Hacker News Reader

Comment by robotresearcher | Hacker News Reader

Where did you learn that?

He is almost certainly referring to factual type queries of the sorts including 'how to do x in y.'

About the zillionth time you get 'I'm sorry, you're right, [blah] doesn't exist. Here's [something else that doesn't exist]', it gets really frustrating. The worst part is you can often tell the software is referencing some relevant page(s), but just inappropriately mixing them with other stuff. And so if you could simply get the link it'd be far more helpful than listening to the program continuing to describe in immaculate detail how to use an API that does exactly what you're looking for, with the slight problem that it doesn't exist.

Bing chat has references to relevant pages

I believe these are added after the response is generated. I.e the response is used as the query and the results are returned.

I've found that they're often not what they're claimed to be or not relevant to what Bing is saying.

Oh, that’s easy, most of my opinions are regurgitations of other comments I’ve read, or headlines I’ve skimmed.

In other words, you learned it “somewhere”. Not sure that is a better answer than what AI can currently give.

So I guess that means that we shouldn't expect LLMs to be any more reliable than some random guy who "heard it on the internet".

It is frustrating to see this idea that human learning and machine learning are essentially the same repeated uncritically.

Humans can't possibly remember provenance of all information. Machines possibly could. There's a significant difference in capabilities and it would be unwise to ignore this.

I don't know that they could, well not LLMs, as they aren't picking from linked data just from the subsequent model generated by combining information derived from that data. I suspect a map of all those links would be bigger than the input data.

They could have 10,000 occurances that told them to use the word dog in a response to the question "what is man's best friend?"

Also, the answer to where did they learn which town to use for "Where and when were grapes introduced into Australia?" seems to be "Actually, I didn't know, I just picked from a list of Australian towns and made the factual link up"

> Humans can't possibly remember provenance of all information. Machines possibly could

Machines can already do that. We have many, many memory mechanism for llm these days. There are already search engine like sites like phind.com which are rigorous about the sources. Langchain has tools to retrieve and cite memory from doc stores, APIs etc. If I ask my langchain agent "where did you learn that" or "why are you saying this" it does a decent job of citing the source.

Now, if your objection is that these are not perfect, then I would like to welcome you to the rising part of the sigmoid function of progress.

Access to openAI API directly gives better explanations to logic that was applied too.

I do not understand this god of the gaps type debate that is going on regarding LLMs. Everyone just seems to be interested in pointing out flaws. They are all using the "I'm witholding my judgement words" but are in the "this is just garbage" tone.

I did not write "fundamentally defective" -- I'm not OP.

My point is merely that we should have high requirements and not excuse flaws in language models just because humans have them. That thinking is akin to apologetics and not constructive.

Thinking that LLMs should behave like humans is flawed logic in my view. Instead of pointing out what we can do with them, everyone is busy talking about how is different from humans. Would you be calling the C language faulty because you can't express fft in a single line like MATLAB? No, because we look at C for what it is, we work with its advantages.

The phind.com thing is an improvement. The same goes for perplexity.ai.

However, for the few queries I made, there are still all sorts of issues involved, such as how the reliability of the sources are determined. For instance, I asked phind.com to give some information with references to arXiv, and it concluded that arxiv-vanity.com was the proper source ("rigorous about the sources" indeed...). Then I tried a few queries about jurisprudence and without hesitation it went to reference questionable commercial sites instead of the primary sources. Furthermore, it seems that phind.com is quilty of 1:1 plagiarism in many cases.

My god man, did you not read my full comment?

> Now, if your objection is that these are not perfect, then I would like to welcome you to the rising part of the sigmoid function of progress.

Edit: Also phind.com does better with coding docs. I think it is focused on code so asking is questions outside of that domain is kinda unfair.

I read it. And as I wrote, things are improving so it seems some things discussed are taken into account. But regarding the general criticism of many bystanders, including journalists, it seems that the industry throughout the world is ruthlessly rushing half-baked things prematurely due to the usual hype/VC/first-mover/monetization/etc. reasons without any considerations of the implications.

I dont know what implications you are talking about, but the mode of operation you described has been how start up are created for ever no?

The sum of information I take in is way smaller (by a magnitude 8 or 9) than what the AI takes in. Partly there are clear trails of what I learn (browser history), and also partly my learnings come from the circles I am in - colleagues, friends, education... I can't regurgitate the whole content like a machine, but can constrain the options strongly and also most often say "I am 100% confident this is true, this probably sure but not fully, and this I don't know".

In the case of the GP, checking where something has been learned or inferred by AI is something I learned from reading articles as well as participating in discussions with colleagues over AI explainability/transparency, as well as reading articles and listening to podcasts over journalism and truth in the age of AI. In fact I know what my 3 to 5 biggest influences/sources on that topic are. They might not be the one who invented the topic and answer, but I know who passed it to me.

Unlike an LLM you have been in a sensorimotor loop with the world for a couple of decades at least. The effective bitrate of your senses is probably pretty high. You’ve seen a thing or two, even if you haven’t read the entire web. You’ve seen and participated in causal processes over and over.

I know the question is meant to be flippant, but are you not aware of where you learned most things? For any of my post-1st-grade vocabulary, I remember the first time I saw the word in print, or where I heard it. I'd qualify that to say for most non-daily words, except that I don't believe there were many of those for me after I started school.

I am absolutely not aware of where I learned most things.

I acknowledge you as my superior.

Not flippant at all. I’m interested in the answer, and whether the answer supports the strong claim or not.

Do you know where you learned your opinions on a novel question? How?

There's a difference between accumulating raw knowledge and deriving or performing logical inference to reach conclusions. I can tell you where I learned many things, cite papers off the top of my head or at least retrieve them, but I can't tell you about knowledge that I derived / synthesized on my own.

So your trains of thought are ephemeral too?

It's more that chains of trains of thought are used to derive knowledge, and said chains can be distributed across space, and across time, so you can't necessarily attribute a single place.

If this is a normal capability of most people, I am horribly broken.

What you describe is probably uncommon.

I'm sure there are more robust neuroscience papers, but as an example: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4409058/

Your average human does a lot of optimisation to storage things (sight sound smell, etc) in the old wet ware, and retrieval can be hit and miss.

Reply on news.ycombinator.com