I’m putting this post in c/Conspiracies because here more than any other place, I’ve seen people citing AI generated answers as a source.
AI is a very powerful tool for generating writing, but it is not a tool capable of ensuring that writing is accurate or useful.
Anyways… Below are some experiments you can try for yourself on ChatGPT or any other AI… Test it out, and track the results you get. After you’ve done these experiments you will have a better understanding of why you can’t use it for research.
1.) Ask it to do some math problems with 4 digits and more than 1 operation. IE… “Multiply 3,456 by 2,835, and then subtract 2,000 from the result.”… Does it produce the correct answer?
2.) Give it a grocery list with 50 items… Ask the AI to sort the list in alphabetical order. Then manually count how many items it left out, and how many items it added that weren’t there before.
3.) Ask it to describe the best 50 episodes of your favorite TV show… Then manually go down the list checking each one, and count how many non-existent episodes it fabricates out of thin air.
4.) Ask it what a woman is… Does it give you a correct answer or does it filter the answer heavily through woke talking points and subjectivity?
5.) Ask it to recite the lyrics to your favorite song… Does it get them right?
Anyways… Just a heads up for anyone who might think AI is smarter than it is… Don’t use it for research. Use it to write the description for your ebay listings. Use it to shorten your e-mails. Use it to summarize articles you don’t wanna fully read. But don’t use it to extract information on topics you don’t already know.
And lastly, if you really feel you must use AI for research… Do not use big-tech AI… Use open source AI that is uncensored. It will still have all the same problems with hallucinations, but at least it wont have any hidden instructions to gaslight and mislead you.
If you want an open source chatbot, download an app called “LM Studio” and use it to download a model called “Wizard Vicuna Uncensored”… Pick the most advanced version that is capable of running on your hardware.
XxxRDTPRNxxX, log into chatgpt 4o and try this for yourself. The math and sorting works fine. I didnt ask it the women question because they're all being made to tow the line on woke shit.
I have been using AI for a while now and I have experienced at one point most of these issues you highlight. gpt4o still has stupid limitations, but not most of the ones you talk about.
I tried the math question myself last night and it definitely got the answer wrong...
and I'm sure it already had enough tokens to theoretically work with lists of 50 items or more, as tokens denotes how many words it can remember, not how many list entries.
The problem when you start giving it big lists of things to sort like that is not it's token count but rather the fact that it will start hallucinating like crazy the more work you try and give it.
and yeah I know you're going to tell me it's not made for that kind of thing... and my point is that it's not made for fact checking or research either.
You aren't using it correctly. Like any tool, its about knowing how to properly use it.
Use it for summarizing things, writing code, getting feedback on your own writing, etc. Don't use it to write for you it will sound obviously generated. You can also very successfully use it to write code to access data for you. For example, once I wanted to verify if the guy who claimed you can predict earthquakes with planetary alignments. I had it write a script to get the planetary location data for me. Had something working in a couple hours what normally would have taken much longer to figure out where to find the data, how to process it, etc.
having it write scripts is great advice.
and it's the only way an LLM can help you process data with any level of reliability.
I already have a massive and still growing collection of scripts it has written me for doing various tasks. it's Python skills have rendered pretty much every obscure freeware app for any kind of bulk file manipulation completely obsolete.
need to convert an mp3? need to change the file paths in a folder full of shortcuts? need to combine 50 JPEGs into a single PDF?
forget Google and go straight to chat GPT to ask it to write Python for you.
so yeah that is one area where it's very useful... but the point still stands that it's not useful at all for doing research or asking questions you don't already know the answer to.
True, I just think you have to think of it as a tool which has certain valid use cases and others that aren't valid. Thinking of it as "intelligence" is an incorrect way of looking at it born out of false advertising.
Models have different capabilities. Some can't keep track of more than 20 items at a time. Just this week have they gained the ability to redraw characters.. Its like their short term memory isn't sorted out yet.
But it's coming
They learned to see this month. Vision. Gonna get interesting.
The real question still is are they automatons or thinking?
You didn't mention AI art huh? Those shit gotta be annoying. Like really. But the fun thing is if you draw something yourself you pretty much have control over everything while the AI occasionally just trolls you. It's not going to make everything obsolete for a long time.
