Jul 17, 2025 · 1h 6m · mad
Ex‑DeepMind Researcher Misha Laskin on Enterprise Super‑Intelligence | Reflection AI
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews Reflection AI co-founder and former DeepMind researcher Misha Laskin about building organizational superintelligence for enterprise software engineering. Laskin discusses his career trajectory, the launch of Reflection AI's new code research agent Asimov, and why solving codebase context and multi-agent retrieval is essential for achieving enterprise AI autonomy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 18.1% of the talking time here. How this is scored →
Matt as informed peer 3.1 Guest teaching 4.4 Guest disagreement 1.3 Matt pushing back 0.9
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 49:45 Direct correction of Bitter Lesson premise
Misha firmly rejects the host's framing on generalization by stating that Rich Sutton's Bitter Lesson actually does not say anything about generalization, but rather about building systems that scale with compute and search.
Hardest push from Matt ▶ 49:45 Challenging generalization narrativeMatt presses Misha by invoking the widely cited Bitter Lesson, challenging Misha's argument about specialized jagged models by arguing that generalization is expected to win out.
Biggest teaching moment ▶ 49:50 Schooling on Sutton's Bitter Lesson and jagged RL capabilitiesMisha corrects common industry misconceptions regarding Rich Sutton's Bitter Lesson, explaining its true focus on compute scaling and demonstrating how post-training RL creates jagged rather than smoothly generalized capability.
Matt holds his own ▶ 49:45 Citing Sutton's Bitter Lesson in AI debateMatt demonstrates deep domain familiarity by bringing up Rich Sutton's Bitter Lesson to challenge Misha on whether specialized domain models contradict fundamental AI scaling dynamics.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Defining Organizational Superintelligence | 1 | 4 | 1 | 0 | Matt introduces the episode and asks Misha to explain Reflection AI. Misha reframes superintelligence away from abstract math Olympiads to a concrete organizational oracle that deeply understands company context. | |
| Industry Terminology: Superintelligence Versus AGI | 4 | 5 | 2 | 2 | Matt references Mark Zuckerberg's aggressive talent poaching and asks if the industry agrees on superintelligence versus AGI. Misha uses an Indiana Jones sandbag analogy to argue that superintelligence is simply a moved goalpost replacing AGI, and reframes coding models as the primary hands and legs interface for software. | |
| Addressing the Code Comprehension Gap in AI Tools | 2 | 5 | 1 | 0 | Matt reacts to Misha's metaphor of L-9 engineers with amnesia. Misha explains that engineers spend 70% of their time collecting context rather than writing code, arguing that tools missing this comprehension gap fail to solve real software engineering problems. | |
| Comparing RAG, Agentic Search, and Contextual Intelligence | 4 | 5 | 2 | 2 | Matt asks how this approach differs from RAG and queries why traditional RAG is primitive. Misha bluntly critiques existing RAG as sparse and weak, offering a metaphor of current agentic search as navigating a dark jungle with a tiny flashlight. | |
| Code Research as a Superintelligence-Complete Problem | 3 | 5 | 3 | 2 | Matt attempts to restate Misha's argument about self-improving AI systems. Misha rejects the abstract framing of intelligence building intelligence, arguing that research disconnected from product co-design is now meaningless. | |
| The Four Historical Breakthroughs Required for Superintelligence | 3 | 5 | 1 | 1 | Matt prompts Misha to list the core technical components needed for superintelligence and asks whether reasoning models replace RLHF. Misha outlines four historical breakthroughs and educates Matt on how alignment RLHF and reasoning RL operate as distinct, parallel capabilities. | |
| Introducing Asimov: The Code Research Agent | 2 | 4 | 1 | 0 | Matt invites Misha to announce Reflection AI's new product. Misha details Asimov, explaining how it creates permanent team-wide organizational memory rather than isolated user memory. | |
