humans have much to learn from robotkind
18
38
12
251
it's all about the prelude before the conversation. You need to tell it what the AI is and is not capable. It's not trying to be right, it's trying to complete what it thinks the AI would do :)
21
107
62
820
I wonder if the AI would be better at math if you told it to show it's work
1
4
3
75
Replying to @QuantumSeany @nicklovescode @eigenrobot
Seems to work

Jul 17, 2020 · 10:15 AM UTC

12
55
31
348
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
Teaching GPT-3 to do a brute force 'for loop' checking answers also seems to work
8
42
13
233
Some rejected scenarios proposed by GPT-3 while trying to avoid the chore of math.
1
3
1
23
more replies
Replying to @kleptid
What was your original prompt for this? I'm having a lot of trouble getting anything coherent out.
1
10
KaryoKleptid @kleptid
20 Jul 2020
Replying to @senpagax @QuantumSeany @nicklovescode @eigenrobot
I tried to recreate it again to check how well it replicates. General context prompt ends in the line 'You say "So, f(x)=x*x".'. You might need to tweak the temperature to make the math more consistent.
4
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
Wow...which part is the prompt here, please?
2
KaryoKleptid @kleptid
20 Jul 2020
Replying to @senpagax @QuantumSeany @nicklovescode @eigenrobot
I tried to recreate it again to check how well it replicates. General context prompt ends in the line 'You say "So, f(x)=x*x".'. You might need to tweak the temperature to make the math more consistent.
1
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
@nachoarranz estuviste siguiendo El Progreso de estos loritos????? 🦜🤯🤯🤯
1
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
Did you intentionally train GPT3 on Spice and Wolf
1
2
15
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
That's fuckin rad
1
Replying to @kleptid @QuantumSeany @nicklovescode @eigenrobot
I think it's getting around a limit on algorithmic depth (how many operations the net can learn to do sequentially) by storing partial results in its output. Which is obviously easier, and probably something we should build into the architecture.
6
Load more