From research physicist to building a $2.1B superintelligence startup

Executive overview

Most AI labs are building better chatbots. Reflection AI is pursuing autonomous coding as the direct path to superintelligence. Misha Laskin left Google DeepMind after Gemini 1.5 because large-lab incentives steer toward incremental product improvements, not the autonomy problem he actually cares about.

Solving autonomous coding is the same problem as solving superintelligence — because code is the natural language of AI interacting with computers.

The move 37 thesis

• AlphaGo's move 37 looked like a bug; 10 moves later it was revealed as the most creative move in Go history
• It was so smart that everyone thought it was dumb — a signal of genuine superhuman intelligence
• The same phenomenon will emerge across all knowledge work: AI returning "move 37s" in math, research, engineering
• When it happened in Go, it expanded human knowledge of the game — players learned from the AI
• Laskin expects this pattern across domains within a few years, marking the arrival of ASI

Why autonomous coding unlocks general intelligence

• Language models were trained on the internet, not spatial experience — code is their natural embodiment
• Humans interact with computers via spatial UI (mouse, cursor); LLMs will interact via programmatic interfaces
• Most software will develop "language-friendly" APIs; the dominant interface for AI will be code
• Autonomous coding is not just a tool for software engineers — it transcends engineering and applies to every computer-based task
• Solving autonomous coding on a computer = solving intelligence on a computer

Why leave DeepMind to start a company

• Post-Gemini 1.5, the lab's paradigm was focused on more capable chatbots, not autonomy
• Real-world evaluation is the only evaluation that matters — you need customers and product to know if your technology works
• A small, focused team moves faster on research when tightly coupled to product feedback
• A large organisation moving in a fixed direction is nearly impossible to course-correct from the inside
• The wedge at Reflection: focus solely on autonomous coding, nothing else, as the bet toward superintelligence

Lessons from building Gemini at scale

• Pre-scaling era: complex, mathematically sophisticated ideas won at small model sizes
• Scaling era: the opposite — simple ideas implemented with extreme craft and detail win
• Gemini RLHF algorithms are very simple relative to what RL researchers theorised a decade ago
• Simple objectives (next-token prediction, basic RL) trained on massive scale outperform complex architectures
• Investment in infrastructure to run large models efficiently matters as much as algorithmic choices

Asking the right questions

• The hardest skill in research and company-building is picking the right problem, not executing the solution
• A highly-cited paper (Curl, ~1,000 citations) asked a locally correct but fundamentally wrong question; the right question would have yielded citations in the millions
• Example: researchers asked "how do we make RL 10x more data-efficient?" — the right question was "how do we build language models?"
• Picking the right thing requires clarity of thought; writing forces that clarity by exposing holes in reasoning
• Discussing with trusted, critical people — ones who challenge, not validate — is the second key method
• Laskin spent months self-teaching RL and producing independent research before approaching top labs — demonstrating intent through action, not words

Building the team and company

• First challenge: reduce the blank slate to a directed, concrete short-term thesis aligned with the long-term mission
• Second challenge: getting the best people requires already having the best people — calibre compounds
• First three hires must be extremely high-calibre people with strong trust; quality begets quality
• Ambitious mission (superintelligence ≈ "taking people to Mars") attracts talented people who want the hardest problems
• A smaller, less ambitious previous startup was harder to recruit for — people optimise for impact over comfort

On human-AI collaboration and the future of work

• Intelligence is not zero-sum — every technological advance increases total output, not just reallocates it
• Knowledge workers will become architects managing an AI workforce, not replaced by one
• The scarce resource becomes asking the right questions and designing the right problems
• Execution will increasingly be delegated to AI; judgment on what to build is what compounds in value
• Setbacks reframe as learning signals — only doing the wrong thing forever is a true setback

Advice for founders and researchers

• Surround yourself with people who are better than you at what you care about — environment shapes trajectory more than raw talent
• Demonstrate desire through action: build something, bring concrete work, don't rely on words or cold emails alone
• Deeply caring about the problem and the people you work with makes setbacks feel like data, not failures
• Momentum — constantly learning and taking action — is more important than avoiding mistakes
• Access to talented, ambitious people is available with sufficient effort and demonstrated commitment