Google has asked staff to test and improve its internal chatbot, Bard, before it’s launched to the public, but many employees are already playing with a superior version.

Googlers built “Big Bard,” as the company refers to it internally, on the same language model as Bard, known as LaMDA, but with seemingly larger parameters that make it appear more intelligent.

Insider viewed examples of users asking both versions similar questions, and Big Bard produced richer and more humanlike responses. It’s also generally chattier, more informal, and prone to swearing.

When Google revealed Bard, Sundar Pichai, the CEO of Google, said the first version to be released to the public would be based on a “lightweight model” of LaMDA, which would require less computing power. Big Bard appears to be a preview of what a more advanced version of the chatbot might look like.

Google might want to launch a limited version of Bard because of cost. John Hennessy, the chair of Google’s parent company Alphabet, recently said a search using AI could potentially cost Google 10 times more than a normal key-word search. 

Google is also wary of any public snafu that could generate bad headlines, such as inaccurate answers or rogue behavior, which a more limited initial launch may help prevent. 

When employees test out Big Bard, they’ll see a note steering them to the regular version of Bard for testing, and some leaders have asked employees to only work with the standard model, one employee said. An earlier version, tested internally, was named, “Apprentice Bard.”

The AI race shows no sign of slowing. Andreas Braun, the CTO of Microsoft Germany, just announced that OpenAI will release the next version of its language model, GPT-4, next week. GPT-4 will be multimodal, meaning it will also encompass video, images, and audio.

A Google spokesperson did not respond to a request for comment.

Got a tip? You can reach him via encrypted email (hlangley@protonmail.com) or encrypted messaging apps Signal/Telegram (+1 628-228-1836).