Meet Ioannis Antonoglou, a founding engineer at DeepMind who's now building the Western answer to DeepSeek. In this exclusive interview, he breaks down why Reflection — backed by Nvidia and others with $2 billion in funding — believes open-weight AI models are the future.
What you'll learn:
- Why Reflection left the closed-model revolution to build truly open intelligence
- How DeepSeek's success exposed the flaws in the "closed AI for safety" argument
- The bet on AI sovereignty for nations and enterprises
- How reinforcement learning will transform coding, agents, and AGI
- Reflection's plan to build the most powerful open-weight model
Antonoglou argues that open intelligence isn't risky but essential. From democratizing access to empowering nations to maintain sovereignty, he makes a bold case against the incumbents. "It's all about the mission," he says.
This is the story of how the next phase of the AI race is being written not by OpenAI or Anthropic, but by a new startup that wants to change the game entirely.
Full interview conducted by Sources founder and ACCESS co-host Alex Heath on the sidelines of the 2026 World Economic Forum in Davos, Switzerland.
Chapters:
00:55 – Founding of Reflection AI: Giannis discusses his background as a founding engineer at DeepMind and how he and his co-founder Misha decided to start Reflection AI with a focus on reinforcement learning.
03:16 – Shift to Open Science: The motivation behind staying dedicated to open model science while other major labs have moved toward closed systems.
04:03 – The "DeepSeek of the West": Addressing the company's $2 billion funding round and its positioning as a Western alternative to Chinese open-weight models.
05:30 – Democratization and Safety: Why Reflection AI believes open intelligence is the best way to accelerate scientific progress and ensure AI safety through transparency.
11:30 – Competing for Talent and Capital: How a mission-driven open lab attracts top researchers despite the massive capital advantages of closed incumbents like OpenAI and Anthropic.
14:48 – 2026 Goals: Details on the upcoming 2026 release of their fully open-weight model and the ambition to build the most powerful open model in the world.
20:31 – The Path to AGI: Defining digital vs. physical AGI and explaining why coding and tool-calling are the most natural ways for models to interact with computers.