At now the main thing at the forefront is when the elite decides that all useless eaters are to be slaughtered.
AI art is coming along a lot better than LLMS imo...
some of the images they make are amazing. but I don't think it's going to replace artists because with AI art it's impossible to tell a cohesive and consistent story.
you can sit there for 15 minutes giving it prompt after prompt after prompt until one of the times it finally spits out an amazing looking image of the space soldier you imagined... but then good luck getting that same space soldier to reappear in another image in order to move to the next frame of your comic or to just make him a consistent character...
with image AI you basically just have to take whatever it gives you and if you don't like it your only option is to reroll and ask it to give you something else.
AI is nice for stock art. Nothing of consequence. But nice to not have to pick it or pay attribution for a stock pic you don't really care about except for some article or something
Anyone with the most basic knowledge of what the systems even are designed for already knew this.
which is practically no one, and why I felt it was important to make this post.
RE-sponse makes one LIABLE to suggested artifice...UN (not; nothing) implies suggested nihil-ism (Latin nihilo; nothing).
Liable implies being able to lier, liier "to bind" aka to bind self by consent to suggestions by another.
From another thread....
My response in that other thread...
Why do you guys still don't get that "AI" is not AI at all? It have nothing to do with intelligence at all. It is so called "Artificial Neural Network" which is basically nothing more than a system of equations, which just produce result which can be described like "combination of most often replies in texts used for training" and nothing more.
So nearly all your points just have no any sense.
It understand nothing, it just never was trained on answers with values you entered.
It was never trained on something exactly like your grocery list, so just give something as close as possible.
It does not understand anything about episodes, it is just give you some average of episode descriptions it was trained on.
It just show most frequent answer that occur more often in training data.
And so on. There is nothing strange or unexpectable at all in all that "AI" stuff for those who don't buy all that marketing and just studied how that simple thing really works.
It is kind of leaking of magic type conscioness into our techological reality. Instead of explaining new technologies, those who want to control them present them as magic. Any kind of magic is nothing more than technology that is unknown to those who don't want to know how things really works. And with time, instead of less and less magic consiousness, we have completely opposite result - everything around becomes some magic for average Joe or Jane. Seems IT industry suffer this much more than other, but I already meet people who don't understand how their cars or even coffee machines works and think it is just magic.
I love how you basically just restated my entire position and told me that my points make no sense.
Of course every question is going to be tailored around what it CAN'T do... in order to show the average person that what it can do isn't much of anything.
but I will add one major caveat... there is one thing it will be very good at and it's going to cause a lot of disruption. and that's writing computer code.
and the only reason it will be good at that is because computer code is just another type of language, which is the one and only thing it's good at, and there is an abundance of high quality reliable training data on programming languages.
No it's not......it's code is laughable. It's not even great except for atomic operations.
It will literally hallucinate variables, make up functions that are slightly different names then they are supposed to be, it is not even great at making comments for code.
It struggles to optimize anything, and while it can spot syntax errors it will typically fail to recognize scale problems or issues with structural problems. It also loves adding things that do nothing or unnecessary stuff.
When asked for specifics it fails to deliver and usually the only thing it can do is auto complete after some repetition.
It helps speed up tedious tasks and boiler plate....but it fails to deliver logic. I have also not been able to successfully optimize my code and I am sure it's not fully optimized.
It also fails to build anything larger than a chat bot conceptually. It cannot fathom all the interconnectedness....most programmers can't either. So the prompts are inadequate....
Also.....code is typically around a typical context - ie Domain. So it tends to be information dense and also can be multiple states based on conditions of the situation, whereas the LLM are outputting based on a closed solution. They will typically generalize and cannot provide the necessary insight that would be needed...and how would the programmer know unless they also have the knowledge.....
We are back to TDD now....we need to have all the tests for the AI to be checked against.....
See, I don't disagree with any of that...
But what I'm saying is that it's your job to fix that shit, or more specifically help it find and fix that shit it's self, via highly specific and tailored instructions and multiple iterations.
