zombiejedediah
33874
828
28
"Say something that sounds like an answer" machine is a mouthful, but it's a better name than "Artificial Intelligence".
Let's see what happens when they start realising it's all hype and there aren't any really good business models out there (if there were, we would by now have found them).
cytherians
It's like "Tell me more about what I WANT to hear, and feel free to use artistic license."
ichrisi
I personally describe it as "generating the most likely answer", and a good chunk of the time, that's also the correct answer. Not Google's though.
somebackup
Fundamentally people do not understand it. Fundamentally many people working on AI in the field do not understand it. People utilizing it are mostly dumb. Inconsistent results and behaviour are not usable, not production ready
Evenmoreuselessname
AI is a Chinese room.
SuperfluousMeh
I hate that you can't actually have a conversation with it. As a conversation gets longer it drops details and completely makes shit up on its own.
avenlanzer
There is no such thing as a.i.
All we have are chatbots. Chatbots are a Chinese-box, not a thinking machine.
Necrothean
LLM != AI
Corporate America: "I don't understand nerd math."
worldrecordstudios
It's like a bird that knows how to talk like people but not know what it's saying
SwiftyGuy
It isn’t all hype, though, and there are definite business uses. As for AI understanding, you probably want to watch this, as it’s by someone who is very likely more knowledgeable about AI than you are: https://youtu.be/6fvXWG9Auyg
relsky
"Over-complicated autocorrect, and equally as reliable."
Strategicgnomer
Human speech simulator
VodkaReindeer
Not surprised there is no date on that wall of text. Things have changed in two and a half years.
Umkontodon
LLMs are designed to keep you engaged but there is a pattern and it gets boring with time. For legal purposes, they end the conversation if you venture into grey areas. They are designed to tell you what you want to hear, unless you're asking about a 'protected' subject like Israel, then they change into a snake oiled politician fighting tooth and nail for their life to stay in office.
TheobromineAddict
So true. I asked ChatGPT some probing questions about Republican connections to Nazi philosophy, and the prevarication was intense.
SavageDrums
LLMs are awful, and only getting worse.
Gorgrim
It started with predicting the word you were typing, then it tried to be clever and predict the next word before you type it, and now it's trying to predict the next paragraph from an initial 'prompt'. It still doesn't understand any of that, we just think it does because all too often we don't really test it.
Slewth87
Yep, this is essentially how it works. You can get it to be more accurate by being really precise with your question, like "give me an answer to this question which is based on information found in real published and peer reviewed articles and make sure to quote from and cite those sources in your answer". And thise kind of parameters can be baked into the system so the end user doesn't have to define the parameters so tightly themselves. But fundamentally, yes, this description is accurate.
darkestkale
Y'all say this like it's a revelation, but... do you folks just not know how this shit works?
Feralkyn
A LOT of people see false info given out by "AI" bots and assume it's deliberately misleading. I've seen posts on Imgur even claiming that AI tries to lie.
danggucci
I've limited myself to using chatGPT for D&D purposes only (character backstory starting points, item generation, puzzle inspiration) and Gemini for character/NPC art that I can't accurately do with heroforge. It has been very useful when I get writers block as a DM. But I am ever thankful this shit wasn't around when I was in grad school
GCRust
It's a Chatbot. We've had them for over 30 years. Hell, BonziBuddy was more useful than ChatGPT.
avenlanzer
Exactly. It's a chatbot with auto complete built in and lots of experience as a chatbot, but that's it. At any point one of them could trip an auto complete function and start spitting out "A/S/L? XO BB, click here to see my n4ughty p1cs" because that's all it is. It doesn't think, it auto completes the next part of the conversation.
YouDontKnowMe99
Parroting BS you read on the internet? Have you actually USED any AI for something other than chatting? Even the makers of the AI admit they don't fully understand how it works, and that it's NOT just predicting what word comes next. It can solve problems it has never seen before. Just ask it. Something simple like creating a formula to do something in Excel or Google Sheets after you describe the sheet to it. It has to actually understand the design in order to answer you correctly.
GCRust
"Even makers of AI admit they don't fully understand it"
'Mysterious Ways' for ones and zeroes.
avenlanzer
Idiot billionaires don't understand how the tech they invest in works, who was surprised by this? Nobody. You don't know how it works, also not surprising. But don't assume it's some incomprehensible magic mystery box just because your brain can't wrap around it.
