Google’s LaMDA to Make Assistant Less Dumb
May. 18, 2021Language Model for Dialogue Applications (LaMDA) is Google’s latest language model built on top of its open-source Transformer neural network architecture.What makes LaMDA different from existing language models such as BERT and GPT-3 is the fact that it is trained on dialogue.
Google says LaMDA picked up on“several of the nuances that distinguish open-ended conversation from other forms of language”during the training process. This is really promising as we often get irrelevant responses from Assistant if it loses the context.
However, there’s a catch. Like most AI systems, LaMDA isn’t perfect just yet. Google even showed a slide where it managed to mess up the conversation and you can check that out below.
Google is also working on multimodal models that utilize images, text, audio, and videos. When this becomes a reality, interacting with Assistant should be a seamless experience. For example, you can just ask Google to find a route with beautiful mountain views for your road trip. You caneven search for specific portions of a video with voice commands. How cool is that!
With natural conversations as the key focus, LaMDA should theoretically help significantly improve Assistant and might as well change the way we interact with it. Unfortunately, LaMDA is not immediately rolling out to current Google products just yet.
The companyplans to use LaMDA on many of its products, including Google Assistant, Search, and Workspacein the future. Google is currently focusing on making LaMDA’s responses both compelling and factually right before leveraging its capabilities.
Subin writes about consumer tech, software, and security. He secretly misses the headphone jack while pretending he’s better off with the wireless freedom.
