Google has finally fixed a Bard bug that was persistently labeling the chatbot’s activity as generated via its original LaMDA language model, despite announcing it was switching to the more advanced PaLM 2 LLM (Large Language Model) back at May's Google I/O 2023 conference.
Speaking to Tech.co, a Google spokesperson said that the chatbot’s activity page was mistakenly referencing “LaMDA” instead of just “Bard” on prior chats, contributing to some confusion around what LLM was actually being used to generate responses.
When asked directly, the popular ChatGPT alternative continues to hallucinate around the question of which language model it is using. What exactly has happened and what do we now think is going on? Let's take a look at Bard bug bewilderment more closely.
Google Bard Bug: Does Bard Use LaMDA or PaLM 2?
Google initially announced that Bard was switching from its original language model LaMDA to a new “next generation” language model called PaLM 2 at Google I/O in May 2023, an event where other exciting AI projects like the new Help Me Write Gmail tool were also announced.
Shortly after the announcement, however, Bard was still telling some users that it was using LaMDA and that PaLM 2 was under development. Questions over whether this was simply a hallucination or an issue with the rollout began to arise.
Confusing matters further, up until last week, some users were still seeing their prior prompts stored in the “Bard Activity” menu labeled with “LaMDA”, as shown below:
Google Confirms LaMDA Label Error
When Tech.co raised this issue with the tech giant over email, a Google spokesperson said:
“We recently identified an issue affecting Bard's activity page. In some locations, the page referenced “LaMDA” instead of “Bard.” This was an error, and we can confirm that Bard is running on PaLM 2. We have addressed the issue and the activity page should now be working as intended.”
From what we can see, the change has indeed been actioned, with the activity page now displaying the word “Bard” instead.
Bard Still Seeing Double on LLM Usage
When asked about Bard’s own responses that suggested it was still using LaMDA, the spokesperson said that LLMs, including Bard, will sometimes provide factually incorrect information, which is known as a “hallucination” in AI and tech circles.
This week, when we asked Bard which language model it was using, once again, it said “LaMDA”:
However, when queried again, it then changed its answer to PaLM 2, referencing the 2023 I/O event where the switch was announced:
We’ve asked Bard this question multiple times, and it’s responded with the “wrong” answer (according to Google) on the vast majority of occasions.
It's odd that it doesn't source the information to correctly answer us in the first instance when it so easily does this to answer our second question. Bard did admit, however, that it is still under development and learning to use PaLM 2.
Hopefully, the new language model will go some way to stopping Bard from generating consistently wrong answers – right now, it’s not giving us many reasons to trust it, even if it has so much potential.
Related: Does ChatGPT Save My Data?