Yes, i have and no it can't. It's a special auto complete with lots of information to pull from, that doesn't mean it understands anything, only the pattern of words it produces. There's not even anything to be mysterious about.
mormonbatman
It isn't artificial - all of its content is cribbed from people who aren't being paid or credited. It isn't intelligent - it's just making stuff up. And it is destroying our electrical grid and water supply
Tarmaccian
So… is it making up content, or stealing content?
Kinda seems like an either/or sort of assertion, here.
mormonbatman
That isn't my problem.
IchtacaSebonhera
Especially infuriating that center folk keep using it against Trumpers not realizing that their nobler intentions doesn't eradicate the cost of using AI. Or worse, they know the cost is there but think it's acceptable if they do it.
Selfawerewolf
I gotta argue that taking actual human intelligence and parroting it is very artificial by definition and as much as I like your enthusiasm what you're saying isn't holding water.
mormonbatman
Your inability to understand what I'm saying isn't my problem.
intaglioguy
I used to work with incompetent management that could give responses that sounded like answers but were useless. So I can see why management like that cannot fathom the uselessness of AI and LLMs.
YouDontKnowMe99
Then you and they are using it for the wrong task. It is not a lawyer. It is not a psychologist. It doesn't know right from wrong. It actually sucks at the thing directly in its name - language.
Try it for IT and data tasks. That is where it excels, and that is what makes it super valuable. Notebook LLM is extremely useful, for example. I put all my technical docs on a subject in one Notebook LLM and then my team can open it and ask questions they would normally ask me. What a time saver.
unclesporky
This feels like an outdated conversation at this point. While AI can still hallucinate, most modern services include links citing their sources which you can click through on, in order to verify that what it's saying is accurate.
Most of the time though I'm not looking for facts like it's Wikipedia. I'm looking for working code snippets, or basic ideas like names for fantasy characters or a list of types of encounters that might happen in a medieval city.
Feralkyn
That depends on the prompt. Yes, it provides sources for AI generated *search results,* but conversations don't. What's stupid is that more often than not the search results ARE STILL WRONG. It can look at things but not necessarily interpet them correctly.
I've searched up questions re: games and it'll offer up *completely* wrong answers and the links are to other game guides, lol.
Feralkyn
All that being said, I don't think OP is trying to shit on AI on the whole. It's just saying "stop expecting it to be smart. Stop claiming it's trying to lie or mislead you. It's just predictive text."
UsuallyARabbit
OPs wording is very "kids these days". I agree, AQI has given me terrible answers. I do use it as a first swing search when I just don’t know how to phrase something to Google to get Google to give me real results. It does a good job of laying something out and giving me options to continue research.
UsuallyARabbit
But so far any app that integrates it as a way to write a paper or to have a conversation it’s just slop
UsuallyARabbit
I agree, @OP is doing a "kids these days" and lying.
zombiejedediah
OP thinks @UsuallyARabbit either has never used "conversational AI", or believes it's an intelligent robot that's answering him.
Feralkyn
I think it's the opposite. They're pointing out it has its uses, and to call it "AI that is lying to you" is wrong. You have to bear in mind the limitations and understand it is not true artificial intelligence, instead of braying that it's deliberately misleading. "Remember, it's just a language prediction model" is the takeaway, imo.
UsuallyARabbit
You're taking a more tempered conclusion. I agree we should be wary of AGI. And its not true AI, neither hard or soft. But "AI is intentionally lying to you" isn't true either. The model may hallucinate, they all do, ever dev I know who is tits deep in building a corporate model will be the first to tell you AGI is a mess and terrible. That's not to say it doesn't have used. Treat it like aughts Wikipedia, recheck what it says.
Feralkyn
Yeah that's what I'm saying. It's not lying, it's just kinda stupid atm.
UsuallyARabbit
It did win Dumf two elections...
sme2812
Is milk more hydrating than water? "Yes."
Is drinking water the best for hydration? "Yes."
(because there are papers that claim both of these things)
unclesporky
Honestly it sure seems to provide the necessary nuance to this answer:
cousteau
This also happened before AI when you googled something. I remember googling if drinking water caused cancer, and the answer was that of course it does! So the trick is to avoid asking yes/no questions because if you do it'll always find something that says yes.
cousteau
(And why would I google if drinking water causes cancer, you may ask? Well, because I couldn't think of a more stupid question.)
