to happen, the claim has to survive one of these constraints.
So, the first claim, hardware constraints, and this is the one where a lot of people just don't fully understand. Most people think that AI's problem is just GPUs, which are graphic processing units. But the real problem with AI is memory. When people think about AI scaling, it's not just added more GPUs. Two things specifically need to happen. You need to have strong memory capacity, how much data can fit on one chip, and you need to have memory bandwidth. How fast the data moves between memory and compute units. If your model doesn't fit in memory, you literally cannot run it efficiently. If data moves too slowly, your processors sit idly waiting by. It doesn't matter how fast they are. So, let's make this concrete. Capacity is the size of your bucket. How much data can physically sit on one chip at once. Bandwidth is the width of the pipe. How fast can data flow from memory to processors? And latency is how long you wait before the first drop of water arrives. You can have the fastest processor in the world and it will still be slow if your memory pipe is just too narrow to carry the information. That's called being memory bound. And it's a very dominant bottleneck for most large AI inference workloads. Meaning a lot of AI companies have extremely great processors, but they just aren't able to get the memory out fast enough for you to get an answer. And let's use real hardware because you probably think I'm speaking in abstract, but no. The Nvidia H100 chip is one of the most widely deployed AI GPUs right now. It has 80 GB of high bandwidth memory and about 3.35 terabytes per second of memory bandwidth. The newer H200 bumps that up to 100 GB of higher bandwidth. And it's a real improvement, but it's still a physical hard ceiling per chip. These numbers are set by chip design and memory technology. You can't just wish for them to get bigger. Meaning this is not a software problem. This is real hardware constraints, physics constraints. And there's a mathematical equation that that shows this and it's called the weight memory. So let's write it out. Weight memory is approximately P * B. What this means is P is the number of parameters and B is the bytes per parameter. An example would be 100 bytes or 100 parameters times two bytes equal 200 GB and this already exceeds a single Nvidia chip. This basic equation already exceeds a single Nvidia chip. So 80 GB, remember the H100 Nvidia, the weights alone are more than double what fits on a single GPU. And that's just the weights. We haven't had a single input yet. So what am I saying? cuz it's a lot of nerd engineering talk. What exactly am I saying? For a model to run or an AI model to run on a bigger chip or or a bigger size or more data, you have to split it across multiple GPUs. That means you need fast interconnects between the GPUs. You mean you need more hardware, more cost, more complexity. And we're still talking about a model that's considered mid-size. We're not talking about crazy things. I'm talking about a 100 billion times two, which is just 200 billion bytes and 200 gigabytes. That's mod that's modest size when you're talking about putting this out for everyone's jobs to be gone. I'm unsure if if technology companies even have what it takes to literally do this in modern times. Meaning, it's not literally possible today. And when you're talking about training memory, it's much worse. During training, you don't just store weights onto chips. You also store activations, gradients, optimizer states. And these aren't things that you can just make up in society. You have to actually have them on every single chip in all of your training exercises.
So, we talked about all the hardware constraints, and we talked about attention, which is what transformers use, where every token looks at every token, and that's how models understand context. But here's the problem. When you're chatting with AI and it's generating a response, it needs to remember everything that came in the conversation. Specifically, it stores what are called key and value vectors when you're talking to it. And they do it for every single previous token. That's why it doesn't have to redo all the math or attention math from scratch after it generates an entirely new word. The stored memory is called the KV cache. That's like, think of it like a running tab. For every message you send, every response it gives, the tab or the receipt just gets bigger and bigger. And the KV memory is the KV cache can be represented with a pretty basic formula, which is just approximately 2 * the number of layers * the context or the the context length, how long is the conversation, * the hidden size, * the bytes per value. And what this means is factor of two equals keys plus values stored separately and grow and this model grows linearly with context length. So basically longer conversations have more memory per user and fewer users per GPU. The takeaway is relatively simple. The longer you talk to AI, the bigger the tab is for your conversation. This is why context windows are not free. When a company says, "Our model supports 1 million tokens of context," that sounds amazing, but what they're really saying is each user's conversation can eat a massive amount of GPU memory. And here's why doomers are wrong. Again, if each user's conversation takes up more memory, you can fit fewer users on every single GPU, which is what I referenced earlier. It's a trade-off. Longer context means fewer simultaneous users. So, picture this. If longer context equals fewer simultaneous users, millions of AI agents running all day, replacing human workers, handling complex multi-step tasks, maintaining long conversation histories, how how is that going to happen? Every single one of these agents is a KV cache sitting in GPU memory. They're not just abstract things. They're on literal GPU memory. Every one of them running up a tab and you're paying for that tab in hardware. Multiply that by millions of agents, you have an infrastructure build that makes current AI buildout look extremely small. So that reality would be catastrophic relative to our current reality, making it improbable for everyone.