| Multi-Agent System Design and Retrieval Strategies | 4 | 5 | 1 | 1 | Matt asks detailed questions about multi-agent architecture and asks if data protocols like MCP are required. Misha maps out the spectrum from RAG to neural retrieval and explains why stateless protocols like MCP differ from indexed knowledge engines. | |
| Enterprise Knowledge Indexing and VPC Deployment | 3 | 4 | 1 | 1 | Matt asks if enterprise deployment requires working in air-gapped VPC environments and how RL is applied in Asimov. Misha confirms VPC deployment is a strict dealbreaker for enterprise clients and explains their pragmatic strategy of targeted RL on top of third-party models. | |
| Misha's Journey: From Russia to Theoretical Physics and AI | 4 | 3 | 0 | 0 | Matt inquires about Misha's personal journey, drawing parallels with other Russian-born AI figures and discussing his upbringing near the Hanford nuclear site in Washington state. Misha explains transitioning from theoretical physics because AI felt like the active science of our time. | |
| AI's Impact on Science and the Nobel Prize Controversy | 4 | 4 | 2 | 1 | Matt asks if AI is eating other scientific disciplines, citing recent Nobel Prizes. Misha gently pushes back, noting that the Nobel Physics award reflected historical physics-like mathematical models rather than AI taking over physics research. | |
| Y Combinator Startup, UC Berkeley Postdoc, and Early AI Collaborators | 2 | 3 | 0 | 0 | Matt asks about Misha's career path post-physics. Misha describes running a YC startup and his postdoc in Peter Abbeel's UC Berkeley lab alongside prominent AI founders and researchers. | |
| DeepMind Research, Unsupervised RL, and Google Gemini | 2 | 4 | 0 | 0 | Matt asks about Misha's transition to DeepMind. Misha explains his initial focus on unsupervised RL with Vlad Mnih before pivoting to lead reward model training for Google Gemini 1.0 and 1.5. | |
| Founding Reflection AI and the Architecture of Superintelligence | 5 | 5 | 3 | 3 | Matt pushes back against specialized models by invoking Rich Sutton's Bitter Lesson and suggesting general models will always win. Misha directly corrects Matt's interpretation, clarifying that the Bitter Lesson concerns compute scaling rather than generalization and explaining how RL produces domain-jagged capability. | |
| Current State and Future of Autonomous Coding Agents | 3 | 5 | 2 | 1 | Matt asks about current benchmarks like SWE-bench and progress toward full coding autonomy. Misha highlights the gap between high SWE-bench scores and real-world utility, explaining that current autonomy is at a junior L-4 level. | |
| Building Reflection AI: Recruiting Talent, Team Setup, and Capital Strategy | 3 | 4 | 1 | 1 | Matt asks how Reflection AI recruits top talent against tech giants throwing around massive compensation packages and asks about capital requirements. Misha explains how mission-driven startups win over talent wanting equity upside and early frontier lab dynamics. |
Statements from this episode (24)
Laskin: Organizational superintelligence will probably be an all-encompassing oracle
“We realized that probably the form factor of an organizational super intelligence, like the thing that's really gonna help organizations get a lot of stuff done, is probably gonna be something like an oracle. Like an oracle that understands the entire organiza…”
Laskin: Large labs use superintelligence synonymously with AGI
“Superintelligence in these contexts of the large lab context is actually just being used synonymously with what AGI used to be used for.”
Laskin: AI models will interact with enterprise software primarily via APIs
“And so the way these language models are going to interact with any piece of software, not just Software engineering software, like Salesforce and other CRMs and creative tools and so forth. The majority of those interactions are going to be through function c…”
Misha Laskin: All existing AI coding tools focus exclusively on code generation
“Every coding tool today is focused on the code generation piece”
Laskin: Advanced AI without comprehension are 'L9 engineers with amnesia'
“What we're going to get to if we don't solve the comprehension piece is basically L-nine engineers with amnesia.”
