Stability AI Launches Open Source StableLM to Challenge GPT-4
Stability AI releases StableLM, a suite of open source text-generating AI models designed to compete with proprietary systems like GPT-4. The alpha models demonstrate that smaller, highly efficient AI can still deliver strong performance in generating both code and text.
Stability AI openly releases a suite of text-generating AI models called StableLM to compete directly with systems like OpenAI's GPT-4. Available in alpha on GitHub and Hugging Face, these models generate both code and text while demonstrating how small, efficient models deliver high performance with appropriate training. The company trains the models on an expanded version of The Pile, a dataset that includes internet-scraped text from websites like PubMed, StackExchange, and Wikipedia.
Like many pretrained large language models, StableLM likely suffers from common limitations such as generating toxic responses and hallucinating facts. Because the training data contains profane and abrasive language, the models occasionally produce offensive content without additional fine-tuning. Early testing proves difficult as the Hugging Face hosting page repeatedly shows "at capacity" errors, likely due to high popularity or the computational demands of the models.
Despite these early limitations, the fine-tuned versions of StableLM show impressive capability in following complex instructions. Tuned using a Stanford-developed technique called Alpaca on open source datasets from organizations like Anthropic, these specific models behave much like ChatGPT. They successfully handle prompts that require writing cover letters or composing creative rap lyrics, proving that open source alternatives effectively challenge proprietary AI systems.