Let's start with what an LLM actually is: a massive mathematical function that predicts the next word in a sequence. That's it. No consciousness, no understanding in the human sense, just statistical pattern matching at a scale that starts to look eerily intelligent. When you type "The capital of France is" into Claude or ChatGPT, the model doesn't "know" Paris exists. It calculates that based on billions of training examples, the token "Paris" has the highest probability of appearing next. It's autocomplete that read the entire internet. The architecture that makes this possible is called a transformer, invented by Google researchers in 2017. Here's the step by step breakdown of how text becomes a response. First, tokenization: Your input text gets chopped into pieces called tokens (roughly 3/4 of a word each). "Hello world" becomes ["Hello", " world"]. Second, embedding: Each token converts to a vector, a list of hundreds of numbers representing its meaning in mathematical space. Words with similar meanings cluster together in this abstract geometry. Third, the attention mechanism: This is the secret sauce. The model calculates which tokens should "pay attention" to which other tokens. In "The animal didn't cross the street because it was too tired," attention helps the model link "it" to "animal" rather than "street." Fourth, layers of processing: Modern LLMs like GPT-4 or Claude 3 have dozens of transformer layers, each refining the representation. Early layers spot syntax patterns, middle layers grasp semantics, late layers handle reasoning. Fifth, prediction: The final layer outputs probabilities for every possible next token. The model samples from this distribution (with some randomness for creativity) and outputs a word. Then it feeds that word back in and predicts the next one. Repeat until done. The math behind attention is where the magic happens. For each position in the sequence, the model computes attention scores using queries, keys, and values. What this means: Q (query) asks "what am I looking for?", K (key) answers "what do I contain?", and V (value) provides "here's my actual information." The dot product between queries and keys measures similarity between them. Positions with high similarity get high attention weights after applying a softmax function. Those weights then pull relevant information from the values. For example, processing "The cat sat on the mat because it was soft": When predicting after "it", the query for "it" computes high similarity with the key for "mat" (softmax might give 0.7 weight) and low similarity with "cat" (0.1 weight). The value vectors get weighted by these scores, so "it" pulls mostly from "mat"'s representation. This flows through dozens of layers, each refining the associations. The training process is where things get expensive and wild. These models learn through a process called unsupervised learning on massive text datasets: books, websites, code repositories, scientific papers. During pre-training, researchers show the model text with some words masked out and make it predict them. Every wrong guess generates an error signal that nudges billions of parameters (the adjustable numbers inside the model) slightly closer to better predictions. The process uses gradient descent at massive scale. Every single parameter update is trying to minimize the cross entropy loss: how surprised is the model by the actual next token? If it predicted "Paris" with 95% probability and "Paris" was correct, loss is tiny. If it predicted "London" with 60% probability but "Paris" was correct, loss is huge. Do this across trillions of examples and billions of parameters, and the model learns the statistical structure of language. GPT-4 reportedly has over 1 trillion parameters and cost over $100 million to train on tens of thousands of GPUs running for months. After pre-training comes fine-tuning, where human trainers rate the model's responses, teaching it to be helpful rather than just accurate. This is why Claude won't write your manifesto or ChatGPT refuses certain requests. Reinforcement learning from human feedback (RLHF) literally rewires the probability distributions to avoid toxic outputs. The model learns that certain response patterns get higher scores from humans, so it favors those patterns. Now here's why predictive industries are getting steamrolled. Any domain where success means recognizing patterns and forecasting outcomes is fundamentally an LLM problem at its core. Weather prediction? Pattern matching historical atmospheric data. Stock trading? Finding correlations in price movements and news sentiment. Medical diagnosis? Matching symptoms to disease patterns from millions of case histories. Legal document review? Identifying relevant precedents and contract clauses. Customer service? Predicting what response solves the problem based on past tickets. The AI advantage is brutal and simple: scale. A human radiologist might review 50,000 x-rays in a career. An AI model trains on 50 million. A hedge fund analyst reads hundreds of earnings reports quarterly. An LLM processes every earnings call transcript, SEC filing, and analyst note published globally in real-time. The pattern recognition advantage compounds until human expertise becomes decorative rather than functional. We're already seeing the takeover accelerate. Bloomberg GPT, trained specifically on financial data, outperforms human analysts at predicting market movements from news. DeepMind's AlphaFold solved the protein folding problem that stumped biologists for 50 years by treating molecular structures as a prediction task. GitHub Copilot writes 40% of code in repositories where it's enabled. These aren't narrow tools helping humans, they're replacement technologies that make entire skill sets obsolete. The uncomfortable truth is that most knowledge work is prediction dressed up as expertise. Lawyers predict case outcomes and draft documents following templates. Accountants predict tax liability and audit risk. Marketers predict which message will resonate with which audience. Teachers predict which explanation will clarify a concept. Doctors predict which treatment will cure which ailment. If your job involves looking at data and saying "based on X, I think Y will happen," you're in the blast radius. The only question is timing.
