No Priors Ep. 123 | With ReflectionAI Co-Founder and CEO Misha Laskin | No Priors: AI, Machine Learning, Tech, & Startups

In this episode of No Priors, Misha Laskin, co-founder of ReflectionAI and former Google DeepMind researcher, discusses building superintelligent autonomous systems with a focus on co-designing products and research. Laskin details ReflectionAI's mission to create an "omniscient oracle" for organizations, starting with Asimov, a code comprehension agent designed to deeply understand large codebases. The conversation covers the importance of focusing on understanding and context in coding tools, the challenges and strategies in post-training models, and the significance of evaluations driven by customer needs. Laskin also shares his views on the future of AI, the potential for achieving ASI in specific domains, and the competitive landscape for AI companies, emphasizing the need for vertical integration and owning the intelligence behind products.

Outlines

Part 1: Introduction and Vision

Part 2: Asimov's Design and Evaluation

Part 3: Scaling RL and Future Outlook

Part 4: Conclusion

Sign in to continue reading, translating and more.

Open full episode in Podwise

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

No Priors: AI, Machine Learning, Tech, & Startups

Part 1: Introduction and Vision

Introduction to ReflectionAI and the Mission of Building Superintelligent Autonomous Systems

Misha Laskin's Journey into AI and the Vision for Scaling Reinforcement Learning

Introducing Asimov: A Code Comprehension Agent Focused on Understanding Complex Systems

Part 2: Asimov's Design and Evaluation

Differentiating Asimov: Combining Product Differentiation with Research Innovation

Evaluation Strategies and the Importance of Tight Coupling Between Product and Model

Examples of Asimov's Superior Performance and the Design of Team-Wide Memory

Designing Permissions for Team-Wide Memory and the Availability of Pre-trained Models

Part 3: Scaling RL and Future Outlook

Challenges in Scaling Reinforcement Learning and the Importance of Reward Models

Hot Takes: Synthetic Environments, Generalization, and the Definition of ASI

Jagged Superintelligence and the Verticalization of Coding

Data Collection and the Differences Between RL in Robotics and Language Models

Reflections Path Forward and the Timing of Deployment

Part 4: Conclusion

Conclusion and Podcast Information

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

No Priors: AI, Machine Learning, Tech, & Startups

Part 1: Introduction and Vision

00:04Introduction to ReflectionAI and the Mission of Building Superintelligent Autonomous Systems

Introduction to ReflectionAI and the Mission of Building Superintelligent Autonomous Systems

03:30Misha Laskin's Journey into AI and the Vision for Scaling Reinforcement Learning

Misha Laskin's Journey into AI and the Vision for Scaling Reinforcement Learning

07:36Introducing Asimov: A Code Comprehension Agent Focused on Understanding Complex Systems

Introducing Asimov: A Code Comprehension Agent Focused on Understanding Complex Systems

Part 2: Asimov's Design and Evaluation

11:15Differentiating Asimov: Combining Product Differentiation with Research Innovation

Differentiating Asimov: Combining Product Differentiation with Research Innovation

16:16Evaluation Strategies and the Importance of Tight Coupling Between Product and Model

Evaluation Strategies and the Importance of Tight Coupling Between Product and Model

21:11Examples of Asimov's Superior Performance and the Design of Team-Wide Memory

Examples of Asimov's Superior Performance and the Design of Team-Wide Memory

27:32Designing Permissions for Team-Wide Memory and the Availability of Pre-trained Models

Designing Permissions for Team-Wide Memory and the Availability of Pre-trained Models

Part 3: Scaling RL and Future Outlook

33:39Challenges in Scaling Reinforcement Learning and the Importance of Reward Models

Challenges in Scaling Reinforcement Learning and the Importance of Reward Models

37:21Hot Takes: Synthetic Environments, Generalization, and the Definition of ASI

Hot Takes: Synthetic Environments, Generalization, and the Definition of ASI

42:41Jagged Superintelligence and the Verticalization of Coding

Jagged Superintelligence and the Verticalization of Coding

48:10Data Collection and the Differences Between RL in Robotics and Language Models

Data Collection and the Differences Between RL in Robotics and Language Models

55:12Reflections Path Forward and the Timing of Deployment

Reflections Path Forward and the Timing of Deployment

Part 4: Conclusion

1:02:30Conclusion and Podcast Information

Conclusion and Podcast Information