Hacker News Reader

Comment by belter | Hacker News Reader

I confirm GPT-4 solves this correctly. Makes me immediately doubt everything else in the article...

When calling the GPT-4 API I get this answer:

Sally has 2 sisters. Each brother counts Sally and two other girls as sisters, which means there are three girls in total. Since one of these girls is Sally herself, that leaves two others as her sisters.

When I said 'nope', it corrected itself and gave the right answer.

Have you tried replicating via the API with a temp of 0?

No I did not.

Using a temp of zero usually returns garbage results from most models, so it would likely do so in case of GPT 4 as well. Any other great ideas?

What do you mean? It works fine for me when I’ve tried it

Temp of 0 gives the least random and most predictable results

That's true, but those results are rarely the correct ones, at least for v1 llama models. In my experience each model has an optimal temperature at which it performs vastly better. I'm sure OpenAI have the best config they know set up for ChatGPT but let people generate trash through the API if they want to waste their credits on it.

Why would the accuracy decrease with lower temperature? Setting temperature to 0 just means at each step the model will emit the token with the highest likelihood.

Yes that's what I'm saying, to reiterate: The likeliest token does not lead to the highest performing result. Otherwise temperature wouldn't even be an option. I would imagine things like language word frequency affect the token rating a lot while having nothing to do with the task at hand except providing a correctly formatted answer, but it's probably not the whole story.

OpenAI (and others that know what they're doing) always do their benchmarks in a multi-sampled way, by running 5 or 20 times at optimal temp. Using a wrapper that runs these samples and then another pass that judges self-consistency for a final answer can give you a correct answer 100% of the time for a question that would be wrong 100% of the time with temp at zero.

I had a conversation with a friend regarding this exact question and my understanding is that model trains to optimize the distribution of all texts, therefore when you restrict it to deterministic sampling that is not representative of inputs you select the slice of the distribution that model learned that conveys much less information than the full distribution, and hence has poorer results.

The point isn't that temp 0 should be used, the point is that anyone surprised that they get different results should realise that there is an element of randomness involved by default.

Even repeating the same question in a single chat can have GPT-4 vary on its output, though it will often settle on a particular output due to context informing the output (which is why adding context is so important for these models)

Not in my experience, in fact I find that when I need precise, realistic, and reliable results temp 0 is needed. For example, here is a bunch of names, gather the names of specific plastics under headings matching their common acronym - if I don't use temp 0 I might get nonsense out. Temp 0? reliably correct.

Interesting, that's the exact opposite of my experience.

Reply on news.ycombinator.com