PERMANENT COLLECTION / ENGLISH EDITION

Start with a question,
not a term.

The English collection opens twelve complete paths. Every listed exhibit must be understandable, operable, and explicit about what was computed, recorded, documented, or only illustrated.

complete exhibits
English

counterpart exhibits
Simplified Chinese

Spanish and Japanese remain unpublished. We do not present machine fallback as finished editorial work.

03—05, 10—11

MEMORY, ROUTING & HARDWARE

Where do the hidden engineering costs move?

Change one variable, inspect the account it affects, then compare the claim with a recorded Python run.

03 What does a KV cache save—and what keeps growing?Separate repeated projection work from retained history with exact browser arithmetic and a measured toy model. MECHANISMINTERACTIVE6–10 MIN 04 How does one token choose only a few experts?Move top-k, keep two parameter accounts separate, and inspect a real routing-imbalance simulation. MECHANISMINTERACTIVE6–10 MIN 05 Why does moving a model to another accelerator take more than three code edits?Pass four engineering gates and explore a real multi-metric numerical comparison. MECHANISMINTERACTIVE7–12 MIN 10 What is an operator—and why can one missing implementation stop a model?Follow an operation through schema, backend dispatch, kernel coverage, and scoped measurement. COMPUTEINTERACTIVE5–8 MIN 11 What does CANN do between an AI framework and Ascend hardware?Separate graph, operator, runtime, compiler, communication, and version responsibilities. STACKINTERACTIVE6–9 MIN

02, 09

MODELS & GENERATION

How does an answer emerge one token at a time?

Separate the model’s candidate scores from the runtime’s selection and the final visible text.

02 How do language models write? What does temperature change?Follow a five-step generation loop, then reshape a six-candidate distribution with one control. MECHANISMINTERACTIVE4–7 MIN 09 When someone says “AGI,” what exactly are they claiming?Separate capability breadth, performance, autonomy, and the evidence needed to place a claim. CLAIMINTERACTIVE5–8 MIN

01, 06—08, 12

AGENTS & ACTION

How does a request become an action?

Separate the model’s proposal from the program’s decision and the tool’s effect.

01 How can AI operate software? What actually does the work?Walk through a six-step tool call, then run it again without the tool and compare the records. MECHANISMINTERACTIVE3–8 MIN 06 What makes an AI agent more than a chat reply?Trace a bounded task across model proposals, program decisions, tool effects, and an explicit stop rule. SYSTEMINTERACTIVE5–8 MIN 07 Who keeps an agent’s loop turning—and decides when it stops?Inspect the software layer that assembles context, gates tools, stores results, and controls continuation. SYSTEMINTERACTIVE5–8 MIN 08 What does MCP standardize—and what does it leave undecided?Separate protocol roles, capability negotiation, host authorization, and the tool call itself. PROTOCOLINTERACTIVE6–9 MIN 12 What is OpenClaw—and what gives it authority to act?Inspect the product boundary across channels, model, harness, tools, requester identity, and permissions. PRODUCTINTERACTIVE6–9 MIN