3 ms·

> Over the next 20 years the war between Google and SEO spammers meant finding things got a bit harder. Along with that Google's proliferation of adverts on the

by rchaud 4y ago

> Over the next 20 years the war between Google and SEO spammers meant finding things got a bit harder. Along with that Google's proliferation of adverts on their search result pages meant it actually became harder to even find the website link I wanted in the page itself. Yes, SEO spam has largely made the search experience a lot worse and ripe for change. But ChatGPT crawls these terrible links too! Many of the links Google spits out are themselves summarizing information found elsewhere, like a book or academic paper. So why would you trust GPT to provide an authoritative answer more than your own instincts? It's omniscience is a function of everything it's crawled; that doesn't mean it has any idea of how to weave them together into a coherent response.

visarga 4y ago
I don't think spam is an unsolvable problem. I think Google didn't want to solve it. I could find whole networks of spam sites in minutes, they could too. But the web is larger than what can be fixed manually. They should make an AI bot to search through the search box and decide if the results are crap, and train it with human preferences. Or set the bot to find difficult to find answers and see how many it finds and how high they are ranked. Probably working in search at Google was a dead end career, the search department wasn't sexy anymore and didn't have green field projects. Maybe they were just afraid to change anything as long as it makes money for them. So the good people left and they have been coasting for a decade.
- pixl97 4y ago
  To believe spam is solvable you must also believe capitalism is solvable. Spam will always be working on the next one up of the system there is too much money involved. Also by needing an always increasing amount of money to make the market happy Google will gladly become part of the problem.
mustacheemperor 4y ago
>So why would you trust GPT to provide an authoritative answer more than your own instincts To use a highly specific example, I've recently been doing a lot of work in Excel and the search space for Excel related queries has been almost completely destroyed by SEO adware sites. If you're lucky, the answer you're looking for is at least buried somewhere on the page you click on, but there's just as much a chance you'll get a semi-gibberish pseudo-article that itself is probably AI generated. Using my instincts/intuition to locate the correct answer is a non-zero cost of time and effort, made intentionally more challenging by site owners who want me to linger and consume advertisements. On the other hand, asking ChatGPT how to match two columns with XLookup or what have you almost always returns a correct, concise answer on the first try. If it's not correct, the formula doesn't work, and I can tell ChatGPT it gave me the wrong formula, and it corrects itself. Even in the rare event that happens, it's a fraction as much effort and frustration as it takes to rummage through the top search results for the same query.
onion2k 4y ago
So why would you trust GPT to provide an authoritative answer more than your own instincts? I wouldn't right now, but I believe the accuracy of an LLM is a function of its size and how much you can reduce the loss function. I think OpenAI can improve those things faster than Google based on the things that have been shown so far. I could be completely wrong though. I also suspect there's no benefit to corrupting a GPT model except for the lulz. If a search tool isn't showing links then there's no financial reason to put effort into setting up link networks and SEO content farms. GPT based search has an innate anti-spam advantage because it renders spam futile if the spam isn't getting visitors. It also renders a lot of non-spam content futile too, which is problematic. It's not a magic bullet.
- jononor 4y ago
  For ChatGPT and similar LLMs there is the loss function does not reward accuracy/truthfulness/correctness. But instead answers that seem correct to a superficial human check. These rewards making something plausible up instead of saying "I do not know", over-simplifying things, and passing off things with high confidence regardless of how much evidence backs the statement. Any successful deceit will bring down the loss. The overconfidence and lying is not due to model size beng too low - it is due to misalignment. If not corrected it will just get worse as size goes up. A recent Computerphile video with Robert Miles discussed this.

by rchaud 4y ago

visarga 4y ago

I don't think spam is an unsolvable problem. I think Google didn't want to solve it. I could find whole networks of spam sites in minutes, they could too. But the web is larger than what can be fixed manually. They should make an AI bot to search through the search box and decide if the results are crap, and train it with human preferences. Or set the bot to find difficult to find answers and see how many it finds and how high they are ranked. Probably working in search at Google was a dead end career, the search department wasn't sexy anymore and didn't have green field projects. Maybe they were just afraid to change anything as long as it makes money for them. So the good people left and they have been coasting for a decade.

pixl97 4y ago
To believe spam is solvable you must also believe capitalism is solvable. Spam will always be working on the next one up of the system there is too much money involved. Also by needing an always increasing amount of money to make the market happy Google will gladly become part of the problem.

mustacheemperor 4y ago

>So why would you trust GPT to provide an authoritative answer more than your own instincts To use a highly specific example, I've recently been doing a lot of work in Excel and the search space for Excel related queries has been almost completely destroyed by SEO adware sites. If you're lucky, the answer you're looking for is at least buried somewhere on the page you click on, but there's just as much a chance you'll get a semi-gibberish pseudo-article that itself is probably AI generated. Using my instincts/intuition to locate the correct answer is a non-zero cost of time and effort, made intentionally more challenging by site owners who want me to linger and consume advertisements. On the other hand, asking ChatGPT how to match two columns with XLookup or what have you almost always returns a correct, concise answer on the first try. If it's not correct, the formula doesn't work, and I can tell ChatGPT it gave me the wrong formula, and it corrects itself. Even in the rare event that happens, it's a fraction as much effort and frustration as it takes to rummage through the top search results for the same query.

onion2k 4y ago

So why would you trust GPT to provide an authoritative answer more than your own instincts? I wouldn't right now, but I believe the accuracy of an LLM is a function of its size and how much you can reduce the loss function. I think OpenAI can improve those things faster than Google based on the things that have been shown so far. I could be completely wrong though. I also suspect there's no benefit to corrupting a GPT model except for the lulz. If a search tool isn't showing links then there's no financial reason to put effort into setting up link networks and SEO content farms. GPT based search has an innate anti-spam advantage because it renders spam futile if the spam isn't getting visitors. It also renders a lot of non-spam content futile too, which is problematic. It's not a magic bullet.

jononor 4y ago
For ChatGPT and similar LLMs there is the loss function does not reward accuracy/truthfulness/correctness. But instead answers that seem correct to a superficial human check. These rewards making something plausible up instead of saying "I do not know", over-simplifying things, and passing off things with high confidence regardless of how much evidence backs the statement. Any successful deceit will bring down the loss. The overconfidence and lying is not due to model size beng too low - it is due to misalignment. If not corrected it will just get worse as size goes up. A recent Computerphile video with Robert Miles discussed this.