Comment by Tiberium | Hacker News Reader

Comment by Tiberium | Hacker News Reader

Original: If anyone's curious about the (probable) non-humorous explanation: I believe this is because they set the frequency/presence penalty too high for the requests made by ChatGPT to the backend models. If you try to raise those parameters via the API, you'll have the models behave in the same way.

It's documented pretty well - https://platform.openai.com/docs/guides/text-generation/freq...

OpenAI API basically has 4 parameters that primarily influence the generations - temperature, top_p, frequency_penalty, presence_penalty (https://platform.openai.com/docs/api-reference/chat/create)

UPD: I think I'm wrong, and it's probably just a high temperature issue - not related to penalties.

Here is a comparison with temperature. gpt-4-0125-preview with temp = 0.

- User: Write a fictional HN comment about implementing printing support for NES.

- Model: https://i.imgur.com/0EiE2D8.png (raw text https://paste.debian.net/plain/1308050)

And then I ran it with temperature = 1.3 - https://i.imgur.com/pbw7n9N.png (raw text https://dpaste.org/fhD5T/raw)

The last paragraph is especially good:

> Anyway, landblasting eclecticism like this only presses forth the murky cloud, promising rain that’ll germinate more of these wonderfully unsuspected hackeries in the fertile lands of vintage development forums. I'm watching this space closely, and hell, I probably need to look into acquiring a compatible printer now!

I don't think it's a temperature issue because everything except the words is still coherent. It's kept the overall document structure and even the right grammar. Usually bad LLM sampling falls into an infinite loop too, though that was reported here.

Always needs oneself a good eldritchEnumerator! Sorry, gotta go feed the corpses, sorry corpuses for future scraping.

Correct me if Im wrong: Temperature is the rand function that prevents the whole system from being a regular deterministic program?

Close. Temperature is the coefficient of a term in a formula that adjusts how likely the system is to pick a next token (word/subword) which it thinks isn't as likely to happen next as the top choice.

When temperature is 0, the effect is that it always just picks the most likely one. As temperature increases it "takes more chances" on tokens which it deems not as fitting. There's no takesies backies with autoregressive models though so once it picks a token it has to run with it to complete the rest of the text; if temperature is too high, you get tokens that derail the train of thought and as you increase it further, it just turns into nonsense (the probability of tokens which don't fit the context approximates the probability of tokens that do and you're essentially just picking at random).

Other parameters like top p and top k affect which tokens are considered at all for sampling and can help control the runaway effect. For instance there's a higher chance of staying cohesive if you use a high temperature but consider only the 40 tokens which had the highest probability of appearing in the first place (top k=40).

> There's no takesies backies with autoregressive models

Doesn’t ChatGPT use beam search?

Almost certainly not.

It's absolutely just sampling with temperature or top_p/k, etc. Beam searches would be very expensive, I can't see them doing that for chatgpt which appears to be their "consumer product" and often has lower quality results compared to the api.

The old legacy had a "best_of" option but that doesn't exist in the new api.

Pretty much.

The model outputs a number for each possible token, but rather than just picking the token with the biggest number, each number x is fed to exp(x/T) and then the resulting values are treated as proportional to probabilities. A random token is then chosen according to said probabilities.

In the limit of T going to 0, this corresponds to always choosing the token for which the model output the largest value (making the output deterministic). In the limit of T going to infinity, it corresponds to each token being equally likely to be chosen, which would be gibberish.

https://medium.com/mlearning-ai/softmax-temperature-5492e400...

Azure OpenAI seemed to have temperature problems before, i.e. temp > 1 led to garbage, at 2 it was producing random words in random character encodings, at 0.01 it was producing what OpenAI's model was producing at 0.5 etc. Perhaps they took the Azure's approach ;-)

That might explain why I found GPT4 via Azure a bit useless unless I turned the temperature down…

Last time I tried a temp above 1, I almost instantly got gibberish. Pretty reliable parameter if you want to make the transformer output unusable.

Landblasting eclecticism is always worthy of pressing forth the murky cloud.

wow this really makes me think the temperature on my brain is set higher than other sapients

The Murky Cloud sounds like a great sarcastic report on how cloud things explode in the style of the old Register.

Reply on news.ycombinator.com