I think that Scratch chatbots are boring, so can you please add AI? From all the JavaScript I know, just connect it to a neuron network using

import ‘neural-network’
let <name> = neuralNetwork()

, and then train it by either:
1. Putting a database of basic stuff, and then teach it how to speak
2. Telling it questions and answers
3. Getting it to go through loads of projects, see the code, find out what it does, and put it in the database.

NOTE: <name> is supposed to be the AI's name, but it has no spaces.

let mode = ‘TRAINING’

to show that it is training so it knows it is being teached, to stop it saying nonsense while you are teaching.

H__4 wrote:

I think that Scratch chatbots are boring,

What chatbots? Everything commenting on this website is either human or not allowed (with the exception of the chatbot that has been used on @ScratchCat's profile in the past on April Fools' Day, but that's run by the ST and gives nonsense answers anyways).

H__4 wrote:

so can you please add AI? From all the JavaScript I know, just connect it to a neuron network using

import ‘neural-network’
let <name> = neuralNetwork()

(…)

1. You really want to run AI on computers with the processing power of a potato? Most Scratchers aren't using devices that are anywhere near fast enough to do the math required for a neural network running a LLM- or, if they are, in many cases websites can't use the parts of them designed to do it anyways. Popular LLMs (such as ChatGPT) send your prompt off to a server, and the AI's response back to you.
2. It's not that simple to make AI.

H__4 wrote:

I think that Scratch chatbots are boring, so can you please add AI? From all the JavaScript I know, just connect it to a neuron network using

import ‘neural-network’
let <name> = neuralNetwork()

, and then train it by either:
1. Putting a database of basic stuff, and then teach it how to speak
2. Telling it questions and answers
3. Getting it to go through loads of projects, see the code, find out what it does, and put it in the database.

NOTE: <name> is supposed to be the AI's name, but it has no spaces.

let mode = ‘TRAINING’

to show that it is training so it knows it is being teached, to stop it saying nonsense while you are teaching.

This is dangerous because chatbot ChatGPT has cloned itself many times And this could get out of hand for scratch

H__4 wrote:

I think that Scratch chatbots are boring, so can you please add AI? From all the JavaScript I know, just connect it to a neuron network using

import ‘neural-network’
let <name> = neuralNetwork()

, and then train it by either:
1. Putting a database of basic stuff, and then teach it how to speak
2. Telling it questions and answers
3. Getting it to go through loads of projects, see the code, find out what it does, and put it in the database.

NOTE: <name> is supposed to be the AI's name, but it has no spaces.

let mode = ‘TRAINING’

to show that it is training so it knows it is being teached, to stop it saying nonsense while you are teaching.

Hi, by some coincidence I made a snarky (not helpful) AI. You can use it if you give me credit. You can find it here if you are interested.

Connerfurr13 wrote:

(#3)
This is dangerous because chatbot ChatGPT has cloned itself many times And this could get out of hand for scratch

…. was it ever successful? How do you imply this one could do any better?

Anyways, I thought this was gonna be a cool thing for predictions or game AI like a client-side tensorflow extension for Scratch. But specifically for a chatbot? Nah. OpenAI already did all this. They used millions of dollars and tons of GPUs and tons of Energy. I don't think anyone has a computer powerful enough at their house that can even close to making a coherent chatbot. It feels like a waste of time and you would just be reinventing the wheel.

You could make an extension that harnesses something like open Ai api and sends that info back to the project

Dagriffpatchfan wrote:

You could make an extension that harnesses something like open Ai api and sends that info back to the project

Which bills you per request, is pretty biased, and doesn't really have that many filters built-into it?

Probably not the best idea- and how would the ST pay for that anyways?

https://scratch-mit-edu.ezproxyberklee.flo.org/discuss/topic/798324/#:~:text=1.%20You%20really%20want%20to%20run%20AI%20on%20computers%20with%20the%20processing%20power%20of%20a%20potato%3F
I'm using a 2014 dell that can withstand my friends btd6 “Experiments” (Spamming high projectile towers)

OP_GAMER123456 wrote:

https://scratch-mit-edu.ezproxyberklee.flo.org/discuss/topic/798324/#:~:text=1.%20You%20really%20want%20to%20run%20AI%20on%20computers%20with%20the%20processing%20power%20of%20a%20potato%3F
I'm using a 2014 dell that can withstand my friends btd6 “Experiments” (Spamming high projectile towers)

1. Please use the quote button rather than a highlight link URL (which doesn't even work on all browsers).
2. Finding one example of a computer that has a decent GPU does not mean that either the majority of Scratchers have access to one that can, or that it would be able to run AI.

A lot of people using Scratch are doing it on school-issue Chromebooks and iPads, which tend to be lower-end devices with not much storage or RAM.

By contrast, running a neural network requires doing a lot of matrix multiplication- and RAM to store the results. Devices with built-in AI acceleration chips use analog computers to speed this up (and save a lot of power in the process), but there are literal millions (if not billions) of devices that can run Scratch but don't have an AI acceleration chip. Most of the devices running Scratch fall into the category of not having one of those- not that we browsers can use it anyways: AI has evolved so fast that web standards can't keep up, and browsers take around 2 years to fully implement a standard once it becomes standard (although it can take more- in one extreme case, major browsers started to fully implement a feature- except Safari, which finally implemented it in 2023).