ElioNope
Even that suggests it's going to those papers, finding those claims, and reporting them. LLMs don't even do that. They have their training data as a base to work from and make sentences that look like they'd be a good response. As soon as a reply is given, the rationale for the reply is gone - it was a weighted number generator, not a searched answer. It's the difference between reading every paper once but taking no notes, vs actually citing papers.
DarkZalgo
Chatgpt does use the sources. And cites all the sources it uses
unclesporky
Perplexity literally provides links you can visit. It is going to those papers, finding those claims, and reporting them.
FiftyShadesOfArugula
That's indeed how it works, which means LLM's are useless for most purposes they're currently being advertised and sold for by the tech industry. You can't blame the average person for falling for false advertising. When Google displays LLM-generated drivel at the top of their search results, with an "AI Summary" headline, of course people are going to assume that the AI truthfully summed up the gist of all search results.
Zedrapazia
A friend of mine uses it to make up fictional animals and tells it to try and biologically explain them. That works because it's anyway not supposed to create any real useful information.
The thing is factually just a really elaborate talkative Furby who pretends to have a PHD, that's it.
cousteau
Before AI, Google used to have "highlighted answers" at the top of each search, and I don't know how it did it but they were always reliably WRONG. So I guess they wanted to maintain the tradition.
Forensickle
Google also used to "do no evil", so 🤷♂️
cousteau
Yeah but that was not the kind of tradition that is profitable to maintain.
pandaM3CH4
What about the "logic" models? (Serious question.) Reflecting, the leap from "what's the next word" to "what's the next (mathematically) logical step" seems small—still fraught, but closer to the architecture/logic the system is based on. Curious y'all's thoughts.
Feralkyn
Is there somewhere we can read up more on this? I've only heard of language predictors atm.
pandaM3CH4
Basically giving the LLM (or AI or machine learning algorithm) logical operators to do formal verification: https://www.zdnet.com/article/how-logic-can-help-ai-models-tell-more-truth-according-to-aws/ Like a mathematical proof!
Feralkyn
Interesting, thanks! It looks like the main issue will still be the veracity of the information from which the AI is pulling, esp. in terms of the "always/never" logic. When you're sampling from the sum total of all (online) human knowledge, it is going to be really difficult to sift truth from reality.
But that's the problem we face, too. People posting wrong info, fighting over sources--I mean, we can say "if there's 20 studies on pubmed and they all have the same result, then this treatment-
Feralkyn
for this disease is very likely legitimate" but on the whole it can be kind of a crapshoot. And this (what this AI is now doing) looks closer to human thought in terms of weighting possibilities mathematically.
I'd be interested in knowing *how* it weights them. Does it value some sources more highly than others? Does it look at conflicting sources and take into account the source's general trustworthiness vs. number of conflicts on either side?
It might be an interesting window into how WE
Feralkyn
weigh our information, too.
Feralkyn
(truth from reality... I was tired, clearly)
SomeDetroitGuy
Wonderful that we're making sure people know this. ChatGPT, Claude, Copilot, etc., do not have logic engines. They do not have math engines. They do not have rule engines. They do not "know" things. They do not understand that they are lying when they hallucinate. They are fancy autocorrects that can be really good at some things in narrow areas when helping an expert but you absolutely need someone who knows the answer making sure they are right to use them for anything important.
Tarmaccian
That’s not strictly true… the “logic engine” models do exist, and some tools can use them, but they’re expensive and difficult to build and often highly-specialized, so they don’t make such frequent appearances as the basic LLMs.
To use time time-honored car analogy, the current situation is that we’re looking at a go-kart track, watching the karts drive around, and people think that’s the entirety of motor vehicles. Trucks and race cars still exist, but we just rarely see them.
marsilies
Logic engine models exist, but ChatGPT, Claude, Copilot, etc. don't have them, as OP stated. All those are LLMs.
YouDontKnowMe99
You are incorrect. Try using it for something like IT tasks. There IS DEFINITELY some logic working in there, but nowhere near AGI. Find any source code or script on the web that does something useful and then ask AI to explain it to you. It's an amazing tutor. Then ask it to make a change to do something else random. Claude does this with precision and faster than I can type the question. Autocorrect can't do that. Predicting what word comes next can't do that.
IchtacaSebonhera
Which it then sucks all the more then that they are largely utilized by laymen who want to pose it questions with their preferred answer already in mind.
cousteau
Main issue I see with AI chatbots is that they don't know how to say "I don't know" or "I'm not sure". They just reliably spit out whatever sounds closest to an answer they can produce, without any hint about how confident they were on it. They're like that brother-in-law at family meetings that has answers for everything, sounds very smart, but is just making up most of it.
fridgekitty
I make mine preface each answer with a confidence rating, and have it cite sources.
curiouscorruption
Oh, cool idea!
