Misha Laskin

Misha Laskin is the co-founder and CEO of Reflection, alongside co-founder and CTO Ioannis Antonoglou. His work spans reinforcement learning (RL), language models, and autonomous systems: collaborative research on learning from pixels and adapting through context preceded his leadership at a company that began with autonomous coding and later expanded toward open frontier models. He also publishes as Michael Laskin and uses the handle MishaLaskin.[5][6][7][8][10][3]

At Reflection’s March 7, 2025 launch, the company set out to scale the autonomous capabilities of large language models through reinforcement learning. Its initial focus was a system that could plan, write, and refine complex software. The team’s stated thesis was: “We believe that solving autonomous coding will enable superintelligence more broadly.” It argued that the reasoning and iterative self-improvement needed for coding could extend to other computer-based work.[6]

The launch also tied research to product development. Reflection emphasized real-world evaluations and user feedback, stating that “Groundbreaking AI doesn’t develop in a vacuum; it requires co-designing research and product.” This was a collective company position under Laskin’s leadership, connecting its research ambitions to whether autonomous systems delivered useful results in practice.[6][5]

On October 9, 2025, Reflection announced a broader direction: frontier open intelligence, bringing its methods from autonomous coding to general agentic reasoning and combining large-scale pretraining with advanced RL. The team argued that capable open models could keep AI infrastructure accessible rather than concentrated in a few closed labs. It also framed openness as a way for independent researchers to investigate safety, while acknowledging the risks of making capable models widely available.[7]

Before Reflection, Laskin’s career included theoretical physics, startup founding, and reinforcement-learning research. His personal biography describes him as a theoretical physics PhD at the University of Chicago, founder of a Y Combinator-backed startup, and former UC Berkeley postdoc. These earlier roles provide context for the research that preceded his work on language models.[4][5]

At Berkeley, Laskin and Aravind Srinivas were equal contributors to CURL, with Pieter Abbeel as their collaborator; the work appeared at ICML 2020. CURL extracts features from raw pixels through contrastive learning while jointly training an RL objective. Its representation-learning task is auxiliary to the control algorithm: the paper paired it with Soft Actor Critic for continuous control and Rainbow DQN for discrete control. The authors reported improved sample efficiency on DeepMind Control Suite and Atari benchmarks.[8]

RAD—Reinforcement Learning with Augmented Data explored a simpler intervention: changing the observations used for training. Laskin and Kimin Lee contributed equally, collaborating with Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. In their pixel-based experiments, random cropping and translation enabled Soft Actor Critic to outperform more complex competing methods. Their finding was that data diversity could help agents focus on meaningful information without changing the underlying RL method.[9]

In the October 25, 2022 paper In-context Reinforcement Learning with Algorithm Distillation, Laskin and his co-authors trained a causal transformer on learning histories generated by an RL algorithm. The transformer predicts actions from the preceding history, allowing its policy to improve in context without updating network parameters. The authors demonstrated this in environments involving sparse rewards, combinatorial task structure, and pixel observations, reporting greater data efficiency than the source algorithm in their experiments.[10]

His Algorithm Distillation collaborators were Luyu Wang, Junhyuk Oh, Emilio Parisotto, Stephen Spencer, Richie Steigerwald, DJ Strouse, Steven Hansen, Angelos Filos, Ethan Brooks, Maxime Gazeau, Himanshu Sahni, Satinder Singh, and Volodymyr Mnih.[10]

Laskin subsequently worked on Gemini at Google DeepMind. In his DeepMind-era personal biography, he expressed interest in how RL could unlock new capabilities in language and multimodal models. Lightspeed’s James Alcorn and Raviraj Jain report that he led reinforcement-learning reward modeling for Gemini 1.5 after working under Pieter Abbeel at Berkeley. That responsibility preceded his role as Reflection’s co-founder and CEO.[4][5]