BigNate469 wrote:

(#2)
1. You really want to run AI on computers with the processing power of a potato?

AI can run from the server-side. For example, whenever i go to a seller's website(for example, Lenovo.com) there is probably a support chatbot. it does all of its neural processing from the server-side and sends outputs to the user. however, locally-hosted AIs can still work decently on potatos. My laptop is the definition of potato and it still runs applications like MicroSoft Designer with decent efficiency. Now, there are worse PCs to own, so i cant say anything for them

cool-shop-dude wrote:

BigNate469 wrote:

(#2)
1. You really want to run AI on computers with the processing power of a potato?

AI can run from the server-side. For example, whenever i go to a seller's website(for example, Lenovo.com) there is probably a support chatbot. it does all of its neural processing from the server-side and sends outputs to the user. however, locally-hosted AIs can still work decently on potatos. My laptop is the definition of potato and it still runs applications like MicroSoft Designer with decent efficiency. Now, there are worse PCs to own, so i cant say anything for them

I'm on a Chromebok that has issues running most everything that uses OpenGL or WebGL to render 3D graphics (such as almost all 3D webpages).

Unfortunately, this Chromebook is also representative of most Chromebooks being used, especially in schools (where a lot of people use Scratch).

Anyways, yes, you could do that. It's just that in the OP it implied that the AI would run client-side, which isn't going to happen.

Also, a majority of apps that have AI capability run that AI server-side, because not enough computers have hardware acceleration for AI yet, so they can't rely on it, especially for cross-platform apps.

Quote from Bignate469:
I'm on a Chromebok that has issues running most everything that uses OpenGL or WebGL to render 3D graphics (such as almost all 3D webpages).

Unfortunately, this Chromebook is also representative of most Chromebooks being used, especially in schools (where a lot of people use Scratch).

About schools, they usually don't have those Go Force and acceleration devices, my school has not got any micro:bits, We haven't got those lego stuff, and they are still extensions. So why not ai?

H__4 wrote:

Quote from Bignate469:
I'm on a Chromebok that has issues running most everything that uses OpenGL or WebGL to render 3D graphics (such as almost all 3D webpages).

Unfortunately, this Chromebook is also representative of most Chromebooks being used, especially in schools (where a lot of people use Scratch).

About schools, they usually don't have those Go Force and acceleration devices, my school has not got any micro:bits, We haven't got those lego stuff, and they are still extensions. So why not ai?

My point isn't that schools don't often have what it takes to use an extension- my point was that a lot of devices using Scratch don't.

And to be fair, my school district owns a few hundred micro:bit v2s. Some schools definitely have the equipment.

BigNate469 wrote:

(#13)

And to be fair, my school district owns a few hundred micro:bit v2s. Some schools definitely have the equipment.

my school doesnt have a single one, or anything cool and techy for that reason.

they do have decent windows 10 laptops with 4gb ram, but when they break all the “tech man” does is turn them off and on again and if that doesn't work, straight to the trash. except for a few known issues, where he just uses a default wizard

You likely would have to use the OpenAI api, which costs money, or run a model like qwen2.5 72b on hugging face, which unless you want like a word per second response, you would probably have to get a $10,000 GPU, or run a smaller model like qwen2.5 7b, that you might be able to run off a normal desktop. Sooooo….. I've been trying to have AI on scratch, and I just figured out that you need a suuuper good computer… ): I haven't tried a 7b model yet though.

I personally agree, it can use GPT 4

_Code-Crafter_ wrote:

You likely would have to use the OpenAI api, which costs money, or run a model like qwen2.5 72b on hugging face, which unless you want like a word per second response, you would probably have to get a $10,000 GPU, or run a smaller model like qwen2.5 7b, that you might be able to run off a normal desktop. Sooooo….. I've been trying to have AI on scratch, and I just figured out that you need a suuuper good computer… ): I haven't tried a 7b model yet though.

a 7b model is still 7 billion. It runs at like 1tk/s on my slow CPU only laptop

What if scratch has an openAI account to communicate with the website when prompted?

BigNate469 wrote:

H__4 wrote:

I think that Scratch chatbots are boring,

What chatbots? Everything commenting on this website is either human or not allowed (with the exception of the chatbot that has been used on @ScratchCat's profile in the past on April Fools' Day, but that's run by the ST and gives nonsense answers anyways).

the makeshift chatbot's that use premade responses

_Code-Crafter_ wrote:

You likely would have to use the OpenAI api, which costs money, or run a model like qwen2.5 72b on hugging face, which unless you want like a word per second response, you would probably have to get a $10,000 GPU, or run a smaller model like qwen2.5 7b, that you might be able to run off a normal desktop. Sooooo….. I've been trying to have AI on scratch, and I just figured out that you need a suuuper good computer… ): I haven't tried a 7b model yet though.

I know the cost is high, but I know how to run ollama on any computer, and while it is free and works, has no gui, but I can fix that in five minutes using docker, or a few weeks, making my own gui, to put in Scratch, and we could easily buy some servers for it, as it is lightweight, so it is pretty cheap. Since it will be running on servers, you only need an internet for the ai, that's it. So it is pretty cheap