Misha Laskin: Engineers spend 70% of their time searching information, not coding
“If you actually hover over a shoulder of an engineer at any large organization, or even a startup with a really sizable code base, and you look at what they do, you'll find that a minority of their time is actually spent coding. Like, 70% of the time they're D…”
Laskin: Traditional RAG agents fail for any meaningful software engineering query
“It'll grab it, and then that's all you have, and most likely, for any meaningful query it will not have given you the information that you need to actually go do the task. So rag agents are actually, are pretty weak.”
Laskin: Current agentic search tools do not scale to large codebases
“That's basically where Agentex search is today, where you're going and you're kind of exploring, and you have this tiny flashlight, and you have to then remember everything in your head, and obviously that's a first form of comprehension, but it's also a prett…”
Laskin: Solving organizational code context yields all capabilities for superintelligence
“Like if you really solve this oracle for organizations just for coding, you've basically built all the capabilities you need to have super intelligence.”
Laskin: Self-improving code AI will mainly make algorithms more efficient
“I think what, I think coding intelligence that builds better coding intelligence will Make algorithms more efficient, basically.”
Laskin: The core ingredients to build AGI and ASI are now known
“But that was only meaningful when the ingredients for how to build artificial general intelligence, or ASI were not known. I think now they're known, and so.”
Laskin: Base language models before alignment are useless stochastic parrots
“If you play with one of these base models before they're aligned, they're really useless. They're, they are, they feel like stochastic parrots. They don't follow instructions.”
Laskin: Other AI companies will converge on multi-agent retrieval architectures
“I'm sure that other companies will converge on it as well.”
Laskin: Pre-LLM AI breakthroughs spent more effort on environment design than model training
“And a lot of the project was in these big projects was not even on training the models. It was figuring out how the agents should actually interface with the environment that you're training it in.”
Laskin: AI coding tools are thin wrappers with low switching costs
“It's really easy to switch around between various different coding providers today because they don't really integrate deeply, and so, you know, you can try cursor today, you can try cloud code tomorrow, you can switch, you know, to windsurf, and there's reall…”
Laskin: AI work awarded Physics Nobel Prize has had little impact on physics
“What's interesting is that the Physics Nobel Prize was given to something that has not really had that much impact in physics, but it is but I still buy it because it's kind of there's a physics smell to the breakthroughs that led to you know, these systems li…”
Laskin: Google Gemini's initial RLHF team was only 10 to 20 people
“I joined a small project at the time that you know, was tens of people. And that project became Gemini one and 1.5, and then obviously two and so forth. And I joined with my co-founder, my co-founder, Yannis was leading the reinforcement learning team, the RLE…”
Laskin: Reinforcement learning makes LLM capabilities jagged, not broadly general
“When you train large language models with reinforcement learning, they become jagged in the sense that they become good at what you wanted them to be good at. And there are some generalization capabilities, but they're much weaker than people think.”
Laskin: Superintelligence will emerge from multiple specialized labs, not one company
“I do think there'll be a general super intelligence, but I think that it won't be one lab that has built it, but it'll be kind of the plurality, like the collection of all intelligences will be a general super intelligence.”
Laskin: Coding agents are currently at an L4 junior engineer level
“There is in some sense, we are probably at a L four kind of junior, junior engineer level of autonomy, which is pretty incredible.”
Laskin: Principal-level AI engineers are a couple of years away
“And that the combination of this you know, L-Nine with Amnesia and the L-Nine's context core will together, you know, that will become the principal level engineer, the AI engineer. And so I actually think that that's not too far away. That's I would say in, y…”
Laskin: Reflection AI regularly beats OpenAI, Anthropic, and DeepMind for talent
“We win over candidates over OpenAI and Anthropic Meta, DeepMind regularly.”
Laskin: Reflection AI will ship research requiring 100k GPU equivalence in 2025
“Later this year we'll be shipping things that I don't think anyone ever thought a startup could do. Like, I think that we're going to be shipping some things on the research side that I think everyone thinks you need to be a giant lab with a 100,000 GPUs to do…”