Hands on with Gemini 4 Argon: Amplitude
Read the Portuguese translation →
Author: Google Cloud · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Hands on with Gemini 4 Argon: Amplitude · Channel: Google Cloud · 02:01 · Watch on YouTube
Topics: AI agents, AI evaluation and benchmarks, Data, databases and embeddings, AI content generation
Summary
Section titled “Summary”Amplitude helps customers understand how people use their software by analyzing product data for insights and opportunities. The speakers say Gemini 4 Argon can interpret customer questions and find answers across collected data sources, including cases where other models may give up. They report that Gemini 4 scored over 90% on Amplitude’s evaluations for generating charts that answer customer questions. In a demonstrated evaluation, the agent makes many tool calls, can recover from some failures, and continues exploring data to produce an answer.
Key points
Section titled “Key points”- Amplitude analyzes data about how end users interact with software products.
- The company uses that data to identify insights and opportunities for customers.
- Argon is described as helping interpret what customers mean and make use of many data sources and signals.
- Amplitude says Gemini 4 scored over 90% on its evaluations for generating charts that answer customer questions.
- The agent uses tool calls to explore data and generate visualizations.
- A demonstrated evaluation includes dozens of tool calls and examples of the agent recovering from failures.
- The agent can incorporate past customer explorations into its analysis.
- The speakers describe the agent as persistent in searching for answers that are difficult to find.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Amplitude and its product-data work | Paige introduces Curtis, who explains how Amplitude analyzes software usage data to help customers build better products. |
| 00:30 | Finding answers in customer data | Curtis describes Argon’s ability to interpret customer intent and use collected data, then introduces chart generation with Gemini 4. |
| 01:00 | Chart-generation evaluation and tool calls | Curtis reports Gemini 4’s evaluation result and shows an agent conversation with many tool calls, including recovery from failures. |
| 01:31 | Exploration and persistent analysis | The speakers discuss how the agent explores custom data sources, uses tool calls to support analysis, and searches persistently for answers. |
Quotes
Section titled “Quotes”We help our customers build better products. — 00:00
Argon is doing a great job at understanding — 00:30
Gemini 4 scored over 90% on our evals — 01:00
there are dozens. — 01:31
it’s very persistent in trying to get things. — 02:01
Concepts and entities
Section titled “Concepts and entities”Gemini 4, Argon, Amplitude, Google DeepMind, Paige, Curtis, evals, tool calls
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Hi, my name is Paige from Google DeepMind, and I’m here today with Curtis, one of the co-founders of Amplitude, who’s here today to talk to us about how they’ve been using Gemini 4 to tackle some of their hardest problems. We help our customers build better products. What that means is we help them gather data on how their end users are using their software products, and we mine that data for insights and opportunities. I think one of the hardest things is understanding
00:30 what people actually mean, and I think Argon is doing a great job at understanding that there’s so many data sources and signals that people collect, and customers don’t know what to do with it. Some other models kind of give up when the answer is actually sitting in the data, and Argon was able to kind of collect all that information and correctly. And for what the definition of that should actually be. Excellent. And so you’ve been using Gemini 4 for a while. Yeah, yeah. So one of the things that our agents need to do is build charts to help customers
01:00 visualize what they’re understanding and what the answers to the questions look like. And one of the things we found is Gemini 4 scored over 90% on our evals with being able to generate the chart that answers the the customer’s question. Awesome. Sounds like a lot of combing through logs, a lot of looking through databases, which probably means a lot of tool calls. Is that right? Yeah. Let me show you one of the things that we tested out. So this is a conversation
01:31 that we have in one of our evals. You can see that each of these is a tool call. I love how many tool calls there are. It looks like there are dozens. You can see some examples of when the tool call might have failed, but then it was able to recover quickly and get the right answer and continue on with the analysis and with the final result. Yeah. This is how our customers might interact, given some exploration. They might ask for some information and the agent will go off on its exploration. I love it, I love that it’s connected to custom data sources, that it can
02:01 be informed by external information, and that it can also use tool calls to generate charts along the way to help people understand and deeply interrogate the data that they might want to look at. It’s very good at taking past explorations that customers have had and incorporating that into its analysis, and we’ve seen it do a really good job, you know, being persistent about finding those kind of needle in the haystack kind of answers. It works hard, like it’s very persistent in trying to get things.