Company attributes
Other attributes
Stability AI is a developer of an open AI model for images, languages, audio, video, 3D, and biology. These open AI tools are created to help developers design and implement solutions through collective intelligence and augmented technology. Stability AI uses a cluster of more than 4,000 Nvidia A100 GPUs in AWS, which the company uses to train AI systems, including the company's Stable Diffusion. Stability AI CEO Emad Mostaque has suggested it should be capable of training models more efficiently in the future, despite the cost of running this compute environment.
Stability AI's Stable Diffusion is an example of a text-to-image AI. It is an open-source software, differentiating it from other text-to-image AI systems, allowing a community of developers to build on the company's code or power their own commercial offerings. Stability AI offers a commercial version of Stable Diffusion, called DreamStudio, which works to generate revenue by developing the underlying infrastructure and customizing versions of the software for corporate clients.
The DreamStudio website is designed to enable anyone to access creative tools without the need for software installation, coding knowledge, or heavy-duty local GPUs, but comes with a cost. The cost is in DreamStudio credits, in which users are charged one credit per image. Depending on the image resolution and step count users choose, the cost-per-image can go as low as 0.2 credits to as high as 28.2 credits per image. First-time users are offered free credits, and once those credits run out, the user will have to purchase more.
In March 2023, Stability AI announced the acquisition of the ClipDrop application and the Init ML team behind the development of the tool. As part of the deal, the founders of Init ML and ClipDrop joined Stability AI to continue developing the ClipDrop tool. The acquisition of ClipDrop further expands Stability AI's generative AI imaging tools.
StableVicuna is the first large-scale open-source chatbot trained via reinforced learning from human feedback (RHLF). StableVicuna is based on the Vicuna chatbot, a 13-billion parameter LLaMA model. The chatbot utilizes the three-stage RLHF pipeline outlined in the papers Steinnon et al. and Ouyang et al. and is further trained with supervised finetuning (SFT) using a mix of three datasets:
- OpenAssistant Conversations Dataset (OASST1), a human-generated, human-annotated assistant-style conversation corpus comprising 161,443 messages distributed across 66,497 conversation trees, in thirty-five languages
- GPT4All Prompt Generations, a dataset of 437,605 prompts and responses generated by GPT-3.5 Turbo
- Alpaca, a dataset of 52,000 instructions and demonstrations generated by OpenAI's text-davinci-003 engine.
Stability AI trained a reward model using trlX for further SFT on the following RLHF datasets:
- OpenAssistant Conversations Dataset (OASST1) contains 7213 preferences samples
- Anthropic HH-RLHF, a dataset of preferences about AI assistant helpfulness and harmlessness containing 160,800 human labels
- Stanford Human Preferences (SHP), a dataset of 348,718 collective human preferences over responses to questions/instructions in 18 different subject areas, from cooking to philosophy
Finally, trlX was used to perform Proximal Policy Optimization (PPO) reinforcement learning to perform RLHF training of the SFT model. Stability AI states the chatbot can perform simple math, text generation, and write code. In common benchmarks, StableVicuna demonstrates similar performance compared to previously released open-source chatbots. StableVicuna was released on April 28, 2023. Upon release, users can access a demo of the chatbot on the HuggingFace hub, with plans to provide StableVicuna through a chat interface in the future.