💻 technology
AI Swallowed Math and Now It Wants Everything Else
Large language models aren't magic. They're prediction machines on steroids, trained on trillions of text fragments to guess what comes next. Understanding how they actually work reveals why AI is colonizing every industry that runs on patterns, from weather forecasts to stock trading. Here's the truth behind the curtain.
My Take
The AI evangelists selling you on augmentation and collaboration are lying or deluded. These systems aren't assistants, they're replacements waiting for the cost curve to drop below your salary. Every "AI won't replace you, but people using AI will" platitude ignores the obvious endpoint: why employ the person at all when the AI does 90% of the work? The remaining 10% will get absorbed by the next model generation within 18 months. What genuinely terrifies me isn't job loss. Economies adapt, new work emerges. It's the epistemological crisis brewing underneath. When prediction becomes free and instant, what happens to human judgment? We're building a civilization where increasingly nobody understands why anything works, we just trust the statistical oracle. The model says this drug will work, this defendant will reoffend, this loan applicant will default. We can't interrogate the reasoning because there isn't any, just 100 billion parameters encoding correlations we can't interpret. We're replacing explicable expertise with inexplicable accuracy, and that's a foundation for catastrophe when the training data stops representing reality.
What Happens Next
The prediction economy bifurcates into two tiers within 24 months. Tier one: commodity predictions where AI achieves 95%+ accuracy at near zero marginal cost. Basic legal docs, routine medical diagnoses, financial forecasting, customer service, entry level code. These collapse to utility pricing, pennies per query. Tier two: edge case predictions where the last 5% accuracy matters enough to pay human premiums. Novel drug design, constitutional law arguments, trading strategies in unprecedented market conditions. But here's the squeeze: tier two keeps shrinking as models improve and training data expands. The first Fortune 500 company will announce AI driven workforce reduction exceeding 50% before 2028, and it won't be a tech company. My money's on a major insurance provider or financial services firm, industries built entirely on risk assessment and prediction. They'll frame it as efficiency and shareholder value. The stock will jump 20% on the announcement. Other CEOs will take note. The race to replace humans with prediction engines will shift from quiet pilots to explicit corporate strategy. Meanwhile, we'll hit a training data wall that nobody's pricing in. These models have already consumed most of the internet's text. Synthetic data (AI generated content used to train future AI) introduces degradation with each generation, like photocopying a photocopy. The next leap requires either multimodal training at massive scale (video, audio, sensor data, not just text) or a completely different architecture beyond transformers. Whoever cracks that first doesn't just win the AI race, they obsolete every prediction-based industry simultaneously. The geopolitical implications of that concentration of power make the current regulatory hand wringing look quaint. We're not ready for what comes after prediction becomes free.
What History Tells Us
The mechanization of prediction mirrors the industrial revolution's mechanization of physical labor, but at 100x the speed. When power looms automated weaving in the 1810s, skilled textile workers had decades to adapt, retrain, or retire. The Luddites smashed machines in 1811 to 1816, but the transition took two generations to fully displace handweavers. We don't have that luxury this time. LLMs went from research curiosity to existential threat to knowledge workers in 36 months, from GPT-3's 2020 release to ChatGPT's late 2022 explosion to today's enterprise-wide deployments. A better parallel might be the 1960s to 1980s automation of calculation. Human 'computers' (literally people employed to perform calculations) vanished almost overnight when electronic computers became reliable. NASA employed hundreds of human computers through the early space program, including mathematician Katherine Johnson. By 1970 they were obsolete, replaced by IBM mainframes that calculated trajectories thousands of times faster. The difference: those displaced workers numbered in thousands. Today's prediction workers number in hundreds of millions globally. The lawyers, accountants, analysts, diagnosticians, and programmers whose core skill is pattern-based forecasting represent 30% to 40% of the knowledge economy. We've never automated away that large a chunk of the labor force that quickly. Historical precedent suggests chaos.
Market Impact
NVIDIA closed April 23, 2026 at $1,247 per share, up 340% from two years ago, purely on AI infrastructure demand. That run continues because every enterprise deploying LLMs needs their H100 and upcoming B100 GPUs. Short term bullish through 2027 as the build out accelerates, but watch for the plateau when cloud providers have sufficient capacity. The real asymmetric bet is shorting companies with prediction-heavy workforces that haven't announced AI transformation plans. Legacy insurance companies like MetLife (MET, currently $78) and tax preparation firms like H&R Block (HRB, $51) face existential margin compression as AI handles work currently done by thousands of actuaries and tax professionals. Bullish on AI-native companies like Palantir (PLTR, $34) and C3.ai (AI, $28) that sell prediction as a service directly to enterprises. The consulting firms (Accenture at ACN, $365, and Deloitte's parent entities) face a reckoning. They sell expensive human judgment, and AI is turning judgment into a commodity. Current valuations don't price in 40% to 60% revenue declines by 2029 as corporations realize they can buy the prediction engine directly rather than rent consultants to operate it.