Welcome to Smallville!

[@microwavecoffee]

Love your videos! Thanks for this Edit: I coded a very basic version of this, with a JavaScript RPG library and ChatGPT for an educational game where the villages teach the player about the blockchain (I don't like crypto but it was required for the hackathon). It was fairly barebones without any complex system in the background. Something that I think is currently really underdeveloped in Computer Science are long-term agents with memory systems. Like the idea of memory creation and memory retrieval is incredibly fascinating to me, but also Cognitive Architectures is an under-researched field. What do you think could be a good benchmark for evaluating long-term efficient memory storage in agents? Models for forgetting memories and retrieving memories efficiently are so interesting to me.

[@Thinkstr]

Hi, thanks for watching! SmallVille is really impressive to me even though I complain about parts of it. Like you, I've tried ChatGPT's API myself, and it's hard to work with! For long-term memory, the first thing that comes to my mind is Catastrophic Forgetting, where the benchmark is in sorting MNIST's handwritten digits. If you train a model to tell the difference between 0s and 1s, it'll do fine... but then if you train the same model to tell the difference between 2s and 3s, it'll forget how to tell the difference between 0s and 1s! So it's important to make the model remember old tasks. I happen to have some video about it! :3 haha. I notice the audio isn't great. I've learned a lot about video-editing along the way. The problem: kzbin.info/www/bejne/roCldnhnn6tnsMk&ab_channel=Thinkstr A solution: kzbin.info/www/bejne/m4fYiGahmbWIrsk&ab_channel=Thinkstr That solution in reinforcement learning: kzbin.info/www/bejne/nKncpaeca7lomaM&ab_channel=Thinkstr

[@microwavecoffee]

@@Thinkstr Thanks a bunch!

[@GMatthew-s6q]

i lasted 5 minutes, time to do anything else.

[@ЕвгенийИванов-ю2б]

Hi! I really like your work and recently wrote a paper about an evolutionary inspired way of neural network gradual complexification. But can't publish that because don't have arxiv endorsement. Maybe you know someone to whom I can address?

[@Thinkstr]

Oh, I might be able to do that! What part of arxiv are you trying to publish in?

[@ЕвгенийИванов-ю2б]

Computer science, artificial intelligence

[@ЕвгенийИванов-ю2б]

​@@Thinkstr I've posted that on GitHub and described the experiments. Tried to reach you through LinkedIn 😊