Why faster chips don't always help. The reason why faster chips don't always help is because a lot of AI models are memory bound and buying a faster processor does nothing. It's like upgrading a faster engine, but keeping the same fuel lines in your car. The engine can burn fuel faster, but the fuel can't get to the models fast enough. So, the engine just sits idly waiting, which a GPU will be doing, sitting idly waiting. For large language model inference or when you use large language models, the thing that needs to run those millions of AI worker agents, the workload is almost always memory bandwidth bound means the speed of the AI response is limited by how fast you can actually move the data, not how fast you can do the math on a newer chip. And the speed of data movement is set by memory technology on the chip. That's a literal physics constraint, if I've ever heard one. you can't code your way around it. So throughout this video, the things that I'm saying is that a lot of AI's problems are due to literal physics and chips and engineering constraints, not necessarily the speed of the GPUs or how advanced the transformers are. Well, it is kind of how advanced the transformers are, but yeah. And there's another thing called Amdahl's law, which is why just add more GPUs literally doesn't work because more GPUs doesn't linearly lead to a speed up in how fast AI works. Amdahl's law can be shown with this equation. S = 1 / (1 - P + P / N). What this equation means is S is the actual speed up. P is parallel. P is the parallel fraction. N is the number of processors. So let me plug in some real numbers because I've just been writing equations. So say 95% of your work is parallelizable, meaning it can all run at the same time. That's pretty generous. P = 0.95 and you throw a 1,000 GPUs at it. So you have a 1,000 GPUs. So S = 1 / (0.05 + 0.95 / 1000). The answer is about 19.6. So what does this mean? If you have a 1,000 more GPUs, your system is not a 1,000 times better. Your system is 19.6 times better. Let that sit with you for a second. The reason is that 5% of work can never be ran in parallel. the synchronization, the communication between GPUs, the data dependencies, that 5% becomes a wall over time. No matter how many GPUs you throw at it, the wall doesn't move. So what does this mean? This proves that no matter how many GPUs you add, like, oh, just add more GPUs for faster compute. Um, Amdahl's law proves that there's always a wall. It doesn't just scale linearly. If you add a thousand more GPUs over time, you don't get a thousand times better. If you add a million more GPUs over time, you don't get a million times better. And this will matter if you're trying to replace all jobs. It's not literally possible.
The next thing is just pure manufacturing. This should be common sense for a lot of these engineer like CEOs, but they are still putting it out as if it's not a constraint. Even with unlimited memory, you have unlimited GPUs, you have unlimited processing, you have the most complex and advanced pipes between the GPUs as possible. You have no KV cache. You cannot manufacture AI chips fast enough. Modern AI GPUs require three extremely specialized technologies with limited world supply. First, it requires advanced lithography. There are machines or if you don't know what advanced lithography is, these are basically machines that print circuits onto chips. One company on the entire planet makes the most advanced ones. ASML in the Netherlands. Their EUV machines cost $200 to 370 million each and they weigh 180 tons and they need three Boeing 747s to ship. They contain over a 100,000 individual parts and lead times are measured in years, not months. meaning you have to order these machines years in advance. Second, you need high band memory or HBM. These are stacked memory chips that sit on top of AI GPUs. Think of SK Hynix or Samsung. They make them. They make the most advanced ones. And obviously, production is limited and it's already booked out. And third, you need advanced packaging. The technology that bonds the GPU die to the HBM stacks in one unit. So, you need all these things simultaneously at the same time. And we need them on a massive higher scale than we have them now, which are which is literally impossible.
