EO StudioFrom Google DeepMind to a $8B Superintelligence Startup | ReflectionAI, Misha Laskin
At a glance
WHAT IT’S REALLY ABOUT
Misha Laskin on autonomous coding as path to superintelligence
- AlphaGo’s “move 37” convinced Laskin that superhuman creativity is real and will soon appear across many knowledge-work domains as analogous “move 37s.”
- ReflectionAI’s core bet is that solving autonomous coding effectively solves computer-based intelligence broadly, because code is the most natural interface (“embodiment”) for language models.
- He left DeepMind after leading post-training/RLHF work on Gemini because small, product-coupled teams can iterate faster and evaluate progress via real-world customer use.
- At frontier scale, simple training ideas executed with extreme craft and infrastructure matter more than elaborate algorithms, a key lesson from training massive models.
- For founders and researchers, the hardest edge is picking the right questions; Laskin uses writing and adversarial discussion to sharpen clarity, and emphasizes surrounding yourself with exceptional people.
IDEAS WORTH REMEMBERING
5 ideas“Move 37” is a template for recognizing emerging superintelligence.
Laskin argues that when an AI’s output looks wrong but later proves decisively correct, it signals a qualitative creativity gap—an early hint of how ASI could reshape many fields beyond games.
Autonomous coding is positioned as the shortest path to general autonomy on computers.
ReflectionAI believes code is the natural “hands and legs” for language models, so an agent that can reliably plan, execute, and verify code can likely perform a wide range of computer-based tasks—not just software engineering.
LLMs become truly agentic when paired with reinforcement learning for autonomy.
He frames LLMs as broad general knowledge engines, while RL provides the mechanism to scale reliable action-taking, long-horizon behavior, and self-improvement loops in real environments.
Real-world product usage is an evaluation strategy, not just distribution.
Reflection prioritizes shipping to customers because the most meaningful tests of autonomy are messy, real tasks; lab benchmarks can lag or mismeasure what matters in practice.
At massive model scale, simple methods win—if executed with meticulous craft.
From Gemini, he learned that straightforward objectives and relatively simple RLHF-style algorithms often outperform “fancier” ideas when supported by high-quality infrastructure, data, and careful implementation details.
WORDS WORTH SAVING
5 quotesIt was so smart that everyone thought it was dumb.
— Misha Laskin
What that meant was that an AI system had discovered a strategy that was fundamentally more creative.
— Misha Laskin
Our belief is that if you solve the problem of autonomous coding, you will solve the superintelligence problem more broadly, and that's kind of our path.
— Misha Laskin
The simple ideas implemented at a great level of detail are the things that work.
— Misha Laskin
In some sense, like, boredom is a gift that you only appreciate in retrospect.
— Misha Laskin
High quality AI-generated summary created from speaker-labeled transcript.
