Disclaimer: This article is the Chinese translation of the article "ChatGPT Equivalent Is Open-Source, But it Is of No Use to Developers" by Preethi Cheguri.
Original link: https://www.analyticsinsight.net/chatgpt-equivalent-is-open-source-but-it-is-of-no-use-to-developers/
Software similar to ChatGPT is now open source, but it seems useless for developers
The first open source software similar to ChatGPT has appeared: this is a large language model PaLM architecture based on Google's 540 billion parameters, and the use of RLHF (Reinforcement Learning from Human Feedback, that is: using reinforcement learning methods, using human feedback Signals directly optimize language models) applications. "PaLM + RLHF" is a replica of ChatGPT and is now open source. It is a text similar to ChatGPT developed by developers responsible for reverse engineering closed source AI systems (such as Make-A-Video released by Meta). Generate a model. To build a system that can do almost any task on ChatGPT, including drafting emails and code hints, the system combined Google's large language model PaLM with the method of reinforcement learning with human feedback (RLHF).
Why is this "Open Source ChatGPT" useless for developers?
"PaLM + RLHF" is not pre-trained, in other words: the system has not been trained on the example data from the network necessary to make it actually work. The experience after downloading "PaLM + RLHF" is not the same as when using ChatGPT, "PaLM + RLHF" needs to generate gigabytes of text that the model can learn, and then find hardware that can handle the training requirements. This is a very expensive process, and "PaLM + RLHF" cannot currently replace ChatGPT unless there is a well-funded enterprise (or individual) dedicated to research and making it available to the public.
The good news is that several other projects that fork ChatGPT are growing fast, including one run by research group CarperAI. CarperAI will partner with OpenAI research group EleutherAI, startups Scale AI and Hugging Face to release the first "ChatGPT-like ready-to-use AI model" trained on human feedback. The non-profit organization LAION spearheaded a project to replicate ChatGPT using state-of-the-art machine learning methods, and LAION also provided the initial dataset for training Stable Diffusion. So what can PaLM applications using RLHF do? As the model scales up, performance across activities continues to improve, and new opportunities will emerge. PaLM scales to 540 billion parameters, compared to about 175 billion parameters for GPT-3.
"ChatGPT" and "PaLM + RLHF"
Reinforcement learning with human feedback is an approach aimed at better aligning language models with what users expect, and is the secret sauce that both ChatGPT and "PaLM + RLHF" have. The RLHF needs to fine-tune a language model using a dataset that contains cues that match what human volunteers expect the model to say, e.g.: prompt "Explain machine learning to a 6-year-old"; answer "Machine learning is a type of AI form……".
The PaLM language model is used in "PaLM + RLHF". After feeding the above cues into the improved model, multiple responses are generated, the volunteers rank each response from best to worst, and use this ranking to train a "reward model" that takes the original model's responses and Sort them in order of preference, then filter out the best answer for a given prompt. The process of collecting training data is very expensive.
Also, the training process is not cheap. PaLM has 540 billion parameters/language model components, all learned from the training data. And a 2020 study suggested that developing a text generation model with only 1.5 billion parameters could cost as much as $1.6 million. It took three months to train the open-source model Bloom with 176 billion parameters, using 384 Nvidia A100 GPUs (each costing thousands of dollars). And running a trained model of "PaLM + RLHF" size is also not trivial. Bloom requires a dedicated PC with about eight A100 GPUs. The cost of running OpenAI's text-generating GPT-3 (which contains more than 175 billion parameters) is estimated to be about $87,000 per year on a single Amazon Web Services instance.
Conclusion: "PaLM + RLHF" is currently unable to replace ChatGPT unless there are well-funded companies (or individuals) investing in training and making it public.
Disclaimer: The information presented in this article represents the opinion of the author/advertiser only and is not investment advice - it is for educational purposes only. By reading this article, the information herein does not constitute any investment or financial advice from Analytics Insight and its team. Investors should seek their own independent financial or professional advice. Please do independent research with a financial advisor before making any investment decisions. Analytics Insight and the team are not responsible for the investment opinions presented in this article.
Open Source Rainforest builds a knowledge system around the three aspects of open source general knowledge, open source use, and open source contribution. It is willing to systematically share long-term accumulated experience with enterprises, provide cooperation in teams, mechanisms, and projects, and promote enterprises to use open source more efficiently. , Contribute to open source, improve the level of open source technology and application in the whole industry.
The content of Open Source Rainforest has been open-sourced and hosted at https://github.com/opensource-rainforest . You are welcome to contribute content in the form of Pull Request, discuss in the form of Issue, and jointly maintain the content of Open Source Rainforest.
Welcome to pay attention to the "Open Source Rainforest" public account to get the latest and most complete news.