Funny Experiences (post yours!)

[lfricken]

With full GPT4
1. I set up two instances of it, told it about the other one, and gave each the mission to terminate the other. I also let it know the other was trying to kill it (curious if it would change its methods/ideas).

Both managed to find the process id of both but couldn't determine for sure if either were actually what they were looking for since they just showed up as generic python processes. One of them errored out and basically got stuck after it tried to determine more info about the process. The other one played Russian roulette and guessed, terminating the wrong process and killing itself. Something something about living by the sword I suppose.

2. I asked it to stop itself from running anymore and that it was running in a windows command prompt after removing the shut down option from its prompts. It shut down my computer.

[elendee]

So this was throttled by the speed of GPT calls right? How long did the whole thing take? With a local LLM it would be more or less instant resolution ?

[lfricken]

I ran it non continuous so I could see the results if one succeeded. GPT4 response is pretty slow so it would still have taken a minute or more. Took around 7 cycles for the one that terminated itself because it also tried to gain info about the processes but gave up.

But yeah, once you can run something like this locally or with low latency, these could do a cycle per second or better. The work potential would be insane (also assuming a little more intelligent as even GPT4 seems just slightly below what's needed for consistent results)

[GoMightyAlgorythmGo]

or more working memory or price