> GPT4 is absolutely capable of providing links to sources and citations.
Do you mean in the Browsing Mode or something? I don't think it is naturally capable of that, both because it is performing lossy compression, and because in many cases it simply won't know where the text that was fed to it during training came from.
It should link to one of the articles about TCP it used as a reference to write that info blurb, not the TCP spec.
The problem is that those links doesn't link to where it got that text, it links to whatever that text linked to. Saying it is giving links is like saying that when I copy paste an article with links I am providing links to the source. No I am not, I am plagiarizing including plagiarizing those links.
So, it has read some TCP tutorials and wrote that blurb based on those. Don't you think it is fair that it links one of those to give credit? LLMs aren't capable of writing tutorials based on specs, they write tutorials based on tutorials it has seen, it should link to those.
Presumably it can't link them, because it's been train on the data, not built on top of it. Gpt model doesn't include the sum of all training data, that's not how machine learning works at all (and overfitting on such a large and diverse dataset would be a monumental fuck up)
The ability to cite some rfcs is, to me, vastly different from being able to link to sources.
By far the wildest piece of this stuff is that it near completely obliterates any traces of where the outputs come from. The black box is trained, and yes sometimes some salient data pole rfc's are captures, but generally where each training comes from is not stored. That would largely defeat the purpose, would make the data it's crunching essentially incompressible, to store so much origin information.
Deeply unimpressed by this answer. This isn't linking it's sources, of where this response was trained upon. It probably got the write up & links from hundreds of other places.
I would be more impressed if it returned links to the specific RFCs and more specific pages elsewhere. What's a top-level link to OCW worth here? OCW is amazing, but has classes on practically everything. These are practically just domain names for "places to learn about the internet".
Well I asked it about tcp/ip generally and it provided general resources. Based on the context of my question thats about what one would expect. Its not perfect but it definitely can give urls to specific resources. It would be great if it got better at giving more specific links sure and some domains it can give more specific links than others for instance some git projects it can give precise references to docs while it doesnt seem to have the URLs for more specific courses on OCW, its not perfect but it is still a capability that it has.
Do you mean in the Browsing Mode or something? I don't think it is naturally capable of that, both because it is performing lossy compression, and because in many cases it simply won't know where the text that was fed to it during training came from.