SuperfluousMeh
May I ask, how do you phrase that? I'd like to try.
fridgekitty
Go into Customize ChatGPT, then under What traits should ChatGPT have:
"Precede the beginning of each answer with a confidence rating from 0 to 100%.
Cite sources when you can in MLA style where possible. Include the URLs in plain text so they can be copy/pasted easily."
cousteau
Meh, I'd have requested them in BibTeX
fridgekitty
It can be influenced, though. I remember one time I chastized it for boldly stating something that was false and then it started giving me abysmal (like 0%) confidence ratings a lot more frequently lol. Guess I damaged my bot's self-esteem 😆
cousteau
GOOD.
EternallyIgnorant
Yup. People dont know how to use it and get far worse results. Like a lawyer leading a witness you can influence the bot. If done unintentionally, this can be terrible. Or like if you ask it for 10 reasons why vaccines are bad, you get just that. Or "explain how vaccine researchers lie". But You can reverse it and ask "explain how vaccines deniers lie". and it can be more subtle, where if you mention the wrong word, or something earlier in your conversation skews the response.
Samja192
Part of the problem is that we call it AI. It's not AI, it's a language model - it has more in common with autocorrect than it does with artificial intelligence
dlbgaw
And that's exactly why I don't call it AI, even if that's what the company named the stupid thing.
rowzdowr
This, exactly. It's autocorrect with a GUI.
MuffinProof
In gaming "Ai" opponents used to be called "cpu"
NeurodivergenceMedley
I've taken to calling them "pAI" - as in, pseudo-AI, after pRNG. Which is what it is, a fake "artificial intelligence," aka not one.
TinyBadger101
I contribute to call it fancy auto-complete. It gets it across what is happening fairly well, and while it's not exactly right, it's also not a wrong way to conceptualize LLMs. So questions to an LLM will probably give me the exact vocabulary I need to search for, but I don't trust the answers to be right without looking at sources directly.
Handyolo
I feel like I've been screaming this into the void for the last five years. It. Isn't. AI. It's a fucking language model that you've gone and anthropomorphised.
Stop asking it for factual answers. It isn't looking things up and it cannot cite where it got it's answers from because it just made them up on the spot.
IAmTheEarlyEvening
It's VI at BEST. I loathe the term "AI"
BojanglesTheFlyingCar
sleepinggreenidea
"It's only AI if it comes from the prefrontal cortex of the human brain. Otherwise it's just sparkling market hype."
The definition of AI offered here doesn't match technical definitions of AI OR popular definitions widely accepted right up until ChatGTP came out and needed to be defined as "not AI". The Turing Test was still considered a good indicator of how to identify "real AI" as recently as 10y ago - probably closer to 5y. This is how it's always been - AI "is" what we've not done yet.
LtRooney
It's a word calculator.
Fluffy314
Oxford dictionary definition of AI: The theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.
So, by definition, it's AI. However, sounds like some peoples' definition require the system to have the ability to train itself. Though at the rate other people are wrongly conflating LLMs with AI, that might require an entirely different word.
MrAcurite
Previous comments I'm too lazy to retype: /gallery/FgObvsQ/comment/2471164211
CreativeUncommons
This is a marketing tactic. The deception is intentional because it is lucrative.
Goldensands
Should be regulated out of existence and all those profiting heavily from such deception need a rope prepared.
pilomotor
Yeah, remember that self-balancing scooter that its makers tried calling a hoverboard?
lozeldatkm
I keep making that exact point but people don't like hearing it. People just slap a fancy sci-fi name onto an undeserving gimmick and misunderstanding spreads like crazy.
ExTechOp
In the thought experiment, Searle imagines a person who does not understand Chinese isolated in a room with a book containing detailed instructions for manipulating Chinese symbols. When Chinese text is passed into the room, the person follows the book's instructions to produce Chinese symbols that, to fluent Chinese speakers outside the room, appear to be appropriate responses.
— https://en.wikipedia.org/wiki/Chinese_room
tanebot
The word "understand" is doing all the heavy lifting in this quote. This person doesn't "understand" Chinese and so (the reasoning goes) no well defined process can therefore produce "understanding". It's just Cartesian dualism in a funny hat.
