They cherry picked a comment about AI from some random student and put it on top. Meanwhile, the article says that these aren’t cuts as much as they are either curriculum modifications, or simply program cancellations because of lack of demand.
If we take a look at how the current AI behemoths got there, there’s a trail that goes straight from stolen data to proprietary models. They are charging their users for the privilege of using better aggregated public data. I hope that, when they raise prices once again and more and more users are cut off their larger models, people would understand where their place is, according to corpos.
I cannot make you understand something you refuse to learn anything about.
But now I get it. Your whole point is that you are so lazy, you are willing to trade accuracy of information and your own privacy, for the convenience of not having to read anything for more than ten seconds straight. I should have known from the moment you so confidently said that “AI is costing [you] nothing”. Sure mate.
Given all that, providing sources and facts from experience has been a waste of time from my part. No more.
No worries though, you won’t need to be intentionally obtuse for longer, outsourcing your thought processes to Google will do that for you, and for free.
JFC this is so tiring. I’m trying to explain the issues from a technical point of view and you dismiss all I’m saying like I’m crazy or I don’t know anything about the topic.
I get that you like these summaries because you believe that they make your life easier, but the least you could do is educate yourself, read some books, anything.
I see that you don’t know what an algorithm is. Because every search engine before Google used algorithms, just not Pagerank.
Tell me what you are seeing. This page you linked shows all top models for AI research at above 80% and the top ones at 90%.
On a single benchmark, GPQA, that was first released in 2022 and is now widely considered to be on its way out. This is extremely common in LLM benchmarking: every couple years, old methodologies are weeded out because most recent LLMs score 90% or higher. This could be for many reasons, but more than likely LLMs are trained on the specific corpus these benchmarks test against, which is akin to taking an exam knowing the answers beforehand.
The aggregates tell a very different story.
On the NYT article, you missed where it says that on the latest Gemini version, more than half of the summaries are ungrounded, meaning that users couldn’t possibly verify the accuracy of the response given the linked site. They also often provide additional information that’s not true.
I’m not even going to bother with the “I’m fine paying with my data” bit, because think that’s just a morally broken take.
Even the number is a bit misleading. First of all, anyone who has ever done LLM benchmarking knows that this isn’t an exact science, at all. You can totally get a 99% on a benchmark and fail every single task on another.
But even this particular claim is nuanced. From the original article:
But with Gemini 3, Google’s A.I.-generated answers were more likely to be ungrounded than when the system was based on Gemini 2, meaning the websites they linked to did not completely support the information they provided. In October, correct answers were ungrounded 37 percent of the time. In February, with Gemini 3, that figure rose to 56 percent.
Text search is what Yahoo and Ask Jeeves did. Then Google improved on it by adding algorithmic search.
There’s no such thing as “algorithmic search”. I don’t know where you got that term from, but again, not a thing. What Larry Page came up with was Pagerank, which is a ranking algorithm.
I bring up accuracy because its not 100% accurate. But if it works 85- 90 percent of the time, which it currently does according to benchmarks, that's more efficient than Text search even accounting for times you need to adjust
Citation needed. Where have you seen those numbers? Because there isn’t a single LLM out there that scores above 75% in publicly available benchmarks, for any given task. Meaning that there isn’t a LLM that does any benchmarked task with an accuracy above 75%, see https://llm-stats.com/
And no its not as cost efficient, but again I dont care as the end user because its not mycost.
Right. What do you think it’s going to happen here in the near future? That companies like Google are going to absorb the costs without passing them to customers at all, ever? Let’s say that they don’t, because they are for profit companies after all, what’s your plan? Signing up for a couple dozen free accounts to keep using them and become sort of a “LLM vagrant”?
But let’s say they don’t charge you ever. How do you think they are going to profit from you? Currently we know, they track your every move, you essentially pay with your privacy. Or you think they won’t? That they will forever lose money?
I guess the part I don’t understand here is that you must know that all these companies make money from their users, one way or another, and still you believe you aren’t paying for any of it. Are you ok with how they make that money, then?
I’ll bite. How do you think a LLM could accomplish what you want to do without using RAG, which necessarily involves a search engine?
In other words, what you want is a natural language parser than then goes to a search engine and retrieves content from a website to construct an answer that must quote verbatim the site. Which is what they have been doing for years.
Also, search engines do quote websites, all the time. Are you stuck with Altavista or something?
First of all, you don’t have any notion of how LLMs work, so you cannot “explain” anything.
But even if you did, the idea that a LLM should just interpret a query written in natural language, and return the results for that query quoted verbatim from the source, is so close to how traditional search has been implemented for the past decade or so, it’s hilarious that anyone would propose an alternative to it that is ten times more expensive and not nearly half as precise.
Regardless of all this, “what if we had LLMs that quote text verbatim” is not an explanation, it’s wishful thinking. It’s like saying that if we had flying cars, traffic jams would go away, then proceed to ignore the fact that cars don’t fly, and even if they did, they would be way more expensive, and even if they weren’t, they would create other issues like the need for air traffic control. Silly ideas are a dime a dozen.
I really do not want to be reading through a 300 page manual from my phone
The solution is text search. Has been for decades now. Search is good enough for these use cases, and proof of it is that nobody ever had to read a 300 page manual to fix such an issue.
Also, you know indexes exist, right?
The function is good. It’s arguably the best path forward. The issue is accuracy and cost.
If the issues are accuracy and cost, it means that it is a worse and less cost effective solution than just search. Unless you believe that LLMs can tell what’s true from what isn’t.
I honestly dont care how much it costs a trillion dollar company to provide a service to us if I dont have to pay. Google, Youtube, Amazon, Netflix, all operated in the red for years and years. I don't remember public discourse being omg how is Google going to afford to keep giving us nonshitty search.
You really believe that these multi trillion companies don’t make a profit? That they offer their products for free? Such a naive take.
Is it? Because for any lodging of minimum quality, hotels are cheaper nowadays.