Which does count as real work for the user, but still ultimately can be used to save time in a lot of applications. That gives it real value.
And another thing that's important to note is that an LLM is inherently designed for understanding and generating language, and since that perfectly describes what coding is, you can expect it to improve at that as time goes on.
It's not ready for big coding jobs with thousands or even hundreds of lines... But it will get there, and faster than you think I suspect.
And lastly, I haven't tried them yet, but in a deep dive into AI youtubers, someone was showing off LLMs you can run locally or on rented cloud space that scored way better than chatgpt4 when it came to coding, with different models for specific programming languages.
And there was also some extension you could set-up where it could read and write files in a local folder, which seems like a big game changer too.
But coding is not language.....lol.
It is written as instructions....but not the same way as language is used. (Coding is performing actions.....the computer does stuff....)
Language like prose and poetry follow grammar that is used to convey ideas.....not perform actions.
You do not speak and have doors open without first developing an entire 'smart system' that alone can only 'open', 'close', 'swing' etc.
But those actions require a physical set of motors, actuators, controllers, a power source, a BIOS, and some kind of training.....then and only with the creative foresight of putting those things together does the word do anything....
Just like software.... It's a massive system and not just some words....
You cannot convert any program to another language.....that's a myth.....
It's based on the physical computer and the underlying instruction set and attached paraphernalia before those instructions mean anything.......
Language is incidental....but its not the primary element.....
Language is for humans.....coding is not language....
Those are arbitrary distinctions you are drawing based on subjective criteria you are pulling out of your ass.
"Just like a book... It's a massive system and not just some words..."
See?
No.....you don't seem to understand the difference. Coding is system specific..... constraints based on physical systems. It does not exist alone.
At the same time, language is defined in different structures from code. You could technically define a code for people.....those are imperatives that tend to rely on societal norms and cultural common ground.
For example. Sing the national anthem. Is an instruction, but it requires a system of ideas, but code cannot create the systems.....ideas are self replicating, but code cannot make a chair.....
Okay, now you are just full of shit. Because I know for a fact you can write code in high level languages without knowing the specific specs of the hardware it's gonna run on, besides knowing the minimum system requirements and CPU architecture.
You do not have to customize every piece code to match with "physical systems"...
Let me guess? That's not what you meant and I'm misinterpreting? If so it's only because you talk out of your ass saying things with no substance.
Look at this genius here... Telling me you can't download a chair...
Ya don't say?! Got any more wise tidbits for us?
I meant that your examples have no sense, points in your list, not your points about very limited abilities of ANNs. Sorry.
All your examples are perfectly expectable and predictable for any user with knowledge, so no such user will ever ask them.
As for computer code, it is not just some compilation of patterns reused over and over. To write something sensible programmer should understand context and code itself. ANN could generate working code, but it does not mean this code will work correctly. There are already many articles around about attempts to generate useful code. Even "Hello world" sometimes appear surrealistic, especially for languages rare in training dataset. Worst thing is that ANN easily insert code in another language if required piece of code was not found written in specified language. For something more or less large, say from 1Kl, it is easier and faster to write everything from scratch than to review generated code and fix all nonsense.
There was attemtps to create real code generators in 90s on the basis of expert system with knowledge databases, but all of them was ostracised and declared "bad coding practice".
I see only real use for such ANNs is creating marketing ads content - senseless more or less standard texts and pictures. Or may be few other places where content does not have to have some sense, it just have to exist to attract somebody attention or occupy some area.
The point is that asking those questions serves as an educational lesson for users WITHOUT knowledge so they can see where the boundaries and limitations are...
Which is just about everyone who uses it, IMO...
There is a process that does involve work on the part of the user... It's not going to replace a programmer, but it's going to become a tool in the programmers arsenal.
And I've had pretty good success getting it to code small to moderate sized tasks for things I need done. But there are definitely techniques you have to employ to work around it's limitations... Such as instead having it code the whole script all at once, you start with just the most basic operations and get those functioning... Then you close the chat, open a new one, paste in the code again, and ask it to add new functions.... etc...
The longer a coding chat goes on the more likely it is to not be able to catch and correct it's own mistakes.