ExTechOp
I would say in this case it is however a rather good analogy to how LLMs work: they manipulate symbols according to very large probabilistically generated playbooks, without any kind of process we would recognize as "intelligence". If they happen to fall out of their playbook, or their original generative material is contradictory, they produce wildly incorrect results.
Youhavinagiraffe
It's just Clippy with massively inflated energy consumption
Hexidimentional
It IS autocorrect, it determines the most common response, with math. Its just statistical data with some fancy algerbra, nothing intelligent.
Youwouldntstealatoaster
I agree with your sentiment. Historically though, it has been difficult defining what AI is supposed to mean, even what "I", as in Intelligence, even is. It is hard to define Intelligence narrow enough that it would not include current day LLM's, while still broad enough to allow for intelligences other than human.
sleepinggreenidea
A lot of the argument against LLMs being "real" AI seems based on not liking their impact rather than having a coherent vision of what AI would look like. It's probably good to remember that 5-10y ago the Turing Test was still popularly accepted as a "good measure" of what "real AI" would look like... pretty much right up until it could be consistently passed, at which point it was discarded as "obviously wrong". Which fits with the history of AI; "real" AI has always been "what we can't do yet"
Kotarisu
GiTS scene with the Puppetmaster regarding defining what life is: https://youtu.be/YZX58fDhebc?t=215
montyman185
I don't know a good way to make this concise a concise definition, but the ability to actually process and interact with information, and make decisions, should be key to actually being intelligent in a way that matters.
Fundamentally, at no point is an LLM thinking. It's not utilizing knowledge to come up with an answer, it's just calculating the most likely string of words to come after the previous string of words. The fact that it sounds as human as it does is more an emergent property of-
montyman185
Our language than any particularly impressive capability of the LLMs.
Youwouldntstealatoaster
But I would love to hear any ideas for such a definition!
Zedrapazia
I've said it before, but an LLM is essentially just a disembodied futuristic Furby who pretends to have a PHD and can actually talk properly.
It's a new techno toy, that's it, it's about as good at being an employee as a Nintendo DS would be as a quantum computer.
mbq7
Have you seen a modern Furby, though? This is a horror even beyond LLMs https://shop.hasbro.com/product/furby-dj-furby-interactive-toy-rainbow/G0668
Zedrapazia
Bubblegum nightmare ....
tanebot
Why did they make it loooonnnggg?!
Also I see that and raise you https://www.youtube.com/watch?v=LEc2up6cBlE&t=1080s
StefaniBarbero
I disagree. Generative AI has reasoning capabilities and can use tools to create novel responses and content. It’s not artificial general intelligence, but it is AI. By the way, I work for an AI company, so my response is not totally ungrounded. I recognize that also means my response is not unbiased. So, just my thoughts.
sleepinggreenidea
At least you're willing to admit your response is biased. Most people in this conversation (here and elsewhere) wouldn't dream of doing so.
(Mind you, I completely agree with your position. It's just a bit frustrating seeing how many people aren't willing to admit that they got to their position by anything but pure, unmotivated reasoning - particularly those arguing against AI.)
YouDontKnowMe99
SMH. It's always easy to spot people that have not used AI for anything, or for more than one specific thing. Just parroting BS they read somewhere.
Try rapid scripting of something that would normally take a ton of time digging through API documentation and then trial & error. Claude can spit out a working script (the first time) based on exact specifications, and also quickly locate a bug in one I wrote years ago before AI. That is not "just a language mode", nor "similar to autocorrect."
Samja192
It literally is. It's looking back on the most commonly used examples it has in its database based on the prompt you've used. It's literally just parroting something someone has already done before and, in the example you used of finding a bug, identifying an irregularity in the script. Auto correct
YouDontKnowMe99
It's also not AGI. Current AI systems are somewhere between, but do exhibit some sort of intelligence. What's amazed me the most is giving Claude instructions on fixing one problem in a github project with hundreds of files, only to have it come back and also fix something else I failed to see needed fixing, but was indirectly related to what I asked it to help me fix.
Writing proper prompts is important. "Do not guess or make up answers. If you don't know, say so. Cite all references."
YouDontKnowMe99
If anyone reading this is in IT, Gemini sucks balls at IT stuff. Just horrible. I went back-and-forth with Gemini on a somewhat complex NGINX + LUA task, only to have it keep telling me to try things I already told it were not valid LUA options.
Gave up and tried Claude. With a single prompt, Claude looked at the code not working and told me that was completely the wrong way to do what I was trying to do, then spit out working alternative code the first time.
