Good luck running it on your laptop since it is the first open source equivalent of Openai.
Philip Wang, the developer responsible for reverse-engineering closed-sourced artificial intelligence systems such as Meta's Make-A-Video, released a text-generating model that behaves similar to the one used in chatGPT. The system combines a large language model from Google and a technique called Reinforcement Learning with Human feedback to create a system that can do a lot of things.
It isn't pre- trained. The system hasn't been trained on the data needed for it to work It will take gigabytes of text from which the model can learn and find hardware that can handle the training workload if you download PaLM +RLHF.
PaLM is a statistical tool that predicts words. When fed a lot of examples from training data, PaLM + RLHF learns how likely words are to occur.
Reinforcement Learning with Human feedback is a technique that aims to better align language models with what users want them to do. A language model is trained on a dataset with prompts and what humans expect the model to do. The volunteers rank the responses from best to worst after the prompt is fed to the model. The rankings are used to train a reward model that takes the original model's responses and sorts them in order of preference.
It takes a lot of money to collect the training data. Training is expensive. 540 billion parameters are referred to as the part of the language model learned from the training data. The 2020 study pegged the costs for developing a text-generating model at as much as $1.6 million. It took three months to train the open source model, and a single A 100 cost thousands of dollars.
It isn't trivial to run a trained model of PaLM +RLHF. There needs to be a dedicated PC with at least eight A 100 graphics cards. Back-of-the-envelope math shows the cost of running OpenAI's text-generating GPT3 on a single Amazon Web Services instance to be around $87,000 annually.
The scaling up of the necessary development workflows could prove to be a challenge according to a post by Sebastian. He said that even if someone gives you 500 graphics processing units, you still need a software framework that can handle that. We are developing frameworks to make that simpler but it is still not trivial.
Unless a well-funded venture goes to the trouble of training and making it available publicly, PaLM +RLHF isn't going to replace CHATGPT today.
A research group called CarperAI is leading a group of others that are trying to duplicate the idea. The first ready-to-run artificial intelligence model trained with human feedback is planned to be released by Carperai.
The initial dataset that was used to train Stable Diffusion was supplied by LAION. LAION wants to build an assistant of the future, one that not only writes emails and cover letters but also does meaningful work and uses the internet. It is in the very beginning. A resources page for the project has been live for a few weeks.
All Rights Reserved © 2024
