Stable Diffusion, the company behind the popular AI image generator, recently announced the launch of the open-source chatbot StableLM – a viable rival for OpenAI’s ChatGPT. This small yet efficient model utilises only 3 – 7 billion parameters, making it 2-4% smaller than ChatGPT’s 175-billion-parameter model. The goal of StableLM was to make access to advanced language models accessible to all and guarantee access to these advanced technologies.
StableLM has undergone further refinement by Stability AI, notably by training on a new, experimental data set from EleutherAI, calle “The Pile”. This data set contains 1.5 trillion tokens which is three times the size of the data set used to train most AI models. However, ChatGPT underwent reinforcement learning refinement processes to reduce flawed results.
All features of StableLM are free and open source, just as the image generator Stable Diffusion. The newest version, StableLM Alpha, can work on lower-cost and less powerful GPUs and is accessible online through the Huggins Face demo page. Although with the alpha version the results are yet to be refined, the platform has the potential to compete with ChatGPT’s advanced results.
Stability AI also has larger models with 15 billion, 30 billion, and 65 billion parameters in the works that may help to further improve the capabilities of StableLM. In the future, the StableLM team is planning on creating a 175 billion parameter model.
Stability AI launched a website to provide community support to chatbot developers, researchers, and hobbyists who are exploring open-source chatbot development. Here users can get resources, tutorials and news related to chatbot development, as well as technical support and bug-reporting tools.
The development of StableLM offers much potential for advanced language models to be used regardless of budget and reminds us that when it comes to groundbreaking AI, the possibilities are endless.
