May 15, 2025
Trending News

“Years old, easy to train, much smaller but just as smart”: Databricks reveals accessible ChatGPT competitor

  • March 27, 2023
  • 0

Databricks built an AI system with features reminiscent of ChatGPT, but used a two-year-old open-source model and a single computer instead of a full data center for training.

“Years old, easy to train, much smaller but just as smart”: Databricks reveals accessible ChatGPT competitor

Databricks built an AI system with features reminiscent of ChatGPT, but used a two-year-old open-source model and a single computer instead of a full data center for training.

OpenAI used billions from Microsoft, a cloud supercomputer with tens of thousands of GPUs and a complex neural network with 175 billion parameters to develop ChatGPT. “Overkill” is what Databricks says. The company built an AI system, Dolly, that behaves similarly to ChatGPT, but based on an ordinary computer and a two-year-old open-source model with nearly 6 billion parameters.

Just like ChatGPT, Dolly can generate answers to your questions, allowing you to give a tip on a specific topic, generate a promotional tweet, or write a love letter. Databricks believes Dolly is a breakthrough that can bring generative AI models to a wider audience faster.

Much less effort, similar results

Dolly’s foundation is a two year old open source LLM model created by EleutherAI. Databricks modified this a bit to follow instructions and thus create the association between the “Hard Drive X or Y” input and the output based on this question. Armed with this framework, Databricks trained the model on a small but highly curated dataset. The company did this in three hours on a single machine.

The results surprised Databricks themselves. The model was able to compete with ChatGPT very quickly. It understands questions and instructions and generates useful and meaningful text based on these questions. According to Databricks, the success indicates that the breakthroughs of OpenAI and the like are not only due to the increased complexity of the models used, but also to the quality of the data sets.

Databricks believes that with limited training and based on an open-source model, Dolly works so well that it shows promise for wider adoption of AI. Finally, it opens the door for companies to train models on their own data, without entrusting data to a cloud provider or spending a small fortune on AI hardware in the cloud. Databricks shares a Github project where you can get started with Dolly yourself.

Source: IT Daily

Leave a Reply

Your email address will not be published. Required fields are marked *