Comment on We need to stop pretending AI is intelligent
Geodad@lemmy.world 3 weeks ago
I’ve never been fooled by their claims of it being intelligent.
Its basically an overly complicated series of if/then statements that try to guess the next series of inputs.
ChatGPT 2 was literally an Excel spreadsheet.
I guesstimate that it’s effectively a supermassive autocomplete algo that uses some TOTP-like factor to help it produce “unique” output every time.
And they’re running into issues due to increasingly ingesting AI-generated data.
Get your popcorn out! 🍿
I really hate the current AI bubble but that article you linked about “chatgpt 2 being literally a spreadsheet” isn’t what the article is saying at all.
Fine, *could literally be.
The thing is, because Excel is Turing Complete, you can say this about literally anything that’s capable of running on a computer.
And they’re running into issues due to increasingly ingesting AI-generated data.
There we go. Who coulda seen that coming! While that’s going to be a fun ride, at the same time companies all but mandate AS* to their employees.
I love this resource, thebullshitmachines.com
Also, Anthropic (ironically) has some nice paper(s) about the limits of “reasoning” in AI.
kromem@lemmy.world 3 weeks ago
It very much isn’t and that’s extremely technically wrong on many, many levels.
Yet still one of the higher up voted comments here.
Which says a lot.
Blue_Morpho@lemmy.world 3 weeks ago
Given that the weights in a model are transformed into a set of conditional if statements (GPU or CPU JMP machine code), he’s not technically wrong. (Of course, it’s more than just JMP and JMP represents the entire class of jump commands like JE and JZ. Something needs to act on the results of the TMULs.)
tmpod@lemmy.pt 3 weeks ago
That is not really true. Yes, there are jump instructions being executed when you run interference on a model, but they are in no way related to the model itself. There’s no translation of weights to jumps in transformers and the underlying attention mechanisms.
I suggest reading …wikipedia.org/…/Transformer_(deep_learning_archi…
Blue_Morpho@lemmy.world 3 weeks ago
The model is data. It needs to be operated on to get information out. That means lots of JMPs.
If someone said viewing a gif is just a bunch of if-else’s, that’s also true. That the data in the gif isn’t itself a bunch of if-else’s isn’t relevant.
Executing LLM’S is particularly JMP heavy. It’s why you need massive fast ram because caching doesn’t help them.
elbarto777@lemmy.world 3 weeks ago
I’ll be pedantic, but yeah. It’s all transistors all the way down, and transistors are pretty much chained if/then switches.
Hotzilla@sopuli.xyz 3 weeks ago
Calling these new LLM’s just if statements is quite a over simplification. These are technically something that has not existed before, they do enable use cases that previously were impossible to implement.
This is far from General Intelligence, but there are solutions now to few coding issues that were near impossible 5 years ago
5 years ago I would have laughed in your face if you came to suggest that can you code a code that summarizes this description that was inputed by user.
JcbAzPx@lemmy.world 3 weeks ago
I think the point is that this is not the path to general intelligence. This is more like cheating on the Turing test.