[georgelyon]

Am I correct in understanding that this doesn't actually generate answers based on HN, but instead finds semantically-near comments and sends them verbatim to GPT to summarize? Seems like a good enough hack, though I'd love a detailed writeup of how to actually specialize an existing LLM with additional training data (like HN).

[ilaksh]

Technically it does give a specific answer to the question, but it is based on the semantically similar comments (and the question).

The thing people don't realize is that right now there is a very large gap between the capabilities of a few models including OpenAI's most recent ones, and most of the other LLMs. So there are several options for actually training or fine-tuning with open models, but actually none of them have the language understanding and generation capabilities at the level of those new OpenAI models.

As far as I know.

[jerpint]

The summary itself is still generated, but has all the context to do summarization in the prompt.

It's very difficult to otherwise finetune existing LLMs. GPT itself is closed-sourced, and doesn't allow for finetuning (except via an opaque API and with limited amounts of data). Other open models are either very difficult to load in memory and/or simply not as expressive as GPT

[serjester]

Agreed, I think the better approach is to do some custom tuning but that becomes cost prohibitive very quickly. Not really much different than Algolia with a minor GPT-3 integration but neat project regardless.

[redox99]

You can literally finetune these OpenAI models using their API. In this case it probably wasn't done because the author found it too much work and/or too expensive.

[moremetadata]

>the author found it too much work and/or too expensive.

Military budgets and manpower would make sock puppets interesting. https://www.theguardian.com/technology/2011/mar/17/us-spy-op...

Does anyone find it curious Edward Snowden is mute on the military's use of this technology?

How many people would be surprised to learn they are communicating with a program on websites like this?