The world is experiencing an unprecedented algorithmic explosion, and with it new points of conflict arise. For example, OpenAI is facing criticism for not being "open" at all in its current projects, while other players maintain a media silence that draws a lot of attention. But there are also exceptions like Databricks, an American company working on the Dolly model. What's so special about it? On one hand, it's an open source project, and on the other, it was designed for commercial use.
A word that is being repeated frequently in AI articles is 'democratization'. In fact, that was the expression Databricks used when it presented its Dolly model in March of this year. The parameters of Dolly AI are minimal compared to GPT-3 (6 billion vs. 175 billion), but the idea is that anyone can take an existing model, train it with ChatGPT-like features, and in 30 minutes have something within reach of any company.
Dolly AI 2.0, Open Source and Commercial
Recently, Databricks decided to double down with Dolly AI 2.0. The new version doubles the number of parameters (12 billion) and is based on EleutherAI's pythia model. However, Databricks created a training set derived from demonstrations generated by thousands of employees. Both the dataset, the model weights, and the training code were published under a Creative Commons license, enabling commercial use.
In contrast, OpenAI's model is proprietary, and besides paying for API access, interested parties must also follow its terms of use (read "cannot compete with OpenAI"). Meta's LLaMA model was leaked on the internet in early March, but its commercial use is prohibited. Alpaca, Koala, Vicuna, and GPT4All (which we recently discussed) have similar restrictions.
Anyone interested in evaluating Dolly AI 2.0 can now obtain the weights on Hugging Face, and the databricks-dolly-15k dataset on GitHub. For now there are no online demos or anything like that, but the open source nature of the project will surely give rise to a new round of chatbots and other AI-backed services.
Official announcement: Click here
