Who Is Misha Laskin, the Co-Founder and CEO of Reflection AI?

Misha Laskin is the co-founder and chief executive officer (CEO) of Reflection AI, an artificial intelligence (AI) research laboratory whose latest funding round closed at a $25 billion pre-money valuation.

Before establishing Reflection AI in March 2024 with Ioannis Antonoglou, Laskin worked as a research scientist at Google DeepMind, where he directed reward model training for the Gemini project.

On October 5, 2026, Reflection AI announced Beam. It is an open-weight model with 501 billion parameters.

Who Misha Laskin Is

Venture capital firm Lightspeed Venture Partners lists Laskin as CEO and co-founder of Reflection AI, with Antonoglou serving as co-founder and CTO. TechCrunch called Laskin "the company's CEO" in October 2025, and in October 2026 it described Reflection AI as "the two-year-old, Brooklyn-based startup."

On April 23, 2026, Laskin appeared on CNBC's 'Squawk on the Street'. He discussed Reflection AI's latest valuation. He also discussed which countries are ahead in the global AI race.

• Current role: Chief executive officer and co-founder of Reflection AI
• Organization headquarters: Brooklyn, New York
• Known for: Leading reward modeling for Google DeepMind's Gemini project and co-founding Reflection AI
• Education: Doctor of philosophy (PhD) in theoretical physics from the University of Chicago
• In the news: Announcing Beam, a 501-billion-parameter open-weight model, in October 2026

Academic Training in Theoretical Physics and the Pivot to AI

Laskin completed graduate studies in theoretical physics before redirecting his career toward machine learning.

According to his personal website, mishalaskin.com, Laskin earned a PhD in theoretical physics from the University of Chicago (UChicago), though the site does not list his graduation year. A profile for the Manifold podcast notes that he trained in theoretical physics at Yale University and UChicago before becoming an artificial intelligence scientist.

A profile published by Sequoia Capital states that Laskin had recently completed a PhD in quantum physics in Chicago when DeepMind released its AlphaGo paper. That paper prompted him to change career paths. On the Sequoia Training Data podcast, Laskin said Richard Feynman's lectures on physics got him into physics. He later decided to leave physics because he wanted to work on the root node problems of his time. After his doctorate, Laskin founded a Y Combinator-backed startup. Its name has not been published. Sequoia describes it as "an AI startup," but on the Training Data podcast Laskin called it "my previous startup which has nothing to do with AI."

Postdoctoral Research in Reinforcement Learning at UC Berkeley

Laskin entered artificial intelligence research through self-directed study and a postdoc at the University of California, Berkeley.

Laskin began by working on published research prompts from OpenAI around 2018 or 2019, as recounted on the Sequoia Training Data podcast. He sent cold emails to OpenAI research scientists. One researcher introduced him to professor Pieter Abbeel at the University of California, Berkeley (UC Berkeley). Abbeel invited Laskin to join his research group. There, Laskin spent two and a half years as a postdoctoral researcher improving reinforcement learning (RL) techniques.

While conducting research at UC Berkeley, Laskin published academic papers under the name Michael Laskin. Sequoia says his Berkeley work covered data augmentation and extracting high-level features from raw pixels. Per Sequoia, the most cited paper he co-authored there was Decision Transformer.

• CURL (April 2020): Contrastive Unsupervised Representations for Reinforcement Learning, co-authored on arXiv with Aravind Srinivas and Pieter Abbeel, applying contrastive learning to visual RL
• RAD (April 2020): Reinforcement Learning with Augmented Data, published on arXiv and accepted at NeurIPS 2020, demonstrating data augmentations on pixel inputs
• SUNRISE (July 2020): A Simple Unified Framework for Ensemble Learning in Deep Reinforcement Learning, released on arXiv for ICML 2021
• Decision Transformer (June 2021): Reinforcement Learning via Sequence Modeling, released on arXiv, abstracting RL into sequence modeling via transformers
• URLB (October 2021): Unsupervised Reinforcement Learning Benchmark, published on arXiv as a standardized evaluation suite
• CIC (February 2022): Contrastive Intrinsic Control for Unsupervised Skill Discovery, released on arXiv to discover diverse agent behaviors without task rewards

Google DeepMind Career and Gemini Reward Modeling

At Google DeepMind, Laskin was a research scientist working on what his personal site calls "developing generally intelligent agents."

According to Sequoia Capital, Laskin interviewed in 2021 with Volodymyr Mnih, who was establishing a new general agents team for Google DeepMind in Toronto. Sequoia says he worked on in-context reinforcement learning with Mnih. Laskin is first author of the Algorithm Distillation paper. It was posted on arXiv in October 2022.

When Google initiated the Gemini project, Ioannis Antonoglou led reinforcement learning from human feedback (RLHF). Antonoglou asked Laskin to lead reward model training. On the Training Data podcast, Laskin explained: "Ioannis led the RLHF effort, and I led reward model training, which was kind of a key part of of RLHF." TechCrunch reported that Laskin "led reward modeling for DeepMind's Gemini project." Sequoia says he and Antonoglou worked together to launch Gemini 1 and 1.5. Laskin is a listed author on Google's primary Gemini technical report on arXiv and on research exploring Vision-Language Models as a Source of Rewards, both published in December 2023.

