Who Is Ioannis Antonoglou, Co-Founder and CTO of Reflection AI?

Ioannis Antonoglou is the co-founder and Chief Technology Officer (CTO) of Reflection AI, an artificial intelligence (AI) research laboratory that develops open-weight foundation models.

Before launching Reflection AI in March 2024 alongside co-founder and Chief Executive Officer (CEO) Misha Laskin, Antonoglou spent more than a decade at Google DeepMind as one of its founding engineers. At DeepMind, he co-authored reinforcement learning papers on Deep Q-Networks (DQN), AlphaGo, AlphaZero, and MuZero. He later led the Reinforcement Learning from Human Feedback (RLHF) team for Google's Gemini models.

On October 5, 2026, Reflection AI announced Beam, a sparse Mixture of Experts (MoE) model with 501 billion parameters, after a funding round at a $25 billion pre-money valuation.

Who Ioannis Antonoglou Is

Lightspeed Venture Partners lists Antonoglou as "Co-Founder, CTO" of Reflection AI, which Endeavor Greece says is headquartered in New York with offices in San Francisco and London.

In profiles published by Endeavor Greece, where he was selected as an Endeavor Entrepreneur in 2025, Antonoglou is also referred to as Giannis. An interview with Turing Post published in February 2026 notes that Antonoglou believes Reflection is the only frontier lab truly committed to open science. In October 2026, TechCrunch reported that Reflection AI has raised roughly $4.7 billion, per PitchBook.

• Current role: Co-Founder and Chief Technology Officer (CTO) of Reflection AI
• Organization headquarters: New York, New York, with offices in San Francisco and London
• Known for: Writing DeepMind's DQN library, engineering work on AlphaGo, and co-authoring the MuZero paper
• In the news: Reflection AI announcing Beam, its 501-billion-parameter open-weight model, in October 2026

Academic Background and Early Education

Antonoglou was born and raised in Thessaloniki, Greece, according to his Endeavor Greece profile. The same profile says he studied at the Aristotle University of Thessaloniki before moving to the University of Edinburgh, where he completed his studies in artificial intelligence.

The Endeavor Greece profile does not list graduation dates or specific degree designations. An episode description from the Endeavor Greece Outliers podcast states that his academic work continued at University College London (UCL), where doctoral research bridged neuroscience with machine learning. That same podcast title references the National Technical University of Athens (NTUA).

His Endeavor Greece profile mentions neither UCL nor NTUA, so whether he holds a PhD or an NTUA degree is unconfirmed. His age, date of birth, and net worth are not published.

DeepMind Research Career and Reinforcement Learning Milestones

In late 2012, Antonoglou joined DeepMind as employee number 25 and the sixth member of its research division. In an editorial profile published by Sequoia Capital, Antonoglou said DeepMind was "the only place in the world where people were seriously considering building AGI" (artificial general intelligence). His first DeepMind collaborator, Vlad Mnih, asked him to write a clean implementation of the Deep Q-Network (DQN) algorithm as a library DeepMind could use for deep reinforcement learning research.

Antonoglou is a co-author of 'Playing Atari with Deep Reinforcement Learning', posted on arXiv in December 2013 and presented at the NIPS Deep Learning Workshop. Antonoglou subsequently co-authored 'Prioritized Experience Replay', published on arXiv in November 2015 and accepted at the International Conference on Learning Representations (ICLR) in 2016.

On AlphaGo, David Silver asked Antonoglou to write graphics processing unit (GPU) kernels to speed up the neural networks, according to Sequoia. After Google acquired DeepMind in 2014, he moved on to implementing the system on Google's first Tensor Processing Units (TPUs). He is a listed co-author of the January 2016 AlphaGo paper in Nature, and Laskin has said Antonoglou was in Seoul when AlphaGo played Lee Sedol.

• Playing Atari with Deep Reinforcement Learning (2013): Presented at the NIPS Deep Learning Workshop
• Prioritized Experience Replay (2015): Published on arXiv and presented at ICLR 2016
• Mastering the game of Go with deep neural networks and tree search (2016): Published in Nature
• Mastering the game of Go without human knowledge (2017): Published in Nature
• Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (2017): Published on arXiv
• Bayesian Optimization in AlphaGo (2018): Published on arXiv

MuZero and Leading RLHF for Gemini

Antonoglou is the second of twelve authors on the November 2019 MuZero paper, 'Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model', posted on arXiv. The host of Sequoia Capital's Training Data podcast says he later co-led MuZero's development.

