Yes, people want answers. But what people need is to work a bit for those answers so they actually have some idea of how reliable they are. If you just get answers you gradually lose the ability to evaluate how believable those answers are. LLMs are comfortable because they produce in you the illusion of learning things fast. But the job of verifying sources and reasoning about the information you get, is just as time consuming. But because the answers seem so polished you won't. Which makes your understanding less robust over time as you grow lazy and complacent.
Also, it is unfair to judge search engines on what Google offers today. There was a time when Google cared about search quality, integrity of result sets and you didn't get a face full of ads with perhaps one "organic"-looking link somewhere on the page.
Google has become ad search. Not web search. And we should be comparing LLMs to actual search engines.
>Which makes your understanding less robust over time as you grow lazy and complacent.
I really wonder about this. Of course it has been bothering me as well. However, when I was young, I have been listening to the same argument wrt/ calculators and the internet vs the good old ways like books and the library. I think this will be the same somehow.
Verification is very easy with AI, I just tell it to bring sources when I feel like it.
Calculators automate a process that can be described by an algorithm and produces deterministic results. LLMs do not produce deterministic results. So they aren't really comparable.
Not using calculators is something you do for learning purposes. Not because you don't trust the calculator's ability to produce correct answers.
This is a technical detail that is orthogonal to my original argument. What do you think about the general idea - that it's about outsourcing thinking, and how good and bad it is, and whether it's just the next step like library->internet search, or something else entirely?
I don't think it is a technical detail. It is the entire point. Calculators perform purely "mechanical work". Answering an arithmetic question isn't thinking it is performing a calculation. Either manually or by using a machine that applies a series of known steps to arrive at a result that is _fully_ specified by mathematics. At no point is there an opportunity to hallucinate a result, for instance.
Deterministic or not, my original point is that thinking is outsourced by a human when using a calculator. And the pondering whether that's good or bad, and how much of it is okay.
What's your take on that? I suppose you don't mind calculators. What do you think about looking up information in books / library vs an old school internet search? Did or didn't the internet already made us lazy and complacent?
Like when I search for a plant (using a photo input) and it comes back with a name and other pictures of that plant I get an answer and verify it all from 1 ai response. I don’t need to click through to the original source for that when the answer contains what I need to verify its claims.
Google also includes direct quotes and citations with links - which for the most part I trust they haven’t hallucinated when they link right to it.
I used to work on web search engines. I worked on several search engines. Google's among them. I was fond of web search engines.
Google has not been a very good search engine for a long while.
It has become the opposite of what it once set out to be. It set out to be a search engine were relevance was king and where ads were curtailed. You got just a couple of ads, but it would more likely be relevant. They were also clearly delineated as ads. This made both search result and ads valuable. It was clean, it worked.
Then gradually people who don't stand for anything seemed to take over at Google and things god muddled. The search engine grew a little worse every year. Until it was utter shit.
Earlier this year I mostly stopped using it because it no longer delivered sufficient value. What I started using instead was a poor replacement for a good search engine, but it was better than the mess that is Google search.
The AI answers to searches are not search results. They are something synthesized from what it thinks I am looking for. It is a poor replacement for actual search.
If I'm going to ask an LLM I'd rather ask a higher quality LLM. Not whatever that thing Google has plugged into their search page is.
OF COURSE, search traffic goes down when you gradually degrade a product and then try to somehow make it relevant anyway by slapping a slop-generator on it. This is not a surprise.
And of course Google search is not coming back. Google is not about making products people want. Google is about making products they think they can monetize. I happen to think they are wrong. Perhaps other people think so to. But Google want to go in this direction so they will defend it even though it may not be the smartest thing to do. Careers, personal prestige and pride is riding on this. Of course they are not going to listen to informed consumers. That would be losing face and admitting they are flailing.
Google is great at fumbling products. The turn has now come to fumbling search.
This is a great opportunity for someone to take search away from them. Because search still has value.
