Bard is an experimental Google chatbot that is powered by the LaMDA large language model.
It’s a generative AI that accepts prompts and performs text-based tasks like providing answers and summaries and creating various forms of content.
Bard also assists in exploring topics by summarizing information found on the internet and providing links for exploring websites with more information.
Bard is powered by a “lightweight” version of LaMDA.
LaMDA is a large language model that is trained on datasets consisting of public dialogue and web data.
There are two important factors related to the training described in the associated research paper, which you can download as a PDF here: LaMDA: Language Models for Dialog Applications (read the abstract here).
A. Safety: The model achieves a level of safety by tuning it with data that was annotated by crowd workers.
B. Groundedness: LaMDA grounds itself factually with external knowledge sources (through information retrieval, which is search).
The LaMDA research paper states:
“…factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator.
We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible.”
Google used three metrics to evaluate the LaMDA outputs:
The LaMDA research paper concludes by stating that crowdsourced reviews and the system’s ability to fact-check with a search engine were useful techniques.