Now, let's zoom all the way out because I've been on this long long spew and look at what AI is. AI is electricity that's converted into tokens. At the end of the day, AI is just electricity being converted into computation and heat. That's all it is. Every GPU consumes hundreds of watts. A rack of GPUs consume kilowatts. A cluster consumes megawatts. A data center consumes tens to hundreds of megawatts per data center. You cannot scale AI deployment faster than you can scale power generation, grid connections, data center construction, and cooling. It's impossible. In these timelines, everything that I've laid out in the second half of this video is not just the software or the transformers being limited, which the transformers are semi limited because they hallucinate, but these are hardware constraints. Even if you want to permit a new data center, that takes between 6 to 24 months if everything goes perfectly depending on where you're building. Large power transformers, the electrical kind, not the AI kind, 18 to 16 months lead time. Custom builds require heavy equipment and you cannot rush these off of an assembly line. So just to build the data centers will take months just to get the equipment the the fab centers years. It's not something that you can do in a year or 18 months. And you can't just run these in parallel like you can a software sprint. Some steps depend on other steps being done first.
If you still don't believe me, I'm going to do the math on the energy scale that it would take to replace all white collar workers. Let's say there are 100 million workers. And we're going to replace them all with AI agents because we're tired of people having jobs and buying stuff with their money. So, let's do it. 100 million. I'm just going to write M to make it easier for myself. times 700 watts equals 70 GW plus 30 to 50% cooling overhead gives say let's say cooling overhead because yeah let's factor in cooling cuz people pretend like that doesn't exist gives about 90 to 105 GW in total. There are roughly 100 million white collar workers in the US to replace you need always on AI agents running all day not just a quick chat GBT search, you need you need these people running all day. Let's be generous. Let's say each agent requires one high-end GPU and practice complex tasks with long context. Probably need more than that. But let's just be let's just be generous. Let's say Sam Altman and and the guy who found Claude and Gemini and Mark Zuckerberg, they just found we only need one per person. An H100 draws about 700 watts underload. Meaning that the average standard chip Nvidia makes for AI has about 700 watts underload. 100 million I'm writing this out. 100 million times 700 watts. And that's just standard. I'm I'm being very conservative here equals 70 billion 70 what is it? What is it? 70 billion watts or 70 GW. And this is just for the GPUs needed to replace white collar workers. just for the GPUs. I didn't add cooling here. And here's the thing, data centers typically 30 to 50% overhead. That brings it to roughly 90 to 105 GW. For reference, and here's the crazy thing. For reference, the entire US data center industry right now consumes about 20 to 30 GW total for everything, not just AI. For every single thing that we use data centers for in the United States, it's 20 to 30 GW total. So just in the most conservative explanation possible to replace workers would take a minimum of 70 GW and that's no cooling. That's no overhead. That is more than double what the entire data center industry in the entire United States takes up for everything. So the doomer narrative requires tripling to quadrupling the entire US data center footprint in a year and a half. Cut it in half for efficiency. You still need double the entire grid. That's not software. That's impossible. That's like power. That's power plants. That's transmission lines. That's not even a conclusion someone can come to in a right mind and actually make sense to you. Even if AI models get dramatically better tomorrow. The physical infrastructure needed to deploy them at civilization scale at minimum takes years to build. That's not me being pessimistic. That's not being an anti-AI. That's just how the real world works. You can't just update software overnight and take people's jobs. You can't build a power plant overnight.
