Keep pulling the thread on Misha Laskin.
Reflection AI has launched Asimov, a code comprehension agent designed to function like a principal-level engineer for large codebases.
Misha Laskin's "hot take" is that there is no such thing as true generalization in AI, only the process of bringing the test data distribution into the training data distribution.
Misha Laskin believes the compute required for state-of-the-art reinforcement learning is currently manageable for a startup, requiring about two orders of magnitude fewer flops than pre-training.
A founding bet of Reflection AI was that pre-training for large language models was converging on a known paradigm, allowing the company to leverage open-weight models instead of investing in its own pre-training.
Misha Laskin, who led reward model development for Gemini, believes the primary bottleneck in scaling reinforcement learning is the "reward problem"—the difficulty of creating accurate reward models for arbitrary tasks.
Misha Laskin believes the final paradigm needed to reach Artificial Super Intelligence (ASI) is scaling reinforcement learning on top of large language models.
Misha Laskin predicts that while the technical blueprint for building ASI will be established within the next couple of years, its actual deployment across various industries will be a multi-decade endeavor.
Misha Laskin cites the rapid revenue growth of Anthropic as evidence that a frontier AI lab can become a massive, self-sustaining business without being owned by a cloud provider.
Misha Laskin predicts that within a couple of years, there will be definitive superintelligence in some meaningful categories of work, such as specific sub-domains of coding.
Reflection AI's strategy is based on the belief that "code" is the fundamental interface for AI agents to interact with software, meaning a powerful coding reasoner will be operationally generalizable to many other knowledge work domains.
Misha Laskin believes the problem of creating a perfect reward model is "ASI complete," meaning a neural network that can accurately verify any outcome is likely a superintelligence itself.
Misha Laskin argues that startups in critical AI categories like search and coding face an existential threat if they cannot build their own frontier models and must rely on third-party APIs.