EO StudioFrom Google DeepMind to a $8B Superintelligence Startup | ReflectionAI, Misha Laskin
CHAPTERS
- 0:00 – 1:00
AlphaGo’s Move 37 and the first felt glimpse of superintelligence
Laskin recounts how AlphaGo’s famous “move 37” shifted his intuition from strong AI to something that looked like genuine superintelligence. He uses the moment to illustrate how AI can appear wrong to humans before revealing a deeper, more creative strategy.
- •Move 37 initially looked like a bug or hallucination to observers
- •It later proved to be a brilliant, human-unimaginable strategy
- •The episode suggested AI creativity can surpass human priors
- •“Move 37 moments” could emerge across all knowledge work
- 1:00 – 2:01
From Move 37 to “ASI feels near”: creativity expansion across knowledge work
He extrapolates from AlphaGo to a near-future where superhuman insights regularly appear in math, science, and professional domains. The key idea is that these systems won’t just automate work—they’ll expand what humans consider possible.
- •Anticipates “move 37s” in many fields (proofs, strategies, discoveries)
- •AGI today feels like sparks; ASI could feel imminent within years
- •AI can expand human creativity by revealing new solution spaces
- •Humans will learn from AI the way chess/Go players learn from engines
- 2:01 – 2:31
Reflection AI’s mission: build superintelligence via autonomous coding
Laskin introduces Reflection AI and its core bet: solving autonomous coding is a direct route to broader superintelligence. He frames the company as building not copilots, but capable autonomous systems.
- •Reflection’s stated mission is building superintelligence
- •Core thesis: autonomous coding is the path to ASI
- •Aim is autonomy and task completion, not just chat assistance
- •Team intends to build agents that operate end-to-end on computers
- 2:31 – 3:32
Technical foundation: combining LLM breadth with RL for autonomy
He explains the two ingredients he believes recently became “mature enough” to create autonomous intelligence: large language models and reinforcement learning. He also situates his and his co-founder’s experience building major systems as the backdrop for this approach.
- •LLMs provide broad competence; RL scales autonomy and goal pursuit
- •Belief: LLM+RL enables agents, not only assistants
- •Co-founder’s background includes DQN, AlphaGo/Zero architecture work
- •Laskin and team led post-training work for Gemini 1/1.5
- 3:32 – 4:32
Early life across countries, solitude, books—and why boredom helped
Laskin describes moving from Russia to Israel and later to Washington State, and how frequent moves shaped his childhood. Time alone with books fostered deep curiosity that later translated into technical ambition.
- •Frequent relocation reduced long-term childhood bonds
- •Spent significant time alone reading from parents’ library
- •Loneliness and boredom became catalysts for thinking and learning
- •Early interests split between literature and physics
- 4:32 – 5:32
Choosing physics: chasing “root-node” breakthroughs behind technology
He explains why physics felt like the most foundational lever for progress, using examples like transistors and GPS relativity. The goal was to work on breakthroughs that later generate entire technological eras.
- •Motivation: work on foundational science enabling downstream tech
- •Examples: transistor (Bardeen) and GPS (special relativity)
- •Desired to be at the origin point of future technological artifacts
- •Physics provided a rigorous reductionist mindset
- 5:32 – 7:03
Pivot from physics PhD to AI: AlphaGo as the field-defining proof point
During his PhD he realized the most impactful “science of the time” had shifted, and deep learning was accelerating. AlphaGo convinced him something fundamental was happening, prompting an intense self-teaching sprint to enter AI research.
- •Realization: many physics breakthroughs he admired were decades old
- •Difficulty seeing new “foundational” physics breakthroughs for him
- •Deep learning was taking off; AlphaGo demonstrated superhuman ability
- •He paused his path to self-teach AI and start independent research
- 7:03 – 8:36
Physics mindset applied to startups and research: find the 1–2 needle-movers
Laskin generalizes a framework: complex problems usually hinge on a small number of decisive factors. He argues this reductionist approach applies equally to autonomy research and company-building across product, customers, and execution.