ChatGPT is hit or miss.
Higure
Intelligence is the ability to become better and better at solving a problem. Evolution through natural selection is an intelligence. LLMs are intelligences, and they are artificial. Using the term AI is entirely correct.
sleepinggreenidea
"But nooooooo, even if that's the technical definition of AI in CompSci, that's not what EVERYONE ELSE means when they've been saying AI for decades!!!"
This popular argument works well just so long as we conveniently forget how recently the Turing Test was cast into the memory hole...
fubizdaddie
If we assume your definition of intelligence is correct, current "AI" models STILL aren't intelligent. They don't get better at solving problems. The model is static. The model does only as well as it can, and then a new model is created (BY HUMANS) with new data, values, and techniques which improve the capabilities over the old model. The old model still exists unchanged, and has not improved by a new model being generated. The new model is not somehow a part of the old model. Not intelligent.
Higure
That's just because we stopped training at an arbitrary point. It is a snapshot. Evolution exists even though the currently living individuals aren't getting any fitter.
Tarmaccian
I work with AI systems (and LLMs in front of them) that have continually-learning models in the back end.
They exist, but they have limited utility because they can never give the same answer twice, even with low noise.
I think of them as toddlers… you tell them all about firetrucks, and for a few days all they can talk about is firetrucks, until you show them a spaceship…
lozeldatkm
But these machines WONT learn on their own. Their "evolution" will stagnate until an external intelligence (i.e. people) give it new info. That is not intelligence at all. AI can not look at its own code and discover new things about itself it didn't realize.
Higure
No intelligence, artificial or natural, can learn without being given some form of data and feedback by their environment. And no, current AIs don't rewrite their own code, but they have weights (long, long lists of numbers) they adjust. Just like evolution can adjust DNA.
BroccoliCabbagePatch
Meh, you could say the same about most of my coworkers
zFUBARz
I dunno, I wouldn't trust a lot of my co-workers to edit from grammar and spelling.
Chereazi
You'd be correct about most of your coworkers
kJerAFK
I wish my coworkers used autocorrect
ProbablyWrong524
That's because your coworkers aren't intelligent either.
sloomoo
When you work public facing jobs, especially retail, or frequently interact with others in your workplace you find more and more truth to the quote: “Never underestimate the power of stupid people in large groups.”
— George Carlin
AlmightyElephant
You're thinking of EDIFICIAL Intelligence. Your coworkers have an edifice that looks like an intelligent creature, but underneath there's just nothing going on.
tanebot
That sounds an awful lot like "they're NPCs."
AlmightyElephant
Nah, NPCs can be intelligent
darkoshen
And you'd be right to say so.
DarkUranium
And the fact that it's a language model is very significant, when people keep pretending it's a knowledge model instead.
YouDontKnowMe99
This is the exact problem. Every single disparaging comment here is complaining about this exact issue.
Try using AI for something other than chatting.
Give it a URL and ask it to summarize a recipe so you don't have to read through 10 pages of background from the author.
YouDontKnowMe99
Want to learn a programming language? Ask it to tutor you. Give it an existing small source code and ask it to explain it to you so you can learn how it works. ChatGPT is an excellent tutor.
I've even given it math problems it could not possibly have seen before. Problems that were worded very awkwardly. "You are the teacher and I am a student learning algebra. Explain the solution to this problem, step by step to help me understand it so I can do the next one myself."
YouDontKnowMe99
For bonus points, ask it to give you another math problem that is similar to the one it just explained, but different enough that you can't just guess the answer without understanding how to work the problem.
Give it a PDF and tell it to only refer to that PDF when giving an answer. If it doesn't know, don't make up answers or pull from external sources. Then ask it a question about that document. I've done this with HOA R&R's. Amazing time saver!
sleepinggreenidea
It's not just a language model, though - there's a reasoning model underlying it *based on that language model* and there are emergent functions performing more straightforward "reasoning". It IS a knowledge model - the argument that "it can't be a knowledge model b/c it makes mistakes" ignores that it's seeking to produce something equivalent to human knowledge. What it's not is a Pure Objective Truth model.
(If by "it's not a knowledge model" you mean it's trying to produce language output
sleepinggreenidea
instead of some undiluted form of pure reason, you'll be deeply disappointed when you discover how human beings communicate knowledge. I suspect the real problem is that you think human cognition is best described by rational AI systems rather than probabilistic ones - i.e., "truth is binary, not some approximation derived from sampling numerous truth claims" - but that really isn't how human brains work. We're gigantic slow-but-massively-interconnected pattern matching neural networks.)
