Skerret wrote:look just watch AI, that will explain most of it.
chat-gpt wrote:Application mostly. There's not necessarily a distinction between the two (AI and software) as you have enterprise software with ML algorithms incorporated into it to automate some aspects or expedite bug fixes as the ML portion 'learns' from the data set it's designed to process.
If a tool is able to - or has a component that can - train on/learn from a data set and then feed those learnings back into a system for predictive purposes, it's AI-ish. You can train a machine learning algorithm on Kow posts, write the code to get it to post the responses it generates to the forum and then you can happily retire once it has warmed up. I doubt you'll notice the difference. The real Skerret has been dead for years.
It's not 'real' AI in the sense that each application of it is really walled garden. A true AI that can learn about the world and act independently needs a vast, diverse data set and the means to sort it into some kind of ontological framework.
Kow wrote:Everyone and his uncle seems to be bandying about the term AI these days, I'm getting Homer Simpson vibes from it. What distinguishes something as being AI rather than a normal piece of software that calculates, or decides some stuff etc according to an algorithm or whatever, like we've had for decades?
acemuzzy wrote:I worry like that captcha is basically how to crowd source answers to "is this a hotdog or not"
b0r1s wrote:For your project you’re gonna be best using your RTX card no? That way you keep whatever it learns local. Hugging face has you covered for any models you need I would have thought. Feed it all Badger’s post data and you can then analyse if they are fanboys or not. You could call it ChatShitGPT.
I preferred my vague and hand wavy explanation betterSpaceGazelle wrote:Kow wrote:Everyone and his uncle seems to be bandying about the term AI these days, I'm getting Homer Simpson vibes from it. What distinguishes something as being AI rather than a normal piece of software that calculates, or decides some stuff etc according to an algorithm or whatever, like we've had for decades?
Traditional software is written for all outcomes so those outcomes have to be kept simple or you're writing a possible infinite lines of code.
if SOMETHING:
DO THIS
else:
DO THIS INSTEAD
That's a conditional statement. There's also loops.
while x < 5:
DO SOMETHING AND ADD 1 TO x
Amazingly, with variables, loops and conditionals you can calculate anything that can be calculated.
ML doesn't work like that and the solutions are not made by the coder.
What you do is make a model of nodes that are all connected together. You only really need to specify how many nodes are in each layer and how many layers there are. It's not an exact science and you can be a bit random in how many nodes the model has and it'll still work out pretty good because it'll still find the best solution for any particular model.
Say you wanted to build an Is it a hotdog? The input layer is the pic of the hotdog, which is just a series of RGB numbers that represent the image. Each hidden node has a value (just a number) that you randomise at the start and each arrow has a weight, also randomised.
What you want is to do in this hotdog case is transform the RGB numbers (input layer), via all the hidden layers and nodes, into an output layer that only has 2 nodes - a yes or no. So you can have random hidden layers but you will specifically set the output layer to have 2 nodes.
The you just start training it. Because the weights and values are essentially randomised it will spit out garbage. You feed it the first pic and those values go through the matrix and the output layer reads something like 0.45, 0.55. If the first number represents the chances of it being a hotdog and the second the chances it isn't, then this is saying it's got a slightly bigger chance of not being a hotdog. But it hasn't learnt anything yet, it's just random.
So you have to tell it, NO THIS IS A HOTDOG YOU FOOL. I need the first number to be bigger! So it basically uses differentiation to tweak all the values and weights slightly so as to make the first number slightly bigger and hence the second number smaller for that image (they should add up to one).
And you keep doing this, image after image. Eventually, as if by magic it'll tweak all the numbers of all the nodes and arrows in a way so the output is always correct, or correct most of the time.
Then it'll start showing things like 0.99, 0.01 - which means it has a very high chance of it being a hotdog. Or 0.25, 0.75 - a fairly high chance it's not a hotdog.The more you train it the more extreme the numbers will get and a perfect match would be 1, 0 - deffo a hotdog, but you'll never quite get to that standard.
And that's why it's mysterious. All these values and weights are incomprehensible and complex because they're all linked together in this matrix. The computer will find patterns we can't. Actually in this hotdog case we can, we just look at the pic, but then again the neurons in our brains are linked up like in the ML model.
It looks like you're new here. If you want to get involved, click one of these buttons!