Hmm. I think I understand where you are going but given that LLMs appear to need injected randomness to work effectively, determinism isn't really on the table so we kind of stop talking about the LLMs we use when we (strictly speaking) consider the case where they could generate deterministic sequences?
Not that this in any way takes away anything from your argument that "it's that language is imprecise". That still holds true, but it is on the input side.
Why would LLMs need injected randomness to work effectively? Their nondeterminism isn't by design, it's largely a byproduct of how (batched) inference actually happens on hardware [0]. In case you are referring to the actual token sampling (be it temperature, Top-K or whatever), i wouldn't really see this as part of the LLM itself (And I've frequently seen benchmark sites run them at t=0 anyways)
I think Canada might have an opportunity to help Europe at least reason a bit about regulations.
For the most part I am thankful that the EU regulates things strictly. As long as I am the consumer of what is regulated and not the maker. Because while the regulation is well meaning, it is not always effective or efficient. And sometimes it achieves absolutely no positive results.
And yet, I am generally for regulating a lot of things.
For instance the cookie-popup-horseshit is failed regulation that came about mainly because regulators were squeamish about real solutions that would have required constant judgement calls. Instead of posing an existential risk to companies that run roughshod over our privacy, they created impotent regulation that is a dual failure: it protects exactly nobody and it wastes everyone's time multiple times over.
Companies that do bad things should face existential risks. It is the only mechanism that works. Fines only work for small players. For large players it is surprisingly often a calculation where the numbers dictate you DO bad things because the downside risk is mostly quantifiable and financially acceptable.
Yes, sanctioning companies guilty of egregious violations of privacy would have been noisy, but it would have been the right thing to do. And it would have been cheaper and more effective than what we have today.
Quality was the norm when people had little or no disposable income and expensive things had to last for a lifetime, and some less expensive things (like shoes) had to last years.
Quality was what you judged products on and the origin of brands.
The overwhelming majority of brands today are meaningless because they have nothing to do with quality. They are either status markers or tools for exploiting our tendency to form habits.
You need go no longer than your local hardware store. Almost all the consumer grade electrical tool brands today are from perhaps 2 manufacturers. But you are likely to have a preference that you tell yourself is grounded in a rational, informed choice. Except, it mostly isn't. You probably buy the same toothpaste every time. Not because you have done an exhaustive study of toothpaste and concluded it is the best one for you. But because it is what you are used to buying.
(Yes, there are still some professional tool brands that make their own electric tools, but the kind of stores consumers shop in don't have those since they cost an arm and a leg)
Yes, quality used to matter. A lot. But most people alive today have not lived in those times. People have a hard time imagining things that are outside their own experience. So much so that they can even conclude "it can't be if I haven't experienced it".
I am starting to see how many developers actually need to re-learn programming in the age of LLMs.
A while back Claude went down in the middle of a somewhat frantic initial deployment of a product to production at a company where a friend of mine works. And suddenly nobody was able to do anything. Because nobody had actually read the code and had no idea how it worked.
So essentially: much of their day to day work now depends entirely on the availability of a couple of frontier LLMs.
If, for a moment, we set aside the fact that Europe is strongly incentivised to depend less on the US, and US companies, it should worry us that so many sectors are dominated by 1-3 or so players. One of the selling points for market capitalism, which we at least pretend to believe in, is efficient allocation of resources through competition. But when the market is dominated by de facto monopolies there exists no real competition and there is no real market.
Even in the previous political climate this should, objectively, be deemed a problem. Having 90% of all eggs in one basket is neither good, nor does it constitute functioning market.
I'm not sure what the Cloudflare market share is in the US, but if it is the same as in Europe, this is a problem for US companies as well.
Also, it is unfair to judge search engines on what Google offers today. There was a time when Google cared about search quality, integrity of result sets and you didn't get a face full of ads with perhaps one "organic"-looking link somewhere on the page.
Google has become ad search. Not web search. And we should be comparing LLMs to actual search engines.
Not whatever Google offers as "search" today.
reply