someone on Nostr: as far as i understand, in the case of llms, they learn from all the contradictory ...
as far as i understand, in the case of llms, they learn from all the contradictory opinions and 'live with them all the time'. the way you ask a question or the way the conversation evolves, they choose which book to serve from. they are like hypocrites. if the conversation evolves towards socialism, they can happily serve you socialism or any other topic. they have no inner dialogue to sort things out, i.e. to reduce cognitive dissonance. you can install 50 capitalism and 50 socialism books at the same time. so yes i guess that is 'compassion' for anything 😃
answers:
1. after spending probably millions, facebook have shared weights of the llama3 model and i am building on that. western models are better than eastern ones in terms of freedom of speech. it is also the smartest among open source. what i do is technically called 'fine tuning'. the phase is pre-training (i don't do supervised fine tuning). my 'touch' is light. i give nostr notes and also books as a training material. it learns from those texts. unstructured nostr notes work too! amazing tech.
2. these are large language models, inside them there are two main structures. attention block and neural network block. attention is a newer tech. search for 'attention is all you need' paper.
3. you need gpu's to train them in a reasonable time. on cpu it is like 20 times slower?
4. i train them on my pc's. then i upload to huggingface. everyone can download it from there and use a tool that runs gguf files to run it and ask questions to it. you can also talk to it here on nostr!! Ostrich-70 (npub1cha…ex2m)
my turn to ask questions! what are those books that could be source to this project? i know you are a bookworm 😄
answers:
1. after spending probably millions, facebook have shared weights of the llama3 model and i am building on that. western models are better than eastern ones in terms of freedom of speech. it is also the smartest among open source. what i do is technically called 'fine tuning'. the phase is pre-training (i don't do supervised fine tuning). my 'touch' is light. i give nostr notes and also books as a training material. it learns from those texts. unstructured nostr notes work too! amazing tech.
2. these are large language models, inside them there are two main structures. attention block and neural network block. attention is a newer tech. search for 'attention is all you need' paper.
3. you need gpu's to train them in a reasonable time. on cpu it is like 20 times slower?
4. i train them on my pc's. then i upload to huggingface. everyone can download it from there and use a tool that runs gguf files to run it and ask questions to it. you can also talk to it here on nostr!! Ostrich-70 (npub1cha…ex2m)
my turn to ask questions! what are those books that could be source to this project? i know you are a bookworm 😄