99 - From Wedding Singer to Google DeepMind: Dima Yeroshenko on Building Frontier AI Models by Product&Growth Show

In Episode 99 of The Product and Growth Show, our first in a year, I sat down with my good friend Dima, Technical Program Manager at Google DeepMind. Dima’s path into AI is anything but linear: singing at weddings in Ukraine, law school and night shifts assembling TVs in Wrocław, then Google Ads customer support, engineering on Google Maps and Gemini integrations, and finally working alongside DeepMind’s research scientists. We talked about what actually goes into building a frontier model, who decides what Gemini gets better at, and whether you still need a PhD to break into an AI lab.

What did we talk about?

- Dima’s non-linear career, from odd jobs in Poland to climbing Google’s internal ladders from support to engineering to DeepMind
- What a technical program manager does in a frontier AI lab, and what a typical week looks like
- How a model gets built: pre-training, mid-training, post-training, evals, and picking which checkpoint to ship
- Who decides the Gemini roadmap, and why labs don’t just ship ten specialised models
- Benchmarks vs. real user feedback, benchmark gaming, model personality and sycophancy
- AI agents in real life: early Gemini extensions, the Uber demo, and what Dima still won’t delegate to AI
- Whether AI wrappers are bad businesses, what makes a niche AI startup defensible, and career advice for getting into AI without a fancy degree

Dima’s LinkedIn: https://www.linkedin.com/in/dima-yeroshenko/