In March 2016, something extraordinary happened in a quiet Seoul auditorium. An AI named AlphaGo, developed by DeepMind, made a move in a high-stakes match against world champion Lee Sedol—a move so unexpected, so counterintuitive, that even its own creators thought it might be a bug.
That moment—dubbed “Move 37”—didn’t just win a game. It revealed a glimpse of something far more profound: artificial superintelligence (ASI) in action.
Today, nearly a decade later, that same spirit of discovery is driving a new generation of AI pioneers—like Misha, CEO and co-founder of Reflection AI—who believe we’re on the cusp of a revolution far more transformative than automation or chatbots. We’re entering an era where AI doesn’t just assist us—it expands human imagination itself.
The Move 37 Moment: When AI Outsmarted Human Intuition
To understand the significance of Move 37, you need to appreciate the game of Go. With more possible board configurations than atoms in the observable universe, Go has long been considered the ultimate test of strategic depth and human intuition.
When AlphaGo placed its 37th stone in an empty corner—apparently sacrificing territory—experts were baffled. Lee Sedol himself paused, visibly unsettled. Commentators speculated it was a glitch.
But within ten moves, the brilliance became undeniable. AlphaGo had orchestrated a long-term trap, revealing a strategy no human had ever conceived.
“It was so smart that everyone thought it was dumb.”
That paradox captures the essence of emergent intelligence. The system hadn’t just learned from humans—it had invented something new, pushing the boundaries of what was thought possible in the game.
This wasn’t just a win for AI. It was a proof of concept for superintelligence: systems that don’t merely replicate human knowledge but generate novel, superior insights.
Today, that same principle is playing out across mathematics, drug discovery, code generation, and scientific research. Imagine a mathematician asking an AI to prove a theorem—and the AI returning a solution so elegant, so unexpected, that it rewrites textbook understanding. That’s the “Move 37 effect” scaled across every domain of knowledge work.
Reflection AI’s Bold Bet: Autonomous Coding as the Path to Superintelligence
So how do we get from today’s large language models (LLMs) to true artificial superintelligence?
For Misha and his team at Reflection AI, the answer lies in autonomous coding.
While most AI labs focus on making chatbots better at answering questions, Reflection believes the real breakthrough comes when AI can operate independently on a computer, completing complex tasks from start to finish—without human prompting at every step.
Why Coding? Because It’s AI’s Native Language
Here’s a key insight most people miss: language models don’t “see” the world like humans do.
We navigate software through graphical interfaces—clicking icons, dragging windows, using a mouse. But for an LLM, that’s like asking a poet to perform surgery with oven mitts.
Instead, AI’s natural “hands and legs” are code.
- Code is structured, logical, and abundant in AI training data.
- APIs, scripts, and command-line tools are far more efficient interfaces for AI than pixel-based UIs.
- As Misha puts it: “Coding is intuitive to language models in the same way spatial reasoning is intuitive to us.”
This leads to a radical prediction: autonomous coding won’t just replace software engineers—it will transform every knowledge worker.
From marketing analysts automating data pipelines to biologists running simulations, the future of work on a computer will be mediated through code—even if the user never writes a single line. AI will act as a universal “embodied agent,” executing tasks via programmatic interfaces.
Solving autonomous coding, Reflection argues, is equivalent to solving general intelligence on a computer.
The Physics Mindset: Simplicity, Clarity, and First Principles
Misha’s journey—from Soviet-era Russia to AI entrepreneurship in Silicon Valley—is deeply shaped by his background in physics. And that training informs Reflection’s entire approach.
In physics, complex phenomena are reduced to first principles: fundamental laws that explain everything from planetary motion to semiconductor behavior. Similarly, in AI, Reflection avoids over-engineering.
Their philosophy, forged during the development of Google’s Gemini models (where Misha co-led post-training efforts), is clear:
Simple ideas, executed with extreme precision, beat complex hacks at scale.
Consider this contrast:
- 1997’s Deep Blue (which beat Kasparov in chess) used brute-force tree search with thousands of hand-coded rules.
- Today’s LLMs learn from data alone, using a deceptively simple objective: predict the next token.
