One of the more interesting developments recently is that you can now run the LLaMA-3 LLM locally on your computer's GPU. There are a few interesting takeaways from that. The environmental impact is eliminated (if you were using your computer anyway, that is.) You can also ask it sensitive questions regarding your proprietary intellectual property without risk of leaking it onto the internet. It boggles my mind that this is possible. The reason it can happen is they have reduced the differentiation algorithm down to one byte (aka 8 states of confidence.) So to, say, distinguish between a cat and a dog, there are now only eight choices:
DOG
almost definitely a dog
probably a dog
LLaMA-3 thinks it is a dog
LLaMA-3 thinks it is a cat
probably a cat
almost definitely a cat
CAT
If you can live with that reduced level of nuance, you can get some pretty astounding answers right from your own computer. For example, my team asked LLaMA-3 a few simple questions and learned how we could save thousands of dollars simply by switching which clinic location doctors work from.
Deliver high value solutions to customers - at Juniper Networks
3moMonsieur l’ambassadeur 😉