Hacker News Reader

I get the sense that any accurate portrayal of the model is getting downvoted. Sure, it's going to be a while before open-source models catch up with OpenAI. I'm glad that Stability is working on this, and I look forward to trying out their larger models that they're training right now, but we should be able to discuss accurately the capabilities of an LLM without just automatically jumping on the hype train.

View on HN

That's fair, but this:

> Is this a joke? I tried the HuggingFace demo, and it just spews out ridiculous replies for everything.

Is a shallow and uninformed criticism. The parameter count alone should inform people that this isn't going to match GPT-3.5 or GPT-4. Not everyone knows that of course, but starting a comment with "Is this a joke?" is precisely the wrong way to start such a discussion.

Yes, go try asking GPT-3 Babbage these same questions and let me know if it is even remotely coherent! That is the fair comparison.

I dunno if OP tried raw 3B, but finetuned llama 7B is much better than that.

Reply on news.ycombinator.com