If the hardware changes significantly and those sites don't exist in the future wouldn't that mean gemeni would degrade in quality because it has nothing to pull from?

▲

hydrogen7800 7 hours ago | parent | next [-]

Right, that success story is only because there was "organic" (for lack of a better term) information from an original source. What happens when all information is nth generation AI feedback with all links to the original source lost?

Edit: A question from AI/LLM ignorance- Can the source database for an LLM be one-way, in that it does not contain output from itself, or other LLMs? I can imagine a quarantined database used for specific applications that remains curated, but this seems impossible on the open internet.

	▲	bigthymer 7 hours ago \| parent \| next [-]
		> Can the source database for an LLM be one-way, in that it does not contain output from itself, or other LLMs? I think, for public internet data, we can only be reasonably confident for information before the big release of ChatGPT.
	▲	nsvd2 3 hours ago \| parent \| prev \| next [-]
		Yes, people have likened pre-LLM Internet content to low-background steel. If in the hypothetical future the continual learning problem gets solved, the AI could just learn from the real world instead of publications and retain that data.
	▲	nprateem an hour ago \| parent \| prev \| next [-]
		One reason why Google made that algorithm to watermark AI output
	▲	black_puppydog 7 hours ago \| parent \| prev [-]
		That's exactly why text written before the first LLMs has a premium on it these days. So no, all major models suffer from slop in their training data.

▲

andy81 7 hours ago | parent | prev | next [-]

We've all tried to ask the LLM about something outside of its training data by now.

In that situation, they give the (wrong) answer that sounds the most plausible.

▲

PacificSpecific 7 hours ago | parent | next [-]

That's definitely been my experience. I work with a lot of weird code bases that have never been public facing and AI has horrible responses for that stuff.

As soon as I tried to make a todomvc it started working great but I wonder how much value that really brings to the table.

It's great for me though. I can finally make a todomvc tailored to my specific needs.

	▲	ctoth 4 hours ago \| parent [-]
		I'm not sure what sorts of weird codebases you're working with but I recently saw Claude programming well on a Lambda MOO -- weirder than that?

▲

visarga 2 hours ago | parent | prev | next [-]

> In that situation, they give the (wrong) answer that sounds the most plausible.

Not if you use web search or deep report, you should not use LLMs as knowledge bases, they are language models - they learn language not information, and are just models not replicas of the training set.

▲

NoMoreNicksLeft 6 hours ago | parent | prev [-]

Once or twice, for me it's deflected rather than answer at all.

On the other hand, they've also surfaced information (later independently confirmed by myself) that I had not been able to find for years. I don't know what to make of it.

▲

visarga 2 hours ago | parent | prev | next [-]

> because it has nothing to pull from?

Chat rooms produce trillions of tokens per day now, interactive tokens, where AI can poke and prod at us, and have its ideas tested in the real world (by us).

▲

elictronic 6 hours ago | parent | prev | next [-]

This then becomes the hardware manufacturers problem. If their new hardware fails for to many users it will no longer be purchased. If they externalize their problem solving like so many companies, they won't be able to gain market share.

This creates financial incentives to pay companies running the new version of search. Your thinking of this as a problem for these companies, when in reality it is a financial incentive.

▲

esperent 7 hours ago | parent | prev | next [-]

Presumably companies will still provide manuals.

	▲	SiempreViernes 7 hours ago \| parent \| next [-]
		It'll be a single sheet of paper with a QR code that redirects to a canned prompt hosted at whichever LLM server paid the most to the manufacturer for their content.
	▲	PacificSpecific 7 hours ago \| parent \| prev [-]
		If that was adequate then wouldn't there not be supplementary material? Results vary of course. I have some very wonderful synthesizer manuals.

▲

roxolotl 6 hours ago | parent | prev [-]

Yea so I’ve had an issue getting video output after boot on a new AMD R9700 Pro. None of the, albeit free, models from OpenAI/Google/Anthropic have really been helpful. I found the pro drivers myself. They never mentioned them.

Thats not to say AI is bad. It’s great in many cases. More that I’m worried about what happens when the repositories of new knowledge get hollowed out.

Also my favorite response was this gem from Sonnet:

> TL;DR: Move your monitor cable from the motherboard to the graphics card.