I think the complaints are more about the "use logic" point than the sources, from my limited understanding, I would not say LLMs currently use logic.

There is a general issue where AIs fail in different ways than humans, and the failures look really dumb to a human. So humans tend to ascribe that dumbness from the human scale. Instead, I’d suggest they just have a dramatically different spider-graph of capabilities to a human, and are overall more capable than the “dumb spreadsheet / parrot” narrative admits. (Definitely not human-level IQ yet, to be clear.)

The start of its lengthy response:

“ The use of probabilistic logic models in LLM can lead to more sophisticated and nuanced logical reasoning. These models represent knowledge about the world in a probabilistic form, which allows the LLM to take into account uncertainty and make inferences based on probabilities.

For example, an LLM system trained on a large knowledge base might encounter a question that it has not seen before. Using probabilistic reasoning, the LLM…”

When you ask whether speaking English and using logic can improve the search experience, I wonder what you consider the most important parts of the search experience? I think many people, most of the time, might say that "accurate information" is their highest expectation, with "a pleasant conversation" somewhere below that. Delivering a plausible-sounding and pleasant answer that's completely wrong is... well, that's not a search engine I can depend on, is it?

You're hypothesizing a few things at the end that sound great! It's completely unclear whether any of those things will actually end up happening, so I think the focus on what is available today, with Chat-GPT and Bing, is more apt than a focus on what could be.

And I totally agree that a) a search engine that doesn’t cite its sources is useless, and b) you almost never want to chat with google like it’s a person. So you’re spot on. But the point I was trying to make is that the main use case is in stuff like automated summaries, specialized page rankings, expanding quick informal queries into longer formal-logic ones, etc.

The point is, "accurate information" is hard. Google's solution is snippets and while it might be fine for some cases, it fails terribly for others. There is zero guarantee an AI-based solution would be more precise, but for sure it will be way more confident - just like ChatGPT is.

Can you give me 3 example queries (questions and answers or typical searches) that are clear cut wins for a search engine application?

For one, you can't just "hook up" a language model to some other task, or to the web. ChatGPT is specifically built and trained to have good conversations. The fact that it can sort of appear to do other things is a happy coincidence.

To do any of what you want, new algorithms need to be built, and none of that is "easy". And finally, these models take A LOT of cpu time. They are not going to be reading thousands of posts in a second without serious and expensive compute hardware backing it, and that level of compute isn't remotely feasible to give out to individual users.

Even chatGPT, which is doing a fraction of the tasks you are listing, costs millions of dollars worth of hardware a day. The only reason it exists for free is because Microsoft has donated all that time.

They had to pay someone to label data and filter the worst "content" humanity has to offer. Otherwise it would've ended up like numerous other attempts at exposing "AI" to the Internet.

So it also has a huge human cost that OpenAI is not properly accounting for.

But chatgpt is already pretty darn good at summarizing and communicating. By “hook up” I literally just mean feed text from the internet into the prompt followed by whatever you want it to do with it - summarize, rank, whatever. Ignoring the economics for a moment (paid search engines?) and assuming that GPT 3.5 is the very best LLM we’ll ever get for simplicity: would you still say you don’t think tweaked versions of such models would MASSIVELY improve search?

There is also the business model to sort out. Right now search is primarily driven by ads, which I doubt will cover the costs of the sort of ultra-personalization that you're thinking about. Also, reducing the time you spend on a search engine or looking through results will further reduce ad revenue. However, I can see paid search engines perhaps leading to this.

So yes, eventually these models can help improve search, just not in the form that we have today. In a few years the story could well be different. I'm quite interested in seeing how Bing integrates chatgpt technology. They claim that they've created some new model on top of it that somehow links it to search results.

I question if this something that users need at the scale you're assuming. Wikipedia has existed for 20 years summarizing enormous breadth of human knowledge, with some articles having thousands of human editors. it's a boon for civilization in the way libraries are. But has it disrupted anything besides Microsoft's overpriced Encarta DVD market?

You're putting a lot of faith in computer models to provide accurate, both-sides'ed information on complex topics in a format that amounts to a soundbite.