Stability AI Releases Open Source ChatGPT-like Language Models
Stability AI has recently open-sourced a suite of text-generating AI models, StableLM. The company claims that its models can generate both code and text, and compete with other systems like OpenAI's GPT-4. StableLM is available on GitHub and Hugging Face platforms. It has been trained on The Pile, a dataset consisting of internet-scraped text samples from various websites like StackExchange, Wikipedia, and PubMed. Stability AI says that it has created a custom training set that increases the size of The Pile by 3x.
The StableLM models are suitable for creating cover letters, epic rap battle songs lyrics, and text generation tasks. They were tuned using a Stanford-developed technique called Alpaca on open source datasets from AI start-up Anthropic. As the size of StableLM models increases, community feedback can improve the quality of responses expected from the models, and enhance the optimization and better data of the models.
The growing popularity of generative AI has led to companies like Nvidia, Meta, and independent groups to release models on par with private APIs like GPT-4 and Anthropic's Claude. Although researchers have criticized the release of open source models, as they can be used to create phishing emails or aid malware attacks, Stability AI argues that it promotes transparency, fosters trust, and allows the broad research and academic community to develop interpretability and safety techniques.
Stability AI, known for its generative AI art tool Stable Diffusion, has faced legal cases for infringing on the rights of millions of artists and developing AI art tools using copyrighted images scraped from the web. Additionally, communities have used the company's tools to generate pornographic celebrity deepfakes and violent depictions. The company is under pressure to monetize its sprawling efforts and generate revenue, but it recently was reported that it is "burning through cash and has been slow to generate revenue."