Founding of Reflection AI and Autonomous Agent Development

Laskin and Antonoglou left Google DeepMind to form Reflection AI in March 2024, according to Sequoia Capital. Sequoia says Reflection AI's original plan was to build "its own large-scale LLM and reinforcement learning platform" and to "run its RL work on open base weights it controlled," instead of building on a closed model's API.

In July 2025, Reflection AI announced Asimov. The autonomous software engineering agent was profiled by SiliconANGLE.

Asimov synthesizes software context across codebase documentation, project notes, code, emails, and Slack messages. SiliconANGLE cites Laskin telling Wired: "We are in this kind of semi-autonomous phase, where agents are just starting to work."

Reflection AI Capital Rounds and Data Center Agreements

Reflection AI has raised roughly $4.7 billion, per PitchBook data reported by TechCrunch, and has agreed to buy computing capacity from SpaceX and Nebius.

In October 2025, TechCrunch reported that Reflection AI raised $2 billion in equity funding at an $8 billion valuation, a 15-fold increase over its $545 million valuation seven months earlier. Investors in that round included NVIDIA, Sequoia Capital, Lightspeed Venture Partners, Eric Schmidt, Eric Yuan, and others, per TechCrunch. A blog post on Reflection AI's site titled Building Frontier Open Intelligence displayed an August 9, 2025, date for the announcement. TechCrunch and Sequoia Capital both documented the raise in October 2025.

On April 23, 2026, Reflection AI stated in its newsroom that its subsequent funding round closed at a $25 billion pre-money valuation. On October 5, 2026, TechCrunch reported via PitchBook data that Reflection AI has raised roughly $4.7 billion in total capital. An undated Endeavor Greece profile of Antonoglou lists $2.13 billion. That is likely an older figure. CNBC reported on June 22, 2026, that Reflection AI has agreed to pay SpaceX $150 million per month from July 1, 2026, through 2029. That comes to about $6.3 billion if the agreement runs to the end of its term. In July 2026, Nebius agreed to sell $1 billion in AI computing capacity to Reflection AI.

• October 2025: Raised $2 billion at an $8 billion valuation from investors including NVIDIA and Sequoia Capital
• April 2026: Confirmed a venture funding round at a $25 billion pre-money valuation
• June 2026: Agreed to pay SpaceX $150 million per month for compute through 2029, about $6.3 billion over the full term (CNBC)
• July 2026: Contracted $1 billion in AI compute capacity from Nebius
• October 2026: Roughly $4.7 billion raised in total, per PitchBook data cited by TechCrunch

Beam Model Release and the Open-Weight Philosophy

Reflection AI announced Beam on October 5, 2026, as an open-weight model with 501 billion parameters.

According to the technical introduction on the Reflection AI blog, Beam uses a sparse Mixture-of-Experts (MoE) architecture that activates 23 billion parameters per token. The model trained on 23.8 trillion pretraining tokens. Midtraining extends Beam's effective context to 1 million tokens. Pretraining took under four weeks on 6,144 NVIDIA GB300 graphics processing units (GPUs). Four weeks of reinforcement learning followed on 10,500 GB300 GPUs across more than 100 million rollouts.

Reflection AI designated the weights for release under an Apache 2.0 license, with model artifacts scheduled for distribution later in October 2026. The Next Web reported that early access is managed via waitlist at platform.reflection.ai. Reflection AI reported that Beam matches GLM-5.2 on advanced reasoning benchmarks using three to four times less inference compute. TechCrunch noted these internal metrics remain unverified by third parties. Reflection AI's own figures, as reported by The Next Web, put Moonshot AI's Kimi K3 ahead of Beam on Terminal Bench v2.1. Kimi K3 led Beam 88.3 to 80.1. Laskin told Semafor that the US Center for Advancing Innovation and Standards for Super Intelligence and the UK AI Safety Institute are assessing the model. On the Sources podcast, Laskin stated: "The only way to own intelligence is, by definition, if it's open."

Reflection AI's Business Model and Private Ownership

Reflection AI has not published a revenue model, so the clearest public evidence of its business is the partnership list on its About page.

The Reflection AI About page says Reflection AI is building a Korean sovereign AI cloud described as a 250-megawatt (MW) data center. The same page says it has been validated on the Dell AI Factory with NVIDIA. Reflection AI also provides open-weight models to the US Department of Energy (DOE) Genesis Mission, the page says. That mission serves the 17 US national laboratories.

In October 2025, Laskin told TechCrunch that Reflection AI would release open model weights while "largely keeping datasets and full training pipelines proprietary." He also said: "Once you get into that territory where you're a large enterprise, by default you want an open model." Laskin's net worth and his equity stake in Reflection AI are not published. Reflection AI is privately held. Its shares do not trade on a public exchange.