Philip Wang, the mastermind behind the reverse-engineering of several closed-source AI systems like DALL-E 2, AlphaFold, and Imagen, has just released an implementation of PaLM + RLHF, a text-based AI model that works just like ChatGPT.
This system combines Google’s PaLM, an LLM (large language model) with 540B parameters (3x more parameters than ChatGPT), and a method commonly known as Reinforcement Learning with Human Feedback (RLHF) to allow for the creation of a chatbot that can do just about anything ChatGPT can, including answering general questions, writing emails, and creating computer code.
Note that this version isn’t exactly the same as the PaLM model developed by Google a few years ago, but it has a very similar architecture and approach. Wang is well-known for being able to “port” several other famous architectures, hence, he might know a thing or two about emulating these models.
The Power of PaLm + RLHF
Since its release, ChatGPT, a fine-tuned version of GPT-3.5, has taken the tech world by storm due to its ability to generate human-like text with high clarity, allowing it to respond in a conversational manner. While it may be a significant advancement from earlier chatbots, many proponents in the AI field have raised concerns over the closed nature of ChatGPT.
As of today, the ChatGPT model remains proprietary, which means that the public cannot view its underlying code. Only OpenAI truly knows how it works and what data it processes. This lack of transparency can have far-reaching implications and may affect trust from users in the long term.
Many developers have been eager to build an open-source alternative and now it has finally arrived. PaLM + RLHF is built exclusively for the Python language, and can be implemented for PyTorch. Developers can easily train PaLM like an autoregressive transformer and then train the reward model using human feedback.
What’s The Catch?
Well, you can’t use it today yet. So what would it take for it to be available to the public?
To launch PaLM + RLHF, you will need to compile gigabytes of text taken from various sources such as blogs, social media posts, news articles, e-books, etc. These data are fed to the fine-tuned PaLm model, which will generate several responses. For example, queries like “what are the basics of Economics,” might yield responses like “Economics is the social science that studies…”
Human volunteers will be employed to rank those responses from best to worst, using the rankings to create a reward model that takes the original model’s responses and sorts them in order of preference, filtering for the top answers to a given prompt.
However, the process of aligning this model with what users want to accomplish with ChatGPT is both costly and time-consuming, as PaLM has a massive 540B parameters. Note that the cost of building a text-based LLM model with only 1.5 billion parameters can reach up to $1.6 million.
When Can I Use an Open-Source ChatGPT?
At the moment, it is unclear how many organizations have the technicals and financial infrastructure to run the implementation built by Wang without degrading their user experience. So far, we have three known players working on this open-source ChatGPT alternative:
- CarperAI (in partnership with Hugging Face, Scale AI, and EleutherAI)
- LAION – the non-profit that supplied the dataset used to train Stable Diffusion
- Yannic Kilcher
None of them have given any hard dates for release as of yet. But to give you a rough timeline, it took three months to train Bloom, an open-source model that boasts 176 billion parameters. With that in mind, it might take more than 6-8 months for us to see a worthwhile release.
As Wang AKA Lucidrains says, “good code is just a prerequisite to begin the journey. It will take data, compute, [and ]adventurers to actually set sail…”
For now, all we can do is wait. In the meantime, it might be a great idea to continue using ChatGPT whilst it’s free. It’s also worth noting that by the time an open clone exists, OpenAI might be far ahead in development. There are many rumors of GPT-4 being ultra-powerful and is set for release in 2023.
Join our newsletter as we build a community of AI and web3 pioneers.
The next 3-5 years is when new industry titans will emerge, and we want you to be one of them.
Benefits include:
- Receive updates on the most significant trends
- Receive crucial insights that will help you stay ahead in the tech world
- The chance to be part of our OG community, which will have exclusive membership perks
