Thomas Dietrich
Microsoft has ChatGPT and just invested $10 billion in OpenAI, the company behind the AI chatbot. Google has similar plans but will draw on its internal research: CEO Sundar Pichai introduced Bard, an “experimental conversational AI service”. Bard will become part of Google’s core product, search, and will initially be available to “trusted testers”. The broad rollout is planned “in the coming weeks”.
Instead of typing in search terms to find a webpage, you simply ask the AI. It pulls new information from the web, leverages multiple sources, and distills it into something easy to read. For example, if you want to learn to play a musical instrument, you can use Bard to decide which one to choose. Is guitar easier to learn than piano, how much practice and time does it take – Google provides the info across various websites and Bard forms some easy-to-read paragraphs with the appropriate answers.
The question, of course, is how much one can trust Bard’s answers. For example, the information from the video demo below is incorrect – the JWST did not capture the first-ever image of a planet outside our own solar system. In fact, the Hubble telescope captured the first photograph of an exoplanet.
Bard is based on a simplified version of Google’s LaMDA, short for Language Mode for Dialog Applications. Simplified because less computing power is required to provide answers compared to the full model. If Google tries to scale Bard for millions of searches and users, performance would be a critical issue.
Google wants to send out the first Bard invitations in March
As for accuracy, Google’s in-house team and early testers will provide the necessary feedback. This post from last year outlines key goals, including quality, security, and groundedness, and how Bard achieved them through training and fine-tuning. Google will initially only use its AI technology in its own products. The first invitations to developers, creators and companies will go out as early as March. They can join in and build new things using the Generative Language API, which is at least initially supported by LaMDA. Also, Pichai promised in his blog post that more AI-related announcements are coming soon.
By the way, ChatGPT doesn’t answer the question about the James Webb Space Telescope correctly either. OpenAI’s AI chatbot was trained with old data from 2020 before launching the JWST. Therefore, ChatGPT only talks about what the telescope might discover in the future. Google explicitly points out that Bard uses current data and can therefore also provide current answers.