On the same podcast, Antonoglou said MuZero "has been used in the real world" for video compression at YouTube. He is also a co-author of 'Learning and Planning in Complex Action Spaces', posted on arXiv in April 2021.

Following the release of ChatGPT, Antonoglou shifted focus toward Google's new large language model Gemini. Antonoglou stated to Sequoia Capital that 'everything changed when ChatGPT came out.' Sequoia says Antonoglou was tapped to lead the Reinforcement Learning from Human Feedback (RLHF) effort for Gemini. He asked Misha Laskin to lead reward models, and Sequoia says the two worked together to launch Gemini 1 and 1.5. Antonoglou is a listed author on the Gemini technical report, posted on arXiv in December 2023.

Founding of Reflection AI and Capital Expansion

In March 2024, Antonoglou and Misha Laskin left Google DeepMind to form Reflection AI. Endeavor Greece says Reflection AI raised more than $130 million in its first year and grew to a team of over 30 engineers.

In July 2025, Reflection AI announced its autonomous coding agent, Asimov. SiliconANGLE quoted Antonoglou: "We've built something like Deep Research, but it's for your engineering systems." He observed that in big engineering teams, a lot of the knowledge is actually stored outside of the codebase.

In October 2025, Reflection AI raised $2 billion at an $8 billion valuation, per TechCrunch. Antonoglou told Turing Post it was "our Series B in October, where we raised more than two billion dollars to access the compute we need." According to PitchBook data reported by TechCrunch in October 2026, cumulative venture investment in Reflection AI reached roughly $4.7 billion, with the latest round establishing a $25 billion pre-money valuation.

• March 2024: Co-founded Reflection AI with Chief Executive Officer Misha Laskin
• July 2025: Announced Asimov, an autonomous coding agent for engineering systems
• October 2025: Raised $2 billion at an $8 billion valuation, which Antonoglou called the Series B, to access compute
• October 2026: Roughly $4.7 billion raised in total per PitchBook, with the last round at a $25 billion pre-money valuation

Beam Model Architecture and Open-Weight Strategy

On October 5, 2026, Reflection AI announced Beam, a 501-billion-parameter open-weight model. According to the Reflection AI blog, Beam is a sparse Mixture of Experts (MoE) model with 23 billion active parameters, pretrained on 23.8 trillion tokens. Reflection AI says the weights will follow later in October 2026 under an Apache 2.0 license.

In his February 2026 Turing Post interview, Antonoglou said Llama 4 "not being a particularly strong model" showed that "the whole Western ecosystem was missing a powerful open base model." He said that is why Reflection AI now builds its frontier models from the ground up, doing both pre-training and reinforcement learning. He also said that building the model, even when you know what to do, "is like building a rocket."

At Davos in January 2026, Antonoglou told Alex Heath on Sources Live that Reflection AI planned to release what it hoped would be "the most powerful open-weight model in the world" later in 2026. Asked on the Sources podcast whether a model could become too capable to release openly, he said: "It is possible that you get to a level of capability that you want to just be more careful with how you deploy it." The Next Web reports that Reflection AI is already training its next model, and that Antonoglou said significantly bigger models will come in 2027.

Technical Philosophy on Artificial General Intelligence and Agent Adoption

Antonoglou told Turing Post that AGI is "literally an agent that interacts with software on a computer and can really do what a human can do." Describing reinforcement learning to Sequoia, he said: "Things that it gets right, it should just do more of. And the things that it gets wrong, it should do less of."

When assessing current Large Language Model (LLM) agents on Sequoia Capital's Training Data podcast, Antonoglou emphasized reliability limitations: 'Sometimes they get it, sometimes they don't. You cannot trust them.' On the same podcast he predicted: "I think we are one to three years away from the 50 percent threshold for SWE agents and three to five years from achieving 90 percent."

Antonoglou told Turing Post that being open "doesn't actually slow you down" and that "the main metric that matters is adoption."