Misha Laskin

Co-Founder and CEO of Reflection AI, former AI researcher at Google

Claims

As speaker

Mentioned in

Key positions and views

The final step to achieving ASI is scaling reinforcement learning on top of LLMs, but the primary bottleneck is the 'reward problem'—the immense difficulty of creating accurate reward models for arbitrary tasks.

True AI generalization is a myth; progress is simply the process of incorporating the test data distribution into the training data distribution.

For startups in critical AI verticals like coding and search, building proprietary frontier models is an existential necessity, as reliance on third-party APIs is an unsustainable and vulnerable position.

The vast majority of an engineer's work (80%) is comprehension and collaboration, not writing code (20%), meaning current AI coding tools are solving the smaller part of the problem and thus have a low productivity impact.

AI business strategy should be vertically integrated, tightly coupling research and product development for a specific domain, rather than pursuing general intelligence in a lab and then searching for a product fit.

Podcast consensus on Laskin

Points of consensus

▶Laskin consistently argues that the primary bottleneck to achieving advanced AI, including ASI, is the 'reward problem'—the difficulty of creating accurate reward models to guide reinforcement learning agents [12, 3, 2].Mar 2026

▶He repeatedly emphasizes that current AI benchmarks and tools are misaligned with real-world value, citing the low productivity impact of coding assistants and the weak correlation of benchmarks like the 'Humanities Last Exam' to user needs [4, 27, 28].Mar 2026

▶Laskin's business and technical strategy centers on the idea that code is the fundamental interface for AI agents, making a powerful coding reasoner a generalizable asset for broader knowledge work [22, 21].Mar 2026

▶He consistently views the path to superintelligence as an extension of existing reinforcement learning paradigms, referencing past projects like AlphaGo and Dota 5 as foundational blueprints that were limited by economics and compute, not fundamental design [9, 19, 29].

Points of debate

▶Laskin presents a nuanced timeline for ASI, predicting the technical blueprint will be set within 'a couple of years' and superintelligence will exist in narrow domains, yet argues its full deployment across industries will be a 'multi-decade' effort, creating a tension between imminent breakthrough and slow integration [5, 10].Mar 2026

▶He claims the compute required for state-of-the-art reinforcement learning is 'manageable for a startup' and two orders of magnitude less than pre-training, while also stating that capitalization requirements remain high and startups face an 'existential threat' if they can't build their own frontier models, suggesting a high but perhaps not insurmountable barrier to entry [11, 34, 35].Mar 2026

▶Laskin critiques the low productivity of existing AI coding tools, yet his company, Reflection AI, is building a specialized coding agent, Asimov, suggesting the problem is not the category itself but the current approach which focuses on code generation over comprehension [4, 21, 25].Mar 2026

▶He posits that true AI generalization is an illusion ('there's no such thing as generalization'), while also arguing that a powerful coding reasoner will be 'operationally generalizable' to many other knowledge work domains, highlighting a distinction between theoretical generalization and practical, operational applicability [1, 22].Mar 2026

Key themes

▶The Path to ASI via Reinforcement LearningMar 2026

Laskin's core technical thesis is that scaling reinforcement learning (RL) on top of large language models is the final paradigm needed for Artificial Super Intelligence (ASI). He identifies the primary obstacle not as the algorithms themselves, but as the 'reward problem'—the immense difficulty of creating a perfect reward model, which he deems an 'ASI complete' problem.

Investors should view progress in reward modeling and RL scaling as key indicators for the next leap in AI capabilities, potentially valuing companies focused on these specific, high-leverage problems over those solely competing on pre-training.

▶Critique of Current AI Utility and BenchmarksMar 2026

Laskin is critical of the real-world productivity gains from current AI tools, particularly in coding, citing reports of negligible or even negative impact. He argues that popular benchmarks are often weakly correlated with end-user value and that the industry over-focuses on code generation (20% of an engineer's job) while neglecting the more significant challenge of code comprehension (80%).

Analysts should be skeptical of vendor claims based solely on benchmark performance and instead focus on products that solve deep, systemic workflow problems like system comprehension and onboarding, which Laskin identifies as the true value drivers.

