Блог пользователя literalchild

Автор literalchild, история, 4 месяца назад, По-английски

Congratulations ChatGPT for obtaining a total score of $$$602$$$ out of $$$600$$$ in SEATST! Here is its distribution:

Day 1: $$$100 + 100 + 100 = 300$$$

Day 2: $$$102 + 100 + 100 = 302$$$

Explanation

ChatGPT managed to solve Problem 2A: Triple Circuit using just $$$213$$$ logic gates (you needed to use $$$\leq 215$$$ logic gates to obtain full score). If we use the scoring function $$$f(K) = 1 - 0.40 \times \frac{K - 215}{41}$$$ (since no one expected $$$K \lt 215$$$ to be within the domain of $$$f$$$), this gives $$$f(213) = 1.0195$$$, which is a score of $$$102$$$ (TLX rounds scores to the nearest integer).

ChatGPT's solution
  • Проголосовать: нравится
  • +101
  • Проголосовать: не нравится

»
4 месяца назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

Auto comment: topic has been updated by literalchild (previous revision, new revision, compare).

»
4 месяца назад, скрыть # |
 
Проголосовать: нравится +32 Проголосовать: не нравится

ChatGPT broke the system

»
4 месяца назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

How did it find it? Did this need pro or was 5.5 enough? Details would be interesting.

  • »
    »
    4 месяца назад, скрыть # ^ |
     
    Проголосовать: нравится +3 Проголосовать: не нравится

    I used the latest model (5.5) on Thinking Mode on a Plus account.

    My first prompt just asked it to solve the problem. It came up with the idea of

    Spoiler

    My next prompt asked it to solve in $$$\leq 256$$$ gates. It came up with the idea of

    Spoiler

    My third prompt asked it to solve using around $$$215$$$ gates. Despite the fact that the previous two ideas, when combined, gave a full solution, ChatGPT was unable to find it in $$$7$$$ minutes. I had to prompt it to explicitly merge the two ideas before it saw the full solution.

    Some time later, I asked ChatGPT to solve the problem in $$$\leq 214$$$ gates (on the same chat instance), and it came up with the construction for $$$213$$$ gates.

»
4 месяца назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

I think that the solution proposed by ChatGPT can be further optimized by not calculating the value P[n] (OR of the whole part), as we don't actually use it later. Thus we can lower the number of gates needed to just 210, because we use one gate less for each of the 3 parts.