And even if you look at that standard person's job, a job is not just one task that AI can do. A job that you go to for work, a white collar job is a bundle of tasks. And the narrative behind AI will take jobs gets this completely wrong. So AI handles some tasks well and it handles some tasks extremely poorly. It handles things extremely poorly. We need something called accountability. So who's on the hook when something goes wrong? Because if something goes wrong at your job, you're on the hook. And we need trust and access. So we need someone who's authorized to enter the building, access the systems, sign the documents and liability whose name is on these things. So even if AI might handle 30 to 60% of tasks in a row, the remaining 40% the judgment calls who's accountable, who's liable, who do we sue if something goes wrong, organizations literally, they can't hand that to a model. Someone has to be accountable for these things. And I haven't even got into in detail hallucinations and the uncertainty that transformers have right now in AI in high stake domains such as legal filings, medical decisions, financial compliance, engineering signoffs. A 1% error rate can be disqualifying. So the cost of one wrong answer isn't just oops, let's try again. It's a lawsuit. It's somebody dying at work. It's a it's a regulatory violation. It's a bridge falling. So if a current AI model hallucinates, they give a confident, well-written, completely wrong answer and they have no idea why they're wrong, until reliability reaches the threshold that each industry requires, you need a human checking the work. And a human checking the work means the job still exists. And you're like, well, AI might not hallucinate long term. Wrong. Again, the hallucination problem that comes with AI is not just something that they can just code out. It's a structural issue. It's not just a bug. So, people assume that hallucination will just get fixed in the next update, but it's baked into how transformers work in general. The best training objective for a large language model is token prediction. The model learns to generate text that sound plausible, like something a human could write. It's not literally sitting there checking a fact in a database. It's just pattern matching at an extremely massive sophisticated scale. So companies add layers to improve accuracy and then they train the models to get better responses. And and something called retrieval augmented generation allows the models to look up through real documents and use a like the sim the same way a calculator calculates math or or searches the web or you know something searches the web. And the problem is that transformers are probabilistic meaning that they run on matrices and math probability and statistics. And you can get hallucination down from 15% to 3%. But if you're running on a probabilistic system, there's still a probability that something will be extremely wrong and you won't have any idea why it's wrong. So let's say models hallucinate from at a 3% rate right now for something specific like financial compliance. Getting that down from 3% to 0.01% is not just something that you can do. You need a team of researchers for that one specific area. So, it's a high stakes problem. We don't have the brain power right now to solve it. And in high stake industries, you need a basically 0% error rate. And the gaps between usually right and 100% reliable is where the timeline for this completely dies, especially in the next few years. And when you go to enterprises adopting AI, enterprise adoption is supposed to be slow. We have CEOs right now who are trying to make it faster obviously but in general it's slow on purpose because you need a demo, you need a pilot, you need security review, you need compliance reviews, you need integration correct integration, you need procurement and you need training and then you need a final roll out. So every big company runs through this checklist or they're supposed to run through this checklist and the big one is security review and these pipelines take months to years every single time. And a demo is not a deployed system. So when an industry or enterprise does things the functional way, they have to run through these things. They can't just say, "Oh, we're going to replace jobs tomorrow." Some CEOs have tried to do that and obviously every single one of them have backtracked because they've realized that it's improbable to do so.
And the last thing is like oh I don't care about anything you're saying in this hour long explanation. AI is different. AI will trans it will it will transcend all these hardware problems. It will transcend all these software problems. It is the next coming of a revolution. And before you have that perspective, just know people said the same thing about the internet. They said the same thing about cloud computing. They said the same thing about phones. They said the same thing about electricity. These technologies were obviously transformative. Without these technologies, I couldn't be making this video. And they did reshape the world dramatically over decades though, not months and not a short few years. AI will probably be faster because it's software heavy and it iterates quickly. But faster than historical precedent is not 18 months. Faster means 7 to 15 years and not 30 years. That's still pretty transformative and is still worth preparing for, just not an apocalypse level catastrophe that will take everyone's jobs in the next few years. Obviously, I'm attacking from an engineering side. But now let's look at this from an economic side.
Even if all the doomers and all the CEOs get unlimited compute tomorrow, they get unlimited power, they get unlimited resources, their model of how displacement works is still wrong. The core error actually has a name. It's called the lump of labor fallacy. The lump of labor fallacy is basically the assumption that there's a fixed amount of work in the economy and work is a pie and machine takes a slice, human loses a slice. Basically, they think that the economy is a zero-sum game. This feels right intuitively, but it has been empirically proven wrong every single time people have believed it for the last 200 years. Meaning, this fallacy has been proven wrong for 200 years in a row regarding our economy. So, when productivity increases, costs drop. When costs drop, demand expands. When demand expands, new tasks emerge, new roles get created, entire new industries appear, and the total amount of work is not just a fixed pie. It grows dramatically.
