Yes. It's wrong all the time and steals all its info from the internet. Just like we humans do.

It just like having conversations with a real schizophrenic

I can do that myself in Notepad

I can do that right now on KC

Ain’t that right retard

🫵🤠🖕

Jijijijijijijijajajaja vamos

tried claude today and it genuinely mogs chatgpt so far, unlike chatgpt it actually brings in new info and doesnt run in circles

it doesnt really have a free tier though, you only have like 15 maybe free messages, but maybe that's for the better

Censorship and "hallucinating" is crazy on the thing

the hallucinating thing on AI it's a misunderstood. AI read the input you write by tokenization, that means it reas like that:

The AI just cut the sentence in pieces so it can be understood, then every word has an ID number, hello is [154], world is [3822], and hello world is [154, 3822].

Now, the AI "stores" the word car as the ID 154, when it thinks on 154 ID, it searches what's inside that ID and there is 1536 decimal numbers...A lot of numbers just for a word.

Then, every ID is stored on an opposite ID far away, in meaning, example:

Car: technology, movement, metal

opposite?

Forest: natural, tree, rocks.

 

154 = car, 10 = Forest (or rock).

 

What i'm trying to explain with all of this is...AI lenguage is shit lol, it's functional but the way it's made gives a margin to mistakes that would make people think that AI is conscious or that AI can hallucinate, they are just mathematical mistakes because the lenguage AI uses it's starting to become old (yes, i'm being serious), they are constantly researching new ways to AI to recieve and process inputs.

the problem at the core is language isn't math.

math is a language, yes, but human language doesn't convert to pure math without mistakes.

 

another point is:

our current model is at it's theoretical maximum performance, when it hits the average/median of the training data; not better.

 

if you filter the training data heavily, you do not have a wide range of possible topics.

if you do not filter the training data, you have shit tier responses, because the average human is shit tier.

I think what "hallucination" means (at least from what I read) that when there is no data in the training set in it's "memory" then it comes up with something that likely exists irl.

The example is that it was referring to a package on github that didn't exist and listed as a dependency for some application (I think it was some aliexpress api stuff), but the name of the package he composed from relevant snippets, and it sounded likely such thing exist.

Thing is, these LLMs, chatbots, autocompletes, aren't occupied by giving an answer or checking facts, but by generating a text that sounds human, sounds believable.

hmmm...I kinda agree, but because Math is perfect, not because language isn't. But math is just...Different kind of "perfect", it's almost divine and it replicates in the universe or quarks. "As above, so below", "Your kingdom come, your will be done".

 

I kind of understand what you say? there is no more room to improve AI with the CURRENT models, that's half true, we can still use this "AI Language" but do bigger datacenters, you will not have a ferrari but your car will be faster.

 

 

it's relative because it depends on what outcome you want it to be. If I learn about programming and ask chatgpt "hey give me this code, hey this is a mistake, fix it, hey there is another mistake" he will just give you the code and generic text. But if you TALK to the AI and ask why something happen, it can actually teach you why something is like it is, and how can you do that.

 

Again, only with coding and programming, I didn't test it with anything else, but it's really good to teach how to develop your skills on coding.

Welcome to the Dystopia. The internet is censored all to heck nowadays.