Ioannis Antonoglou and Sofia Puccini discuss Beam, Reflection's mixture-of-experts modelA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. with 501 billion total parameters and 23 billion active parameters. Ioannis Antonoglou describes an emphasis on reasoning efficiency and useful agent capabilities, presenting the performance comparisons as Reflection's reported results rather than independent verification.
Ioannis Antonoglou describes reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. at the scale of roughly 100 million rollouts on 10,500 GPUs over four weeks. The account emphasizes a stable pretrained base, training and inference coordination, weight transfer, sandboxingAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. and managing stale rollouts, alongside proprietary algorithmic changes that are not explained in enough detail to reproduce.
Ioannis Antonoglou tells Sofia Puccini that training on coding and tool-use tasks improved behavior in unfamiliar browsing tasks. They discuss the distinction between task-specific rewards and general problem-solving skills, the tradeoff between model size and reasoning efficiency, and lessons from DQN, AlphaGo, AlphaZero, MuZero and Gemini. Forecasts about more capable autonomous agents remain the speaker's expectations.
Watch on YouTube