In another paradox called Jevons paradox, similar to my name, shows that efficiency creates even more demand. Jevons paradox, which is named after William Stanley Jevons in 1865, was not about AI. It was about coal. Because yes, this issue has been around I mean around for a long time. The doomer narratives have been around for a long time. When steam engines got more fuel efficient, total coal consumption didn't just drop, it went up. And efficiency made steam power cheaper, which made it more accessible to more industries, which expanded more usage. So the same thing can be applied to AI. Let's say AI makes financial analysis five times cheaper and five times easier. Companies won't just fire analysts and do the same analysis that they've been doing. They're going to try to do 10 times more analysis, exploring scenarios that they've never even considered before. That's what will happen with AI expansion. If AI makes legal document review 80% cheaper, law firms will be foolish to just cut paralegals. Many will offer services to smaller clients they weren't able to reach before, and they will expand their total market. Basically, efficiency almost never shrinks demand. And historically, society proves this. When ATMs came out, people thought there would be less bank teller bank tellers, but it was the opposite. ATMs rolled out in the 70s and the 80s, and everyone predicted that bank tellers would just not exist anymore. But what happened? ATMs made branches cheaper. Banks started opening more branches. And the total headcount for bank tellers increased over the decades. The job changed, the headcount went up. And what's funny is that bank tellers work a pretty standard 9 to 5 and a lot of you don't even know this.
So what am I getting at? The people pushing these narratives are going against engineering. They're going against economics. They're going against basic manufacturing processes. They're going against logistical errors. They're going against everything. They're even going against current data on AI because AI has been widely available for the past 3 years. Large language models and very advanced AI has been available for a a while now. So even if you forget everything I said, everything I said was a lie. He doesn't know what he's talking about. Let's talk about what actually happened with the data. If displacement was going to be fast, it should have shown up already. And it has shown up in a few ways. When it comes to freelance copywriting, some of that showed up. And when it comes to some coding tasks, customer support has been automated and some coding jobs have been automated. But broad white collar unemployment, massive AI-driven layoffs, not in the aggregate data. White collar unemployment is historically normal and consistent with previous trends, meaning there aren't crazy spikes in in white collar unemployment before and after AI. That's not proof that it can't happen, but it's proof that it didn't happen. It's proof that the data is not on the doomer's side when they say AI will replace jobs, and it's not on the CEO's sides either.
And I'm not just making this up because I didn't make up any of this. Anthropic's own data proves my point. This graph that's on the screen right now was published March 2026 by Anthropic, the company that makes Claude, which is one of the most advanced AI models currently on Earth. So this is not from a skeptic. This is not from a think tank trying to downplay AI. This is from the people who literally built AI and profit directly from its adoption. Look at this chart. The blue area shows what AI could theoretically do across different job categories. The theoretical capability and yeah, it's a massive capability. Computer and math jobs, 94% of tasks can be theoretically handled by a large language model. Office and admin jobs about 90%. Business, finance, legal, education. The blue reaches way out, meaning that these jobs can theoretically be done by AI. The blue area is what doomers are pointing out when they say all jobs will be gone. Now look at the red area. That's what's actually happening today. That's observed AI usage in real professional settings measured from actual cloud usage data. The red is tiny. Let this let this sit with you. For computer and math workers, 94% of theoretical capability, but only 33% actual coverage. And that's the highest adoption category. Everything else is even smaller than that. And you may think, "Oh, the gap is people just haven't found out about it yet. People haven't been using AI yet." No, the gap is everything I've explained in this video. The gap is memory bandwidth ceiling. The gap is power grids. The gap is enterprise security reviews. The gap is the hallucination problems that transformers naturally have due to their probabilistic nature. The gap is Amdahl's law. The gap is entire deployment realities. Those are the gaps. The doomers look at the blue and they extrapolate information that's not there. Engineers look at the gap and they say why does this gap exist? And it gets even better. Anthropic's own researchers Masenov and McCory looked at unemployment data for workers in the most AI exposed situations. And what did they find? They found that AI in general has not increased unemployment in jobs that are most exposed by it after two plus years. They did find one real signal. Hiring of young workers ages 22 to 25 have slowed down in most exposed occupations. And that's consistent to what I said earlier. Entry-level and routine work will get pressured first. And that's something that you have to pay attention to. But hiring has slowed for some entry-level roles is completely different than all white collar jobs will be gone in 18 months. So, here's something I
can't get over. Anthropic's own CEO said last year that AI could disrupt half of white collar jobs. And here's a clip of him saying it.
