Test-driving Google’s Gemini-Exp-1206 model in data analysis, visualizations

One of Google’s latest experimental models, Gemini-Exp-1206, shows the potential to alleviate one of the most grueling aspects of any analyst’s job: getting their data and visualizations to sync up perfectly and provide a compelling narrative, without having to work all night.

Investment analysts, junior bankers, and members of consulting teams aspiring for partnership positions take their roles knowing that long hours, weekends, and pulling the occasional all-nighter could give them an inside edge on a promotion.

What burns so much of their time is getting advanced data analysis done while also creating visualizations that reinforce a compelling storyline. Making this more challenging is that every banking, fintech and consulting firm, like JP Morgan, McKinsey and PwC, has unique formats and conventions for data analysis and visualization.

VentureBeat interviewed members of internal project teams whose employers had hired these firms and assigned them to the project. Employees working on consultant-led teams said producing visuals that condense and consolidate the massive amount of data is a persistent challenge. One said it was common for consultant teams to work overnight and do a minimum of three to four iterations of a presentation’s visualizations before settling on one and getting it ready for board-level updates.

A compelling use case for test-driving Google’s latest model

The process analysts rely on to create presentations that support a storyline with solid visualizations and graphics has so many manual steps and repetitions that it proved a compelling use case for testing Google’s latest model.

In launching the model earlier in December, Google’s Patrick Kane wrote, “Whether you’re tackling complex coding challenges, solving mathematical problems for school or personal projects, or providing detailed, multistep instructions to craft a tailored business plan, Gemini-Exp-1206 will help you navigate complex tasks with greater ease.” Google noted the model’s improved performance in more complex tasks, including math reasoning, coding, and following a series of instructions.

VentureBeat took Google’s Exp-1206 model for a thorough test drive this week. We created and tested over 50 Python scripts in an attempt to automate and integrate analysis and intuitive, easily understood visualizations that could simplify the complex data being analyzed. Given how hyperscalers are dominant in news cycles today, our specific goal was to create an analysis of a given technology market while also creating supporting tables and advanced graphics.

Through over 50 different iterations of verified Python scripts, our findings included:

• The greater the complexity of a Python code request, the more the model “thinks” and tries to anticipate the desired result. Exp-1206 attempts to anticipate what’s needed from a given complex prompt and will vary what it produces by even the slightest nuance change in a prompt. We saw this in how the model would alternate between formats of table types placed directly above the spider graph of the hyperscaler market analysis we created for the test.  

• Forcing the model to attempt complex data analysis and visualization and produce an Excel file delivers a multi-tabbed spreadsheet. Without ever being asked for an Excel spreadsheet with multiple tabs, Exp-1206 created one. The primary tabular analysis requested was on one tab, visualizations on another, and an ancillary table on the third.

• Telling the model to iterate on the data and recommend the 10 visualizations it decides best fit the data delivers beneficial, insightful results. Aiming to reduce the time drain of having to create three or four iterations of slide decks before a board review, we forced the model to produce multiple concept iterations of images. These could be easily cleaned up and integrated into a presentation, saving many hours of manual work creating diagrams on slides.

Pushing Exp-1206 toward complex, layered tasks

VentureBeat’s goal was to see how far the model could be pushed in terms of complexity and layered tasks. Its performance in creating, running, editing and fine-tuning 50 different Python scripts showed how quickly the model attempts to pick up on nuances in code and react immediately. The model flexes and adapts based on prompt history.

The result of running Python code created with Exp-1206 in Google Colab showed that the nuanced granularity extended into shading and translucency of layers in an eight-point spider graph that was designed to show how six hyperscaler competitors compare. The eight attributes we asked Exp-1206 to identify across all hyperscalers and to anchor the spider graph stayed consistent, while graphical representations varied.

Battle of the hyperscalers

We chose the following hyperscalers to compare in our test: Alibaba Cloud, Amazon Web Services (AWS), Digital Realty, Equinix, Google Cloud Platform (GCP), Huawei, IBM Cloud, Meta Platforms (Facebook), Microsoft Azure, NTT Global Data Centers, Oracle Cloud, and Tencent Cloud.

Next, we wrote an 11-step prompt of over 450 words. The goal was to see how well Exp-1206 can handle sequential logic and not lose its place in a complex multistep process. (You can read the prompt in the appendix at the end of this article.)

