🤖 AI engineering track
Same idea as the coding Lab, for AI: predict, change an input, inspect evidence, then defend the trade-off. Each page labels what is calculated, deterministically simulated, or only illustrated. No page should imply a real model call unless it actually makes one.
Tiny Transformer Evidence Lab
Prompt → ordered merges → embeddings → Q/K/V → causal attention → logits → seeded sampling, with editable fixtures, semantic-step replay, tensor values, and executable counterexamples.
Foundations
— what an LLM call even isit's a next-token loop over a context you control
greedy vs sampled, temperature, seed — watch tokens generate
system + messages assembled into the context
force valid JSON to a schema, validate, retry
model emits a call, host runs it, result fed back
Retrieval & orchestration
— make it grounded and reliablechunk → embed → retrieve top-k → stuff the prompt
fixed graph of steps, deterministic control flow
propose → human approves → then execute
every step logged: inputs, outputs, cost, retries
cases + graders → score a change before shipping
Agents
— loops, memory, protocolsThought → Action → Observation until it can answer
short-term context vs long-term recall, what to keep
the tool/resource protocol agents plug into
planner → workers → judge, fan-out and merge
read → edit → run tests → repeat on a repo
Product & ops
— ship it and keep it aliveroles, queues, handoffs across many agents
the human's view: approvals, runs, results
keys, rate limits, retries, cost, monitoring
Four interactive routes are linked here. Their evidence labels differ: numeric calculation, deterministic rule simulation, or illustrative replay. Planned topics remain marked “soon.”