Just to disentangle this a bit, GP's post was about Gemini and you're responding about Bard. The two are different things. Gemini hasn't been trained yet; it's the next-gen thing that's currently being worked on (Pichai mentioned it at I/O).
It obviously remains to be seen how good Gemini will be. I just wanted to clear up the apparent confusion.
On the other hand, bard performed much better at a more complex task: I provided fake data spanning six years and a few rows of different categories. I left the sixth year blank. Without prompting it, reformatted the from comma do you limited into a nice table. It also gave a brief summarization and analysis of the trends. When I asked follow-up questions he gave detailed responses that were moderately insightful for anyone that would be familiar with the data domain given a few minutes to review the data set. Finally, I asked it to give a projection of what the values would be for the final column, which I had left blank. It did so, with an explanation of the trend it observed, as well as a description of the linear regression model. It used it and apologized for not being able to give more complex modeling due to the Small data set.
GPT-4 provided answers and analysis that were comparable in quality, but required additional rounds of prompting and clarification.