Da_Best_Gamer123 wrote:

AndThenWat wrote:

what kind of AI? the old stuff like NPCs or actual AI?

Actual AI

horse_development wrote:

Da_Best_Gamer123 wrote:

AndThenWat wrote:

what kind of AI? the old stuff like NPCs or actual AI?

Actual AI

I don't think generative AI is possible in scratch, sorry

when green flag clicked
if <> then 

  

else
if <> then

end
end

pio10657 wrote:

here is an example.

So, i made it once well it turned out to work but you have to say the words in order… WELL i tried THIS,

when green flag clicked
if <> then 

  

else
if <> then

end
end

AND IT WORKED! well anyways we have limited answers because OF COURSE, Scratch does NOT have an API. MAYBE JUST ADD MORE ANSWERS LOL

here is the game.

dem_bot wrote:

A language model:
So what you do is you make a neural network and then you find the partial derivative of the loss function and then you multiply the gradient by the learning rate and change every weight by that until there's no more improvement.

A less efficient version is where you use the loss of a model to determine its fitness, then pick the best performing models and mutate or fuse them to make new models that have a chance to have a higher fitness.

The third option is the Markov chain, where no conventional training is required, but instead you look at how many times certain tokens occur in certain orders and make a graph with the chances for each state to flow to the next state. This is relatively lightweight and more suitable on scratch, but generally gets an effect similar to spamming the suggested words on a phone keyboard.

LLMs have billions of parameters and are trained on data centres, so don't expect your home brew to get anywhere close to chatgpt.

An image generator:
The way those are generally trained is by taking an image with a description and adding random noise to it and training the AI to filter out the noise. You gradually increase the amount of noise, until the entire image is noise and the AI is generating an image instead of filtering the noise. This you also do with gradient descend, like the first bit about the language models.
Image generators are also really quite big, so it'll already be quite an accomplishment if you could train one purely to make dog images.