The "Bitter Lesson" is not just about using more compute, but leveraging it scalably. Current LLMs are inefficient because they only learn during a discrete training phase, not during deployment where most computation occurs. This reliance on a special, data-intensive training period is not a scalable use of computational resources.

Unlike simple classification (one pass), generative AI performs recursive inference. Each new token (word, pixel) requires a full pass through the model, turning a single prompt into a series of demanding computations. This makes inference a major, ongoing driver of GPU demand, rivaling training.

Building features like custom commands and sub-agents can look like reliable, deterministic workflows. However, because they are built on non-deterministic LLMs, they fail unpredictably. This misleads users into trusting a fragile abstraction and ultimately results in a poor experience.

Model architecture decisions directly impact inference performance. AI company Zyphra pre-selects target hardware and then chooses model parameters—such as a hidden dimension with many powers of two—to align with how GPUs split up workloads, maximizing efficiency from day one.

Traditional video models process an entire clip at once, causing delays. Descartes' Mirage model is autoregressive, predicting only the next frame based on the input stream and previously generated frames. This LLM-like approach is what enables its real-time, low-latency performance.

GPUs were designed for graphics, not AI. It was a "twist of fate" that their massively parallel architecture suited AI workloads. Chips designed from scratch for AI would be much more efficient, opening the door for new startups to build better, more specialized hardware and challenge incumbents.

Contrary to the idea that infrastructure problems get commoditized, AI inference is growing more complex. This is driven by three factors: (1) increasing model scale (multi-trillion parameters), (2) greater diversity in model architectures and hardware, and (3) the shift to agentic systems that require managing long-lived, unpredictable state.

Traditional ML used "micro-batching" by normalizing inputs to the same size. LLMs break this model due to variable input/output lengths. The core innovation is continuous processing, handling one token at a time across all active requests, which creates complex scheduling and memory challenges solved by techniques like PagedAttention.

AI's computational needs are not just from initial training. They compound exponentially due to post-training (reinforcement learning) and inference (multi-step reasoning), creating a much larger demand profile than previously understood and driving a billion-X increase in compute.