intro

[timm]

how does chatGPT change things?

https://github.com/timm/tested/raw/main/etc/mov/chaptGPT_lua_kmeans.mov

Rahul's first rule: when it is wrong, it so confidently wrong that you cannot tell.

• wrong, on several points
  • security
  • hard-wired constants that should be adjustable
  • does not support symbolics
  • does not support streaming
  • does not support hierarchy
  • nothing about connecting clustering to sampling and optimization

398  {:Acc+ 15.6 :Lbs- 2970.4 :Mpg+ 23.8}
| 199
| | 99
| | | 49
| | | | 24  {:Acc+ 17.3 :Lbs- 2623.5 :Mpg+ 30.4}
| | | | 25  {:Acc+ 16.3 :Lbs- 2693.4 :Mpg+ 29.2}
| | | 50
| | | | 25  {:Acc+ 15.8 :Lbs- 2446.1 :Mpg+ 27.2}
| | | | 25  {:Acc+ 16.7 :Lbs- 2309.2 :Mpg+ 26.0}
| | 100
| | | 50
| | | | 25  {:Acc+ 16.2 :Lbs- 2362.5 :Mpg+ 32.0}
| | | | 25  {:Acc+ 16.4 :Lbs- 2184.1 :Mpg+ 34.8}
| | | 50
| | | | 25  {:Acc+ 16.2 :Lbs- 2185.8 :Mpg+ 29.6} <== best?
| | | | 25  {:Acc+ 16.3 :Lbs- 2179.4 :Mpg+ 26.4}
| 199
| | 99
| | | 49
| | | | 24  {:Acc+ 16.6 :Lbs- 2716.9 :Mpg+ 22.5}
| | | | 25  {:Acc+ 16.1 :Lbs- 3063.5 :Mpg+ 20.4}
| | | 50
| | | | 25  {:Acc+ 17.4 :Lbs- 3104.6 :Mpg+ 21.6}
| | | | 25  {:Acc+ 16.3 :Lbs- 3145.6 :Mpg+ 22.0}
| | 100
| | | 50
| | | | 25  {:Acc+ 12.4 :Lbs- 4320.5 :Mpg+ 12.4}
| | | | 25  {:Acc+ 11.3 :Lbs- 4194.2 :Mpg+ 12.8} <== worst
| | | 50
| | | | 25  {:Acc+ 13.7 :Lbs- 4143.1 :Mpg+ 18.0}
| | | | 25  {:Acc+ 14.4 :Lbs- 3830.2 :Mpg+ 16.4}

Now here's nearly the same algorithm, but know we run a greedy search over the splits. When splitting on two distance points A,B, we peek at the Y values and ignore the worse half.

398  {:Acc+ 15.6 :Lbs- 2970.4 :Mpg+ 23.8}
| 199
| | 100
| | | 50
| | | | 25  {:Acc+ 17.2 :Lbs- 2001.0 :Mpg+ 33.2}

Rahul's second rule: you need to know a lot about the code to to make best use of automatically generated code.

My main problem: cognitive fixation:

"If it was up to current thinking (in the 1920s) to cure polio. . . You'd have the best iron lung in the world but not a polio vaccine."