4 ms·

> Can't cite their sources? Has there been research on neural nets who could cite their sources? In common cases, thousands or millions of items in the trainin

by sampo 4y ago

> Can't cite their sources? Has there been research on neural nets who could cite their sources? In common cases, thousands or millions of items in the training set have contributed to the weights in the pathways that activate for a certain input. But maybe for some inputs, only a small number of items in the training set contributed strongly to the weights on the pathways the are decisive to the output? Could you design a data structure that would keep this information in the neural net?

jesterman 4y ago
perplexity.ai/ specifically uses web sources to construct natural language. I believe it's going to be in Bing soon?
- mynameisvlad 4y ago
  It still has the exact same issues. I asked it to tell me about “Microsoft layoffs before 2022” and it confidently stated that Microsoft had no layoffs before 2022, with one of its sources even talking about the previous layoffs in 2009.
- sampo 4y ago
  > perplexity.ai/ specifically uses web sources to construct natural language. I am going to guess that even though it uses the GPT language model, it works fundamentally differently than ChatGPT. My guess is, Perplexity AI is a traditional search engine, that first finds a list of links that match your questions. And then it feeds text from those links to the GPT language model to write a summary of the results. Whereas ChatGPT produces text based only on the trained deep neural net. Maybe I am guessing wrong how Perplexity AI works. Maybe it works in a different manner.
etrautmann 4y ago
I would imagine that citing sources would look more like finding sources that have a minimum distance to the output in a latent space representation? I don’t think it’ll make sense to track how each connection weight changed over time due to various inputs.
yellow_postit 4y ago
I’d have thought the short lived Galatica from Meta would have demonstrated more capability on this need given it was explicitly positioned as being to help Researchers. I wonder if this is a special case for Explainable Models [1] — once you select an answer, another model helps describe and source the output. [1] https://en.m.wikipedia.org/wiki/Explainable_artificial_intelligence https://en.m.wikipedia.org/wiki/Explainable_artificial_intel...

by sampo 4y ago

jesterman 4y ago

perplexity.ai/ specifically uses web sources to construct natural language. I believe it's going to be in Bing soon?

mynameisvlad 4y ago
It still has the exact same issues. I asked it to tell me about “Microsoft layoffs before 2022” and it confidently stated that Microsoft had no layoffs before 2022, with one of its sources even talking about the previous layoffs in 2009.
sampo 4y ago
> perplexity.ai/ specifically uses web sources to construct natural language. I am going to guess that even though it uses the GPT language model, it works fundamentally differently than ChatGPT. My guess is, Perplexity AI is a traditional search engine, that first finds a list of links that match your questions. And then it feeds text from those links to the GPT language model to write a summary of the results. Whereas ChatGPT produces text based only on the trained deep neural net. Maybe I am guessing wrong how Perplexity AI works. Maybe it works in a different manner.

etrautmann 4y ago

I would imagine that citing sources would look more like finding sources that have a minimum distance to the output in a latent space representation? I don’t think it’ll make sense to track how each connection weight changed over time due to various inputs.

yellow_postit 4y ago

I’d have thought the short lived Galatica from Meta would have demonstrated more capability on this need given it was explicitly positioned as being to help Researchers. I wonder if this is a special case for Explainable Models [1] — once you select an answer, another model helps describe and source the output. [1] https://en.m.wikipedia.org/wiki/Explainable_artificial_intelligence https://en.m.wikipedia.org/wiki/Explainable_artificial_intel...