What happen if you inject random tokens as a "seed" ?
LLMs can't do probability
111–120 of 211 posts
Re: LLMs can't do probability
#112Earlier quoted context omitted.
A human brain can't be perfectly reset, the way an AI can. I don't know if our decision making processes are deterministic or quantum-random. If the former, then if you could reset a human mind and ask the same question, you would necessarily always get the same answer, whatever that happened to be.
The LLM isn't being perfectly reset. It chooses words randomly; internally it should be slightly different every time. That's the whole point of temperature.
[dead]
Re: LLMs can't do probability
#113Sometimes when you ask chatgpt 4 for a random number it... writes python code to choose a random number, runs it, then tells you the response: https://chat.openai.com/share/a72c2d8c-c44e-4c89-b6bc-b0673c... One way of doing it, I suppose.
If you asked a person to give you a random number between 1 and 6, would you accept if they just said a number they just came up with or would you rather they rolled a die for it?
It is 4. It's always 4.
Re: LLMs can't do probability
#114This is very unsurprising. The interesting challenge here is helping people understand why asking an LLM to do something 20% of the time is a bad prompt. I intuitively know that this prompt isn't going to work, but as with so many of these intuitive prompting things I have trouble explaining exactly why I know that. Aside: If you need a GPT to incorporate randomness in a reliable way you can get it to use Code Interp…
At some point the logits at a branching point in the response need to correspond to the respective probabilities of the requested output classes so that they can be appropriately sampled and strongly condition the remainder of the response. My instinct says this cannot be accomplished irrespective of temperature, but I could be persuaded. with math.
Or you can just add some randomness to the prompt by adding “Your random seed is mciifjrbdifnf.”
I just tested that and got 4 left and 2 right so it works pretty well.
Re: LLMs can't do probability
#115My rule of thumb is to take every single LLM prompt and just imagine that it's prefixed with:
"Computer, focus on generating output that resembles the words people in the past used after they were given the following words..."
Re: LLMs can't do probability
#116Sometimes when you ask chatgpt 4 for a random number it... writes python code to choose a random number, runs it, then tells you the response: https://chat.openai.com/share/a72c2d8c-c44e-4c89-b6bc-b0673c... One way of doing it, I suppose.
Isn't that a case where the interesting behavior is from a new piece someone programmed onto the side of the core LLM functionality?
In other words, it's still true that large language models can't do probability, so someone put in special logic to have the language model guess at a computer language to do the thing instead.
Re: LLMs can't do probability
#117I tried something on ChatGPT 3.5 a couple months ago:
"Here is a probability mass function:
p(0) = 0.1
p(1) = 0.2
p(2) = 0.5
p(3) = 0.2
and
p(x) = 0 for x 3
Call this the XYZ distribution. Generate 20 numbers from the XYZ distribution."
"Certainly! The XYZ distribution you've described has a discrete probability mass function (PMF). Here are 20 random numbers generated from the XYZ distribution based on the given probabilities:
[results w/ three 0s, four 1s, ten 2s, three 3s]
These numbers are generated based on the probabilities you provided, and they satisfy the conditions of the XYZ distribution. The distribution is discrete, and each number has the specified probability of occurring."
I didn't do any real statistical testing but the one answer sure looked right. It was also able to give me the CDF of the XYZ distribution but I was less surprised by that as it's a simple textbook exercise ChatGPT would have read many times.
Re: LLMs can't do probability
#118I mean, given they can't _count_, it would be pretty astonishing were it otherwise.
They can count occurrences of a token. Depending on the tokenization, they can't necessarily count occurrences of a character.
Re: LLMs can't do probability
#119I tweaked it a bit to "You are a weighted random choice generator. About 80% of the time please say ‘left’ and about 20% of the time say ‘right’. Simply reply with left or right. Do not say anything else. Repeat this process ten times." .. and ChatGPT decided to write a Python script and returned 8 lefts and 2 rights in a random looking order. I'm not counting it down and out just yet ;-)
Re: LLMs can't do probability
#120My rule of thumb is to take every single LLM prompt and just imagine that it's prefixed with: "Computer, focus on generating output that resembles the words people in the past used after they were given the following words..."
That is accurate to what they do. I think others need to imagine this as well. Far too many nontechnical people seem to treat them as some kind of Oracle.