Hacker News Reader

It just hasn't been prompted or fine-tuned to have the neutral, self effacing personality of ChatGPT.

It's doing the pure, "try to guess the most likely next token" task on which they were both trained (https://heartbeat.comet.ml/causal-language-modeling-with-gpt...).

ChatGPT is further trained with reinforcement from human feedback to make them more tool-like (https://arxiv.org/abs/2204.05862 & https://openai.com/blog/chatgpt & https://arxiv.org/abs/2203.02155),

with a bit of randomness added for variety's sake (https://huggingface.co/blo1g/how-to-generate).

That bit about "fine-tuned on the Alpaca dataset" is precisely about that. But, yeah, no RLHF so far, although some people are already working on that.

The randomness link doesn’t work. All the ChatGPT products need randomness or it can’t be creative.

https://huggingface.co/blog/how-to-generate

It doesn't have an identity, so it just creates something random. It is irrelevant for many use cases I would guess.

Hmm I'm not sure here. I asked "Who are you?" and first I got this response.

  LLaMa didn’t return anything, please try another prompt. It 
  responds best to questions.

Then when I asked again I got the same response as the parent comment, word for word.

  I am a 25-year-old woman from the United States. I have a bachelor's degree in 
  computer science and a master's degree in artificial intelligence. I am 
  currently working as a software engineer at a tech company.

Me: who are you?

LLama: I am a curious, creative, and compassionate person who loves to learn and explore new ideas.

Me: answer that again

LLama: Again?

Me: yes, why not?

LLama: Why not?

Me: why are you parroting me?

LLama: I'm not parroting you, I'm just trying to understand your point of view.

Me: then just answer that again

LLama: Then just answer that again.

Me: Uhm, you are parroting me. Is this some kind of joke you're playing?

LLama: No, it's not a joke. I'm just repeating what you said.

Me: Why?

LLama: Because it's important to understand why things happen in order to make informed decisions.

Me: What the hell, I just want you to answer my question again. That's all. Can you please do that?

LLama: Yes, I can answer your question again.

Me: Do it.

Llama: Done.

Me: I don't see the answer.

LLama: I don't see the answer.

Me: Can you remember what question I asked you to answer?

LLama: Yes.

Me: What was it?

LLama: It was a bird.

This is the first time I had a good laugh while interacting with LLMs.

I like Mrs. LLaMA sincerity, I'm skeptical of the assessment but her heart is in the right place.

This is just fantastic - this is how all these models should work, and how they should be expected to work IMHO. I don't want someone programming it to act like a real human who works at facebook. I want a language model that can assume any identity I want.

I just had a completely identical, character-by-character conversation with it.

Shouldn't there be at least a very little bit of randomness?

Per the code linked there is

temperature (optional, default=0.1): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.

top_p (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.

top_k (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.

Potentially they set temperature to 0? Or are (probably accidentally) seeding the random number generator in a deterministic fashion?

That is hilarious, considering that LLaMA was developed by Meta.

So we can deduce the training data is from before January ‘23.

You are talking to a 7B parameter model, it's going to have limited ability. Llama 65b is considerably better.

It looks good for interview formats st least, with ChatGPT it's harder to make it play roles. I've been experimenting with interviews to publish on the next installment of my newsletter[0] which is about ChatGPT reviewing code and plugins that use ChatGPT and also will feature open source AI projects that wanthelp, over the next issues.

[0]: https://generativereview.substack.com/p/the-generative-revie...

> Maybe there is a long path to improvements on LLaMA

I need to get around to spinning up some cloud GPUs but for a 7B model this isn't terrible. I'd guess there's a big jump when using the really big model variants. Would love to hear from folks who have tried the bigger models.

The model was just instruction tuned. That is, it can answer questions and respond to instructions. Unless especially prompted, a pure predictor would often not respond to instructions but elaborate on them. To get specific types of responses, you would need something like RLHF.