"Dario, you've said that AI could wipe out half of all entry-level white collar jobs and spike unemployment to 10 to 20%. How soon might that happen?"
"Well, let's uh uh uh well, first of all, thanks for having me on the show, but uh uh just to back up a little bit, you know, I've been building AI for over a decade, and I think maybe the most salient feature of the technology, and what is driving all of this is how fast the technology is getting better. Um, a couple years ago, you could say that AI models were maybe as good as a smart high school student. I would say that now they're as good as a smart college student and and and sort of reaching past that."
Then Enthropic's own research team publishes a paper showing actual adoption is a fraction of the theoretical capability. Unemployment effects are indistinguishable from zero. Yes, they literally wrote that it's indistinguishable from zero. And the gap between what AI can do and what AI is actually doing is enormous. The executives are saying one thing and their own researchers are are contradicting them in the same year. I'm not saying executives are lying, but when your own research company contradicts your public predictions, maybe the timeline that you're speaking about deserves way more scrutiny in general.
And the graph doesn't lie. The blue is potential, the red is reality. And these are from actual researchers at Anthropic. Now, I will be fair here. I will be fair. There is a smarter version of the displacement argument that I will respond to. You don't need to replace 100% of a role to eliminate a position. If an AI can handle 60% of a junior analyst task, the company doesn't have to keep the analyst. They can just celebrate the 60% gain and fire two out of five analysts and that will lose 40% of jobs. And that's a reasonable sounding argument. And in some specific niche cases, it can possibly happen like that.
But like I said earlier, the lump of labor fallacy in Javon's paradox explains why that doesn't scale to all white collar jobs disappear. The 60% argument assumes that there's a fixed demand. There's no new task and there will be no market expansion from these companies. Meaning that new projects can get funded, adjacent teams can expand, and new scopes can emerge when it comes to adopting AI. There's not a fixed amount of work that can happen. meaning that why are people losing jobs?
The people are losing jobs because there's an incentive alignment. If you're raising money to for an AI company, it makes sense for a CEO to say this AI is very important and you need to adopt it now. You need to fire people because these CEOs are trying to make money. They don't necessarily have the technological or engineering physics backing to say this will replace jobs, but it helps their pockets. It helps them when they have a fear mongering narrative. when they have extreme predictions regarding AI. It helps when they see companies laying off people even though they know the engineering proves otherwise. It helps them. It makes them look like they're the next coming of an AI or a technology mogul. But that's not reality. Economics don't prove it and engineering don't prove it.
But let me be intellectually honest here. There is one wild card that I will say can happen where AI can genuinely replace everyone's job, but it's unpredictable. And the wild card is a fundamentally new AI architecture. Meaning, we switch from Transformers to something completely new that we don't know exists yet. Because although Transformers change the way AI exists as we know it today, in its current capacity, it literally cannot replace jobs in 18 months. But if a new architecture comes out that does something completely different that scales exponentially without its relevance to GPUs, it's possible that can replace jobs. But I don't think it makes sense to form entire hypotheses and predictions based off of things that are impossible to predict.
So this is the entire video. It was a long one, but I wanted to not leave anything up for interpretation. Thank you all for watching and please make sure you subscribe, show some support. My name is Jovon and yeah, see you on the next one.