IN FACT MUSEUM 中文

COMPARISON ROOM / FOUR WORKING ARRANGEMENTS

Intelligence is not one kind of work.

Four existing objects enter a new comparison. The facts stay in their archive records; this room changes the questions placed beside them.

04records compared
04fixed reading lenses
1984—2022record dates

ONE QUESTION / FOUR RECORDS

When a system appears intelligent, where did people specify the work—and what did the machine actually change or execute?

Read every object through the same four lenses. This is a comparison of working arrangements, not a capability ranking and not a claim that one record inevitably led to the next.

  1. 01 SPECIFY
  2. 02 ADJUST
  3. 03 ARRANGE
  4. 04 RELEASE
  1. 01 SPECIFY

    ARCHIVE OBJECT / AIM-BOOK-1984-001

    Rule-Based Expert Systems: The MYCIN Experiments

    Expertise becomes explicit rules, while the surrounding human work remains visible.
    01What people specify
    Researchers acquire domain knowledge, write rules, and design how uncertainty, explanation, evaluation, and human use are handled.
    02What the system changes or runs
    The program applies encoded rules to a case and produces conclusions that its explanation machinery can expose.
    03What this record lets us inspect
    A retrospective research volume separates the rule engine from the large body of work required around it.
    04What this does not prove
    This research record is neither medical guidance nor a certificate that the system was ready for general deployment.
  2. 02 ADJUST

    ARCHIVE OBJECT / AIM-DOC-1986-001

    Learning representations by back-propagating errors

    Useful representations are acquired by changing weights, rather than written out one rule at a time.
    01What people specify
    Researchers choose a network structure, training cases, and desired outputs against which error can be measured.
    02What the system changes or runs
    The procedure repeatedly changes connection weights to reduce the gap between actual and desired outputs.
    03What this record lets us inspect
    A concise 1986 account states what is adjusted, which difference is reduced, and what hidden units can acquire.
    04What this does not prove
    This one formulation does not show that every later learning system uses the same procedure or that every ingredient began here.
  3. 03 ARRANGE

    ARCHIVE OBJECT / AIM-DOC-2017-001

    Attention Is All You Need

    The distinctive object on view is an architecture: a way to organise learned computation.
    01What people specify
    Researchers specify an encoder–decoder design and make attention, rather than recurrence or convolution, its organising choice.
    02What the system changes or runs
    Within the proposed architecture, learned parameters transform and relate representations without recurrent or convolutional layers.
    03What this record lets us inspect
    The architecture section and Figure 1 provide a compact technical record of the proposed arrangement.
    04What this does not prove
    The paper does not contain every training method, safety layer, product decision, or capability later associated with Transformer systems.
  4. 04 RELEASE

    ARCHIVE OBJECT / AIM-WEB-2022-001

    Introducing ChatGPT

    A public interface turns model behaviour into an encounter that visitors can experience and interpret.
    01What people specify
    A product team shapes the dialogue interface, release framing, examples, stated limitations, and routes for feedback.
    02What the system changes or runs
    The deployed model produces conversational responses inside the interface and behavioural limits described by the release record.
    03What this record lets us inspect
    A dated launch document shows how a conversational research preview was presented to the public on release day.
    04What this does not prove
    A launch page records the 2022 preview; it does not establish how later versions behaved or independently validate every product claim.