Three Explanations of LLM Learning and Generation

1. The Software Engineer

You know how you build APIs by defining contracts—input shapes, output shapes, and rules for transformation? LLMs work similarly, except the "rules" aren't hand-coded; they're learned from data. Here's the key insight: during training, we show the model billions of text sequences and ask it to predict the next token (word fragment). This sounds trivial, but it's actually a compression problem. To predict well across diverse contexts, the model must internally build representations of grammar, facts, causality, and reasoning patterns—because those patterns reduce prediction error.

The architecture (Transformers) is where the cleverness lives. Imagine you're building a system that needs to route requests through a complex dependency graph. The Transformer uses "attention mechanisms"—essentially a learned routing layer that decides which previous tokens matter for predicting the next one. During training, billions of examples teach it which routing patterns work. At inference time, you feed it a prompt, and it applies these learned patterns sequentially: token 1 → attention routes to relevant context → token 2 → repeat. It's not magic; it's just that the learned routing patterns, when composed across 96 layers, capture enough of human language structure to produce coherent text. The skepticism is warranted—it is "just" next-token prediction—but that simplicity is precisely why it scales. You're not building explicit logic; you're learning compressed representations of patterns in data, similar to how your distributed system learns optimal request routing through empirical observation.

2. The Physicist

The core operation is straightforward linear algebra: given a sequence of token embeddings (vectors in ℝ^d), compute a sequence of output vectors via repeated application of a learned transformation. Specifically, each Transformer layer applies: (1) a multi-head attention operation, which computes softmax(Q K^T / √d) V for learned projections Q, K, V; and (2) a feed-forward network (two dense layers with ReLU). Both are differentiable, so the entire pipeline is end-to-end differentiable.

What's genuinely novel—and what separates this from previous neural networks—is the scaling law. Empirically, loss decreases as a power law with model size (parameters N), data volume (D), and compute (C): Loss ∝ N^{-α} where α ≈ 0.07. This wasn't obvious a priori. Previous deep learning intuition suggested overfitting would dominate; instead, we see smooth, predictable improvement. This suggests the function class (Transformer) has favorable inductive bias for language. The second novelty is in-context learning: these models exhibit few-shot behavior without gradient updates. Mechanistically, the attention mechanism can implement simple algorithms (like copying or sorting) in a single forward pass, which is remarkable but not mysterious—it's a consequence of the expressiveness of the function class and what the data distribution incentivizes.

The hype conflates two things: (1) the genuine algorithmic/empirical insight (scaling laws, attention mechanism) and (2) the marketing claim that this is "reasoning" or "understanding." Mathematically, you have a high-dimensional function approximator optimized for next-token prediction. It's powerful because language data is high-dimensional and structured, and the Transformer is an expressive architecture. But it's not solving novel equations or discovering new physics—it's interpolating in a learned embedding space. The real question is whether the learned representations are rich enough to support downstream tasks, which they empirically are, but that's an empirical observation, not a theoretical guarantee.

3. The Venture Capitalist

Here's what matters: Can this be defensibly better than competitors?

The technical moat has three components. First, training data and compute: training a frontier model costs $100M+ and requires access to most publicly available text. This creates a capital barrier—a startup can't casually replicate GPT-4 from scratch. However, this moat is eroding; open-source models (Llama, Mistral) are now competitive for many tasks, and data is becoming a commodity. Second, architecture and training efficiency: marginal improvements in how you train or structure the model (better attention, better optimization) can reduce training cost by 30-50%, which is material. But these improvements are published and copied within months. Third, fine-tuning and alignment: teaching the model to be helpful, harmless, and honest requires domain expertise, data, and reinforcement learning infrastructure. This is defensible if you control the use case (e.g., customer service for your specific vertical) but not defensible as a general-purpose product against larger competitors.

The credibility question: founders should be able to explain why their approach is different. Red flags: "We have a new architecture that's 10x better" (extraordinary claims require extraordinary evidence; show benchmarks, not promises). "We're fine-tuning GPT-4" (you're a services business, not a technology company). "We're using proprietary data" (data advantages are real but often overstated; verify the data is actually proprietary and defensible). Green flags: "We're optimizing for latency/cost in a specific domain" (concrete, measurable, defensible). "We're building tools for using LLMs better" (less exposed to model commoditization). "We've achieved 2x inference speed at 95% of GPT-4 accuracy on our benchmark" (specific, verifiable, valuable). The fundamental question: Would customers switch to a better/cheaper model, or are you entrenched? If it's the former, you're in a race to the bottom. If it's the latter, you have a business.