|
89
|
1 Points to make: BCS observation
|
|
|
2
|
|
|
3 Phase 1: 1960--1970s: AI -> CogSci trajectory from basing AI on what's
|
|
|
4 understood in Psych/Ling/Anthro/Socio etc. and developing
|
|
|
5 computational implementable approaches to understanding e.g. human
|
|
|
6 language. HMMs
|
|
|
7
|
|
|
8 Phase 2, Late 1970s-early 20th: Evidence-based, but with relatively
|
|
|
9 little evidence by today's standards: a few millions of words of
|
|
|
10 written (American) English, a few 10s of hours of spoken, simple
|
|
|
11 probabilistic models, _no_ reference to human process. ANNs.
|
|
|
12
|
|
|
13 Phase 3, 2010s to 2023: Step-change in the size (of data, of models,
|
|
|
14 of compute power) and in the complexity of the architecture (LLMs) and
|
|
|
15 the underlying mathematics.
|
|
|
16
|
|
|
17 Apothesis: Release of ChatGPT in April 2023 (?).
|
|
|
18
|
|
|
19 But one thing hasn't changed: it's all about the _form_ of the data,
|
|
|
20 not the substance (BCS).
|