Inference by Turing Post What Reflection AI offers to beat closed labs
Mar 11, 2026
Ioannis Antonoglou, co-founder, President & CTO at Reflection AI and ex-DeepMind researcher behind AlphaGo/AlphaZero/MuZero. He explains building an open-weight general agent model trained with pretraining plus reinforcement learning. Short takes cover why Reflection shifted strategy, the engineering and scale bottlenecks they face, and how open models might challenge closed labs.
AI Snips
Chapters
Books
Transcript
Episode notes
AlphaGo Predicted Today's Model Training Pipeline
- AlphaGo's development foreshadowed modern scale: training phases mirrored today's pretraining then RL fine-tuning workflow.
- Ioannis Antonoglou emphasizes heavy engineering, massive compute and phased training with human testers as the template for current frontier models.
AGI Might Be Engineering Not Magic
- Ioannis argues AGI may not require a single dramatic breakthrough but a correct assembly of existing components and engineering at scale.
- He frames AGI as an agent that interacts with software and performs human-level computer tasks, shifting focus to integration and execution.
Reflection Shifted From RL-First To Build Its Own Base Models
- Reflection originally aimed to push RL-first agentic models but found Western open base models were too weak for large-scale RL post-training.
- That realization led Ioannis to pivot Reflection to build frontier open-weight models with full pretraining plus RL.