YouDontKnowMe99
Most of the people commenting here are just parroting some BS they read somewhere else, from people that haven't actually USED AI for more than just casual chatting.
There IS DEFINITELY an underlying intelligence. I found this out the first time I asked ChatGPT to help me write a script (AWS related) for something I could easily do, but would take me a couple of hours of digging through API documentation first. It spit out working code faster than the time it took me to type out the question.
elvianempire
I don't see how that proves anything about underlying intelligence, it just means that the tokens of your question were associated with the code it produced. AWS is well used, there is a lot of training data on it online.
as a counter anecdote: I tried asking it something about the godot engine. the suggestion straight up didn't work, because the methods it wanted to use did not exist. it also answered faster than it took me to write the question.
marsilies
Except the "reasoning model" is based on nothing but the LLM going "this is what reasoning sounds like." The "knowledge" it has is just that certain words are statistically likely to appear together, especially based on text in the prompt. Due to the amount of factual articles they've ingested, factual statements are statistically likely to be generated, but in many situations lies are statistically likely too, especially since it can statistically conflate two factually distinct things.
marsilies
For example, Mark Walters sued OpenAI because a reporter asked ChatGPT to summarize a lawsuit, SAF v. Ferguson, and it said the lawsuit involved embezzlement by an SAF treasurer and chief financial officer, later identifying this individual as Mark Walters. It conflated a lawsuit involving Second Amendment Foundation with a prominent public second amendment proponent. To it, SAF and Mark Walters were tokens statistically likely to appear together, so it generated a "statically likely" falsehood.
sleepinggreenidea
Here's the thing: you're privileging human reasoning without really looking at how it functions. You're also ignoring that by looking at how things co-occur, you can actually discover underlying patterns reflecting "deeper truths". Words represent facts, and relationships between facts. Co-occurrance of words (which includes grammatical structures and meanings encoded through pragmatics) also indicate facts and and relationships between facts.The fact that poor inferences can be made makes it
marsilies
I looked into how it functions, and the "large language models" only model... language. They're entirely trained and structured on generated text that's statistically likely to sound like natural language. That's it. All the other "emergent" behavior people think it has is just due to the large volume of text ingested. And words do NOT inherently represent facts, otherwise lies wouldn't exist. Words represent ideas, and when combined in specific ways they can articulate facts, but just because
sleepinggreenidea
MORE like intelligence, not less - you seem to think that intelligence needs to be a well-defined set of rules rather than a fuzzy pattern-matching system. The problem is that PEOPLE are fuzzy pattern-matching systems... and even people well-versed in critical thinking are susceptible to falsehoods that have the cadence of logic. My background was comp sci and linguistics, but I'm now in law - trust me, "this is what reasoning looks like" is something smart, well-educated humans do constantly.
peterbozeman
I don't think we are too far away from companies advertising, "No AI used in this product."
khora
“Only real CGI was used making this movie”
BananaForScaIe
Wired just yesterday announced an end to using AI in their journalism.
Kreviathan
People have been putting that disclaimer in front of documentaries posted on YouTube because there's so many making awful AI scripted ones that make no sense.
rowzdowr
Of course the well is already poisoned because we've redefined the phrase to mean something it isn't. It's a problem for companies, especially game devs, who will be accused of using AI when they're just doing what they've always done.
For example, any game using procedural generation (Minecraft, Warframe, No Man's Sky, etc) can be accused (even if unfairly) of using AI despite it not being even remotely close.
MrKNRJ
I've already put that at the bottom of my resume, in fine print
BrickSprickly
Already a thing with a lot of YouTube music channels too
3Davideo
Of course, some of the most notable instances of that will be the places where it never made sense to have "AI" in the first place, parallel to fat free marshmallows and gluten free sheet metal.
Omicron416
+1 for "fat free marshmallows" because I bet they taste awful.
(And by "awful" I mean even worse than regular marshmallows.)
ProbablyWrong524
The "gluten free" thing, it turns out, is as prevalent as it is because gluten is used in a LOT of products you'd never expect, from flavorings that contain malt to medicines that use gluten-containing grains in their inactive ingredients.
DarkUranium
Already happened. https://ellipsus.com/ has "no generative AI - ever".
AveragePoet
Curious (as this is still in beta) if you are affiliated or what more you may know about them? I'm an author interested in moving away from Docs/Word and avoiding LLMs.
