How does one sentence become a real action?
A model can propose an action. Software must decide whether to execute it. Enter the gallery and follow one request all the way from words to a visible result.
REQUEST / MODEL / TOOL / RESULT English collection All twelve interactive exhibits are open. The complete route spans model generation, agent systems, compute infrastructure, product anatomy, and claim reading.
View collection statusOBJECT ROOM / 1943—2022
Do not just explain history.
Put its records on view.
19 objects—from a 1943 formal neuron model to a 2022 release record—separate feedback machines, physical learners, planning systems, encoded rules, learned representations, specialised computation, data infrastructure, architecture, and a public interface.
COMPARISON ROOM / 1984—2022
The result looks intelligent.
Where did the work happen?
Put four records side by side: specified rules, learned weights, an attention architecture, and a public interface. Read each through the same four lenses without duplicating its archive record.
Enter the comparison roomEVIDENCE NETWORK / 6 OBJECTS · 7 PATHS
Evidence is not a badge.
Follow the line to the record.
Six existing objects connect to seven approved source records. Every line means one documented source path; thin evidence stays visible instead of becoming a confidence score.
Enter the evidence networkNOW ROOM / CHECKED 2026-09-11
What changed this month—and who says so?
Five dated records separate publisher statements from what the page itself establishes and what the museum infers. The newest follows an AI agent entering a physical-lab calibration loop, with the case study’s limits kept attached.
Enter the Now roomALSO SHOWING / EXHIBIT 02
LANGUAGE MODELS · INTERACTIVE · 4–7 MIN Watch an answer emerge one token at a time. Step through a generation loop, then move one control and watch a probability distribution change in your browser. ↗NEW WING / AGENT SYSTEMS
One word hides three layers.
Follow one connected route through the agent that pursues a task, the harness that controls each turn, and the protocol that connects outside capabilities.
COMPLETE COLLECTION / FIELD GUIDES 09—12
Four terms people use too quickly.
Open each one and operate the boundary: classify an AGI claim, diagnose an operator path, read a CANN port, and constrain an OpenClaw action.
- 09 / CLAIMAGIBreadth, performance, and autonomy are not one score.
- 10 / COMPUTEOperatorAn operation, dispatch route, kernel, and benchmark are different findings.
- 11 / STACKCANNGraph, operator, runtime, compiler, communication, and versions.
- 12 / PRODUCTOpenClawThe model proposes; the configured product holds the authority.
THE CENTRAL DISTINCTION
The model writes.
The program acts.
“AI used a tool” compresses several different events into one sentence. This museum separates them: what you asked, what the model produced, what software accepted, and what code actually changed.
See the four-part signal pathHOW TO VISIT
Twelve exhibits.
Three distances.
- 01Understand
Get the distinction in thirty seconds, without a wall of terms.
ABOUT 30 SECONDS - 02Observe
Keep the interface and the hidden process visible at the same time.
ABOUT 3 MINUTES - 03Test
Run both conditions and compare what truly executed with what was only scripted.
ABOUT 8 MINUTES