We next submitted the prompt in Google AI Studio, selecting the Gemini Experimental 1206 model, as shown in the figure below.

Next, we copied the code into Google Colab and saved it into a Jupyter notebook (Hyperscaler Comparison – Gemini Experimental 1206.ipynb), then ran the Python script. The script ran flawlessly and created three files (denoted with the red arrows in the upper left).

Hyperscaler comparative analysis and a graphic — in less than a minute

The first series of instructions in the prompt asked Exp-1206 to create a Python script that would compare 12 different hyperscalers by their product name, unique features and differentiators, and data center locations. Below is how the Excel file that was requested in the script turned out. It took less than a minute to format the spreadsheet to shrink it to fit in the columns.

The next series of commands asked for a table of the top six hyperscalers compared across the top of a page and the spider graph below. Exp-1206 chose on its own to represent the data in HTML format, creating the page below.

The final sequence of prompt commands centered on creating a spider graph to compare the top six hyperscalers. We tasked Exp-1206 with selecting the eight criteria for the comparison and completing the plot. That series of commands was translated into Python, and the model created the file and provided it in the Google Colab session.

A model purpose-built to save analysts’ time

VentureBeat has learned that in their daily work, analysts are continuing to create, share and fine-tune libraries of prompts for specific AI models with the goal of streamlining reporting, analysis and visualization across their teams.

Teams assigned to large-scale consulting projects need to consider how models like Gemini-Exp-1206 can vastly improve productivity and alleviate the need for 60-hour-plus work weeks and the occasional all-nighter. A series of automated prompts can do the exploratory work of looking at relationships in data, enabling analysts to produce visuals with much greater certainty without having to spend an inordinate amount of time getting there.

Appendix:

Google Gemini Experimental 1206 Prompt Test

Write a Python script to analyze the following hyperscalers who have announced a Global Infrastructure and Data Center Presence for their platforms and create a table comparing them that captures the significant differences in each approach in Global Infrastructure and Data Center Presence.

Have the first column of the table be the company name, the second column be the names of each of the company’s hyperscalers that have Global Infrastructure and Data Center Presence, the third column be what makes their hyperscalers unique and a deep dive into the most differentiated features, and the fourth column be locations of data centers for each hyperscaler to the city, state and country level. Include all 12 hyperscalers in the Excel file. Don’t web scrape. Produce an Excel file of the result and format the text in the Excel file so it is clear of any brackets ({}), quote marks (‘), double asterisks (**) and any HTML code to improve readability. Name the Excel file, Gemini_Experimental_1206_test.xlsx.

Next, create a table that is three columns wide and seven columns deep. The first column is titled Hyperscaler, the second Unique Features & Differentiators, and the third, Infrastructure and Data Center Locations. Bold the titles of the columns and center them. Bold the titles of the hyperscalers too. Double check to make sure text within each cell of this table wraps around and doesn’t cross into the next cell. Adjust the height of each row to make sure all text can fit in its intended cell. This table compares Amazon Web Services (AWS), Google Cloud Platform (GCP), IBM Cloud, Meta Platforms (Facebook), Microsoft Azure, and Oracle Cloud. Center the table at the top of the page of output.

Next, take Amazon Web Services (AWS), Google Cloud Platform (GCP), IBM Cloud, Meta Platforms (Facebook), Microsoft Azure, and Oracle Cloud and define the eight most differentiating aspects of the group. Use those eight differentiating aspects to create a spider graph that compares these six hyperscalers. Create a single large spider graph that clearly shows the differences in these six hyperscalers, using different colors to improve its readability and the ability to see the outlines or footprints of different hyperscalers. Be sure to title the analysis, What Most Differentiates Hyperscalers, December 2024. Make sure the legend is completely visible and not on top of the graphic.

 Add the spider graphic at the bottom of the page. Center the spider graphic under the table on the page of output.

These are the hyperscalers to include in the Python script: Alibaba Cloud, Amazon Web Services (AWS), Digital Realty, Equinix, Google Cloud Platform (GCP), Huawei, IBM Cloud, Meta Platforms (Facebook), Microsoft Azure, NTT Global Data Centers, Oracle Cloud, Tencent Cloud.