3 ms·

https://phind.com https://phind.com (YC S22) uses LLMs grounded in internet contexts to answer questions. There's an interesting tradeoff between grounding LLM

https://phind.com https://phind.com (YC S22) uses LLMs grounded in internet contexts to answer questions. There's an interesting tradeoff between grounding LLM answers in web contexts and answering questions from parametric memory like ChatGPT. The former is more accurate and verifiable, but the latter can be more creative and specific to you. We will have both of those answers soon. (Disclaimer: I am the co-founder).

mark_l_watson 4y ago
I had not heard of phind.com either. I asked the same detailed question about myself to both perplexity.ai and phone.com, and I really liked both answers with references. Both are likely more useful to me, right now, than Google search. That said, the technique of a accepting a question, get content text with web searches, and then use LLMs like squad, etc. to answer questions - this is something the Google and Bing search will soon do also, with many of their improvements. As a slightly disinterested third party, it will be fun the watch the competition.
throwaway1851 4y ago
Hi. I remember reading a question answering paper from Google in 2021 that concluded there was little evidence that models were actually grounding the generated answers in the contexts. Do you know if this problem has been solved?
- rushingcreek 4y ago
  Ah yes, it's this paper: https://arxiv.org/abs/2103.06332 https://arxiv.org/abs/2103.06332. With recent advances in alignment, this problem has gotten much better. Furthermore, the LLM used in that paper is only ~400 million parameters, which is orders of magnitude smaller than models being used for this task today. Scale + alignment has been the solution so far.
  - throwaway1851 4y ago
    Thanks for the reply! It’s interesting (counterintuitive?) that the problem of models drawing too much from their stored knowledge is mitigated by using models with much greater stored knowledge.

mark_l_watson 4y ago

I had not heard of phind.com either. I asked the same detailed question about myself to both perplexity.ai and phone.com, and I really liked both answers with references. Both are likely more useful to me, right now, than Google search. That said, the technique of a accepting a question, get content text with web searches, and then use LLMs like squad, etc. to answer questions - this is something the Google and Bing search will soon do also, with many of their improvements. As a slightly disinterested third party, it will be fun the watch the competition.

throwaway1851 4y ago

Hi. I remember reading a question answering paper from Google in 2021 that concluded there was little evidence that models were actually grounding the generated answers in the contexts. Do you know if this problem has been solved?

rushingcreek 4y ago
Ah yes, it's this paper: https://arxiv.org/abs/2103.06332 https://arxiv.org/abs/2103.06332. With recent advances in alignment, this problem has gotten much better. Furthermore, the LLM used in that paper is only ~400 million parameters, which is orders of magnitude smaller than models being used for this task today. Scale + alignment has been the solution so far.
- throwaway1851 4y ago
  Thanks for the reply! It’s interesting (counterintuitive?) that the problem of models drawing too much from their stored knowledge is mitigated by using models with much greater stored knowledge.