Despite their 600+ billion parameters (Gemini Ultra), these models thrive on elegant, scalable algorithms—like Reinforcement Learning from Human Feedback (RLHF)—not labyrinthine architectures.
As Misha notes: “In the era of massive scale, craftsmanship matters more than complexity.”
This physics-inspired mindset also guides startup strategy: identify the one or two core problems that truly move the needle, and ignore everything else.
Why Startups, Not Big Labs? Speed, Focus, and Real-World Feedback
After helping build Gemini at Google, Misha and co-founder Giannis (a key architect of AlphaGo and AlphaZero) could have stayed at one of the world’s largest AI labs.
Instead, they chose to launch Reflection. Why?
Because big labs optimize for chatbots; startups optimize for autonomy.
Large organizations are slow to pivot. Their product roadmaps are set years in advance. But building superintelligence requires rapid iteration, deep coupling between research and product, and—most critically—real-world validation.
“The evaluation that matters most is real-world evaluation,” Misha insists.
At Reflection, AI agents aren’t judged by benchmark scores—they’re tested by actual users solving real problems. This tight feedback loop accelerates learning in ways closed-off research labs simply can’t match.
Humans as Architects: The Future of Work in an AI-Powered World
One of the most persistent fears about AI is job displacement. But Misha offers a more optimistic—and realistic—vision:
This isn’t a zero-sum game. It’s a creativity multiplier.
Historically, technology hasn’t eliminated work—it expands the frontier of what’s possible. The printing press didn’t put scribes out of business; it created publishers, editors, and novelists.
Similarly, superintelligent AI won’t replace knowledge workers—it will amplify their impact.
Imagine a software engineer in 2030:
- No longer debugging syntax or writing boilerplate.
- Instead acting as a “software architect,” defining high-level goals and orchestrating teams of AI agents.
The same applies to scientists, designers, analysts, and entrepreneurs. Execution becomes automated; strategy becomes paramount.
The ultimate skill of the future? Asking the right questions.
As Misha reflects: “If you have a superintelligent AI, the challenge isn’t getting it to work—it’s knowing what to ask it to do.”
The Art of Asking the Right Questions (And Why It’s So Hard)
Here’s a sobering truth: even brilliant researchers often ask the wrong questions.
Misha cites his own highly cited paper (“CURL”)—a solid contribution, yes, but one that addressed a locally optimal problem, not a transformative one. Meanwhile, the teams that asked “What if we trained a model on all human text?” gave us GPT—and changed the world.
So how do you cultivate the ability to ask the right questions?
Misha’s answer: clarity through writing and dialogue.
- Writing forces precision. When you articulate an idea in prose, gaps in logic become obvious.
- Critical discussion with trusted peers exposes blind spots you’d never see alone.
It’s not about having all the answers—it’s about refining the questions until they point toward true leverage.
Building a Startup in the Age of ASI: Mission, Team, and Momentum
Launching a company like Reflection isn’t for the faint of heart. But Misha identifies two non-negotiables:
- An ambitious, clear mission (“Build superintelligence”) that attracts top talent.
- A concrete near-term wedge (autonomous coding) that delivers real value today.
Early hiring is make-or-break. The first three hires must be exceptional—not just skilled, but believers in the vision. Because as Misha puts it: “Good people beget good people.”
And setbacks? They’re not failures—they’re data points. The real risk isn’t pivoting; it’s persisting with a flawed assumption.
“If you’re not experiencing course corrections, you’re probably not learning.”
Final Thought: The Gift of Boredom—and the Pursuit of Impact
Misha’s journey began in solitude—reading his parents’ books in a new country, grappling with loneliness, and cultivating a love for physics and literature.
Looking back, he now sees boredom as a gift: the space where curiosity and deep thinking take root.
Today, that same curiosity drives him toward what he calls “the most impactful science of our time.”
We’re not just building smarter tools. We’re creating collaborative intelligences that will help humanity solve its grandest challenges—from climate modeling to disease eradication.
And it all started with a single, beautiful, misunderstood move on a Go board.
The future isn’t about humans versus AI. It’s about humans with AI—asking better questions, creating more boldly, and reaching further than ever before.
Move 37 was just the beginning.
0 Comments