▶Business Strategy for AI Startups

Laskin advocates for a focused, vertically-integrated strategy where startups co-design research and product for a specific, high-value category. He believes that relying on third-party APIs from frontier labs is an 'existential threat' for startups in critical domains like coding, and that building proprietary models, while capital-intensive, is necessary for long-term survival and success.

The long-term viability of AI application-layer startups may depend on their ability to move down the stack and build specialized models, challenging the notion that a few frontier labs will dominate the entire ecosystem.

▶The Primacy of Code ComprehensionMar 2026

A central theme for Laskin and his company, Reflection AI, is that 'comprehension' is the key unsolved problem in AI, dwarfing the challenge of 'action'. He argues that since code is the primary interface for AI agents to interact with software, building a superior code comprehension agent creates a technology that is operationally generalizable to a wide range of knowledge work.

The market for AI developer tools may see a shift from code-completion 'copilots' to more sophisticated 'comprehension agents' that can navigate and explain complex systems, representing a new, potentially larger, enterprise software category.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Pre-Reflection AI

Laskin worked at Google, where he led reward model development for the Gemini project. He describes the coupling between research and product as 'very loose' and inefficient, an experience that likely informed his later startup strategy [12, 18].

Circa late 2022 / early 2023

Co-founded Reflection AI with Yana Santanaglio. The company was founded on the bet that LLM pre-training was a converging paradigm, allowing them to leverage open-weight models and focus on the next layer of the problem: reinforcement learning and comprehension [13, 36].

Present

As CEO of Reflection AI, Laskin has launched Asimov, a code comprehension agent. His public discourse focuses on the limitations of current AI tools, the importance of solving the 'reward problem,' and the necessity for startups in key verticals to build their own models [21, 12, 34].

Next 1-2 Years (Prediction)

Laskin predicts that within this timeframe, the technical blueprint for ASI will be established, and AI will achieve definitive superintelligence in narrow domains like competitive coding [5, 10, 16].

Next several decades (Prediction)

Despite the rapid technical progress he foresees, Laskin predicts that the actual deployment and integration of ASI across various industries will be a multi-decade process [5].

Suggested prompts

How does Laskin's definition of the 'reward problem' as 'ASI complete' impact the commercial viability and timeline of his own company's reinforcement learning-based approach? &nearr;Given Laskin's assertion that relying on third-party APIs is an 'existential threat,' what specific moats can startups like Reflection AI build against frontier labs who could decide to compete directly in the coding vertical? &nearr;How does Laskin's claim that 'there's no such thing as generalization' reconcile with his strategy that a powerful coding reasoner will be 'operationally generalizable' to other knowledge work? &nearr;If current AI coding tools have negligible or negative productivity impact, what specific evaluation metrics ('evals') must Reflection AI's Asimov agent excel at to prove tangible ROI to enterprise customers? &nearr;

Key concepts

Reinforcement Learning (RL) 7 ep Artificial Super Intelligence (ASI) 6 ep Reward Models 4 ep AI Business Strategy 11 ep AI Coding Tools 8 ep Code Comprehension 5 ep Benchmarks & Evals 5 ep Generalization 2 ep Verticalization 3 ep

Notable quotes

“the final paradigm that we need to have in place before what people used to call AGI, or now I think the goalposts have shifted to ASI, is reached. It's just figuring out how to scale reinforcement learning on top of large language models.”

Misha Laskin · No Priors Ep. 123 | With ReflectionAI Co-Founder and CEO Misha Laskin

“the reward problem in itself is at the time I called I thought it was AGI complete now I'd say it's ASI complete but by the time you have a neural network that can accurately verify any outcome that is probably a super intelligence”

Misha Laskin · No Priors Ep. 123 | With ReflectionAI Co-Founder and CEO Misha Laskin

“Maybe the hot take is that there's no such thing as generalization. There's just bringing the test distribution into train.”

Misha Laskin · No Priors Ep. 123 | With ReflectionAI Co-Founder and CEO Misha Laskin

“it actually puts companies that are in these kind of critical path categories like search and coding in a pretty existential place if they can't build their own frontier models.”

Misha Laskin · No Priors Ep. 123 | With ReflectionAI Co-Founder and CEO Misha Laskin

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report