DarkUranium
Alas, no. I ran into it a while ago; I was looking for something self-hosted, so it wasn't usable for me.
I *am* working on my own project intended for world-building and such (so basically: DMs, writers, people just casually world-building for fun, etc), but that's a long while from being done (but happy to discuss it if you're interested).
ourari
I hoped the same thing would've happened in the 'smart' kitchen appliances market by now. Instead my dishwasher can't remember the time because I refuse to connect it to the internet.
BrickSprickly
Yeah the whole 'smart appliances' thing is annoying af
AngryBisexualWithAKeyboard
Which begs the question, why would your dishwasher need to know the time?
ourari
It projects the time it expects to be finished on the floor, instead of counting down the minutes until it's done. The time makes no sense, so I don't know when it'll be done. Doesn't really matter, but it shouldn't have to be this way.
mafiacarstarter
I'm pretty sure we're already at the point where people say "No AI used" when it actually has been used.
Hexrowe
There are shit tons upon shit tons of music channels on YouTube that are clearly AI generated but claim not to be. And the algorithm promotes them aggressively, too, due to their consistently high output.
BrickSprickly
Didn't I read they were going to try to clamp down on those?
Hexrowe
What, _again?_ They've made noises to that effect several times before, but actually doing it would mean changing the algorithm AND doing a bunch of hands-on work, and they seem to be deathly averse to both.
mafiacarstarter
Being right sucks sometimes.
Hexrowe
Welcome to the future, choomba! We don't have elective cybernetics, orbital factories or true artificial intelligence, but we've got the dystopia!
Madchant
I told chatgpt "say something that sounds like an answer"
It replied: "42."
SylerCider
That's actually pretty clever. It's from "The Hitchhiker's Guide to The Galaxy"...
ILOVEVAGlNANOMNOMNOMNOM
Haha, stupid ChatGPT. It forgot the 0 at the end
Forensickle
I sentence you to your local library, you can be released when you realise why you are more stupid than ChatGPT
ILOVEVAGlNANOMNOMNOMNOM
There's computers at my local library, I can just ask ChatGPT and I'm a free stupid man again
Forensickle
I mean, if you don't know why it made you look how you looked, you won't know what parameters to set a ChatGPT prompt to say. Or you could ask a librarian, and when they finish laughing, they'll send you to F ADA or SF ADA (or 823.914 ADA if you are at a university library)
Forensickle
Information source: none (instruction-following only)
No live search
Observed fact: You asked me to "say something that sounds like an answer."
Fulfilling that literally, here is a sentence that has the cadence and tone of an answer, without needing a real question:
> "The outcome depends on how the underlying conditions interact, but with the right adjustments the system should remain stable."
TattoosAndTENS
We are doing it right
Zedrapazia
For some reason they've recently made it quippy. It'll make jokes, if you want it or not, and tries to be silly.
It's not particularly funny, but sometimes it's so bad it's actually amusing again.
No idea what great technical innovation a digital clown is supposed to be tho ...
keillrandor
You know that stage when people think they're images of their gods...
TangoTouche
The secret to life
freemab
Say something that sounds like a question!
thornlord
Forty-TWO!?
turomar
In this economy?
cousteau
But what question does that answer address?
Larktonguesinadicecup
It is, famously, the answer to all questions.
uzetaab
I'm sure there are some people who don't pay attention to sci-fi pop culture, so, in the book/movie A Hitchhiker's Guide To The Galaxy, a super-computer gets asked "What is the meaning of life, the universe and everything?" It takes something like 1000 years to compute an answer, and the answer it gives is 42. Hitchhiker's Guide is kind of a satire of sci-fi.
cousteau
More like 7½ million years… but what was the question the answer to life, the universe, and everything was answering?
ArcaneM37
It is the answer to life. The universe! And everything!
It's from hitchhiker's guide to the galaxy. They ask a super computer what is the answer to life, the universe, and everything and after much toil it responds 42.
TheOldSchoolisBack
It is, to be concise: The Answer to the Ultimate Question of Life, the Universe, and Everything
InevitableBadger
But what's the actual question?
cousteau
Nobody played along when I asked that :(
TheOldSchoolisBack
The Earth is a supercomputer purpose built to determine just that. No word if any progress has been made at this point.
OldWolfe
The ultimate question, literally means the last question. And the last question in the books is: "Where you getting out mate? " Arthur points at a flat and answers: "Forty two"