- •Assumption: most domains have only 1–2 changes that truly matter
- •Rigor helps strip away noise and focus on fundamentals
- •Applies to research (e.g., autonomy training) and business (product/customer fit)
- •Emphasis on identifying what actually moves the needle
- 8:36 – 10:07
Why leave DeepMind: autonomy focus, real-world evaluation, and moving faster
He outlines the limits of large labs for his goals: big organizations optimize for existing directions, and evaluating autonomy requires product and customer feedback. A startup structure enabled tighter iteration loops and clearer steering.
- •Big labs excel at compute/talent but can be slow to change direction
- •Post-Gemini paradigm centered on more capable chatbots
- •Reflection prioritized autonomy as both a research and product problem
- •Belief: real-world evaluation with customers matters most
- 10:07 – 13:16
Gemini lessons: at massive scale, simple ideas + craftsmanship win
Laskin contrasts earlier eras where complex algorithms dominated with today’s scaling regime where simple objectives and careful implementation outperform. He highlights infrastructure, attention to detail, and straightforward RLHF-style methods as key to success at scale.
- •Modern models are orders of magnitude larger than earlier networks
- •In scaling era, simple objectives (next-token) are surprisingly effective
- •RLHF methods used at scale are relatively simple algorithmically
- •Biggest differentiator: engineering craft, detail, and infrastructure
- 13:16 – 16:17
Why autonomous coding is the fastest route to AGI/ASI on computers
He argues coding is not merely a developer tool but the natural “embodiment” interface for language models. Because LMs lack human spatial priors, programmatic interaction is more native than GUIs, making autonomous coding a general solution for computer work.
- •Autonomous coding goes beyond software engineering into all computer work
- •“Embodiment” for LMs is likely programmatic interfaces, not mouse/GUI
- •Code is intuitive to LMs due to training data; spatial reasoning is not
- •Thesis: solving autonomous coding ≈ solving computer-based intelligence
- 16:17 – 18:18
Non–zero-sum future of work: humans become architects of AI workforces
Laskin predicts AI will expand total production rather than simply replace workers. Professionals will shift from direct execution to directing, scoping, and questioning—managing teams of AI agents like an ‘AI workforce.’
- •Technological advances historically increase total output
- •AI could yield an order-of-magnitude increase in ideas/software/experiments
- •Roles shift toward architecture, orchestration, and strategy
- •Core human leverage becomes choosing problems and asking good questions
- 18:18 – 22:50
Clear thinking framework: writing and adversarial discussion to find the right questions
He discusses how hard it is to select the right research or product questions, giving examples of ‘locally good’ but globally mis-aimed questions. His practical method is to force clarity through writing and to pressure-test ideas with smart, critical peers.
- •Asking the right question is harder than executing once defined
- •Example: optimizing RL sample efficiency missed the coming LLM paradigm
- •Writing exposes holes; revision strips ambiguity and unnecessary parts
- •Seek challengers, not cheerleaders, to stress-test ideas
- 22:50 – 29:23
Building Reflection to $8B: early startup realities—focus, hiring, motivation, and setbacks
Laskin describes the hardest parts of early-stage company-building: turning a blank slate into a focused wedge and recruiting exceptional initial talent. He frames motivation as combining an ambitious mission with a credible near-term plan, and treats setbacks as learning signals rather than failures.
- •Early challenge: define near-term steps aligned with long-term mission
- •First hires are pivotal; “good people beget good people”
- •Mission attracts talent, but short-term wedge must be convincingly right
- •Setbacks reframed as evidence-driven course correction and momentum
- 29:23 – 31:49
Three personal lessons: surround yourself with the right people and prove intent through action
He closes with guidance to his younger self: the people around you largely determine your trajectory. He emphasizes earning access to great mentors and teams by demonstrating seriousness through concrete work, not just words.
- •Most impactful lever: surrounding yourself with high-quality people
- •Example: transition enabled by joining Berkeley lab despite non-AI background
- •Talented people respond to demonstrated effort and tangible outputs
- •Persistence and action can open doors that cold outreach alone cannot
