Summary

Reflection AI’s core bet is that enterprise coding is bottlenecked by comprehension, not generation. Misha Laskin says the world today is focused roughly 80% on action and 20% on understanding, while engineers spend about 80% of their time understanding systems and collaborating. In complex organizations, measured productivity gains can be “sometimes negative, sometimes negligible.”
Asimov is designed as an organizational oracle rather than another autocomplete layer. It targets semantic questions spanning code, chats and project tools—the equivalent of having a “principal-level engineer” explain why a test is flaky or why jobs suddenly run five times slower. Solving that context layer should make subsequent agents more reliable; action is “more or less solved,” though Misha stops short of calling it trivial.
Reflection’s prospective moat is a tightly coupled loop from customer problem to evaluation, agent design, post-training and product. Frontier labs spread effort across hundreds of evals, while a focused startup can reverse-engineer capabilities from concrete needs—for example, compressing engineer onboarding from months to a couple of weeks on a codebase that might contain 100 million tokens. “The only advantage that you’ll ever have as a startup” is focus and velocity.
Open-weight models have created a temporary economic opening for new frontier labs. Their quality exceeded Misha’s expectations, while RL post-training currently needs roughly two orders of magnitude less compute than pre-training. The capital requirements remain high and may converge asymptotically, but Reflection’s thesis is that product revenue can finance the climb without surrendering to a cloud provider.
RL is reward-bound before it is compute-bound. LLM-as-judge rewards are noisy and eventually get hacked, ground-truth rewards are scarce, and a verifier that could score every arbitrary task would itself be superintelligent: “The reward problem…is ASI complete.” Today’s algorithms also lack granular credit assignment, producing “garden path meandering” reasoning instead of identifying which step was right.
Misha expects jagged superintelligence, not a sudden general intelligence that makes every remaining task free. His deliberately aggressive framing is that “there’s no such thing as generalization, there’s just bringing the test distribution into train”; within a couple of years, meaningful slivers such as front-end development might become superhuman because their data is economical to create. The blueprint may arrive quickly, but deployment across work categories remains a multi-decade undertaking.
Verticalization makes model-dependent coding products strategically exposed. Claude Code illustrates that a frontier lab can successfully couple product and intelligence, while labs can subsidize end-user products more deeply than startups paying their inference margins. Misha therefore sees companies in “critical path categories” such as search and coding as existentially vulnerable unless they own frontier intelligence or are deeply embedded in customer workflows; he nevertheless expects the next generational coding companies to be built now.

Deep dive

1. Reflection is building backward from deployed superintelligence

Misha’s path began in theoretical physics, inspired by the Feynman lectures and by physics’ role in technologies such as transistors and GPS. Watching AlphaGo convinced him AI was undergoing the dynamic scientific birth that physics had experienced roughly a century earlier, prompting him to enter Pieter Abbeel’s reinforcement-learning lab.
Reflection’s distinction is less the word “superintelligence” than its method: instead of maximizing benchmarks in an isolated research lab and finding a product afterward, it is “co-designing product and research together.” The cost is narrower initial focus; the requirement is choosing an “ASI-complete” category broad enough to draw out genuinely powerful capabilities.
At Gemini, co-founder Ioannis Tsantéoglou led RL for Gemini 1 and 1.5, with Misha working closely on the team. They concluded that scaling RL over large language models was “effectively the final paradigm” needed before AGI—or the now-shifted goalpost of ASI—even though Misha believes the field remains earlier than most observers assume.

2. Asimov bets that coding’s 80/20 split is backwards

Misha describes Asimov as “the best code research agent in the world”: a comprehension system intended to feel like deep research over a large codebase, with a principal-level engineer’s institutional understanding available on demand.
The enterprise problem differs from startup greenfield coding. Once no engineer can hold the entire system in mind, knowledge fragments across repositories, chats, project-management tools and individual memories; coding-agent productivity then becomes “sometimes negative, sometimes negligible,” consistent with what Misha hears from engineering leaders.
Existing products are roughly “eighty percent kind of action, twenty percent understanding,” but actual engineering work is close to the reverse. Collaboration itself is often one person asking another about an unfamiliar system, so missing context—not an inability to emit code—is what prevents agents from operating intelligently inside organizations.
The endpoint is an “omniscient oracle for organizations” that answers questions at any complexity level using both formal artifacts and tribal knowledge. Once that is solved, reliable action follows more easily: Misha calls agent action perhaps 20% of the total problem and “more or less solved,” while explicitly declining to call it trivial.

3. Customer-derived evals are the startup’s real wedge

The system combines product plumbing with post-trained capabilities: small models with very long contexts perform something like “neural retrieval,” absorbing large amounts of a codebase and reasoning out relevant information, while a higher-level agent handles search, tool use and multi-hop decisions across sources such as Jira.
General training across thousands of environments—the approach illustrated by Kimi 2—provides an agentic base analogous to pre-training. Depth comes from training against the exact tools the product will use; Misha cites Grok 4’s tool-trained improvement on Humanity’s Last Exam, while questioning whether that benchmark has more than a weak correlation with end-user value.
At a large incumbent, a researcher may be one of hundreds of evals and therefore spread thin. Reflection instead translates customer behavior into tests, then adjusts every layer—product features, agent architecture and model training—rather than remaining “at the mercy” of general third-party models.
Onboarding makes the loop concrete: a large company may need months to acclimate an engineer to a codebase containing perhaps 100 million tokens. Reflection’s target is a couple of weeks, deriving long-context and retrieval requirements by inspecting what prevents Asimov from answering the new hire’s actual questions.

4. Organizational memory becomes a Git-like control plane

Asimov is not meant for a quick question about the file already open; waiting tens of seconds or a minute would be unnecessary. Its advantage appears when the query is semantic and the engineer knows the symptom but neither the relevant function nor where the explanation lives.
Reflection’s own example was, “My jobs are running slowly, five times more slowly than usual. Why is that?” Two teams had submitted individually valid pull requests that interacted like a race condition and slowed everyone down—a cross-system bug that might otherwise consume two or three engineers for several days.
Team-wide memory lets engineers teach organizational semantics—for example, what that team means by “environment jobs”—but immediately creates authority and permissions problems. Early customers begin with trusted staff-level or more senior engineers as gatekeepers, then expand participation.
Sarah noted the difference between static role-based access and evolving, Git-like review as expertise changes with the codebase. Misha expects memory edits to resemble pull requests approved by current component owners: “GitHub plus plus,” versioning not code but the metaknowledge models need—effectively “a new Git from scratch.”

5. Open weights create a temporary post-training window

Reflection began with the bet that internet-scale pre-training was converging on a known recipe with diminishing signal extraction from an extremely noisy dataset. It expected adequate open-weight bases; Misha’s update is that the open frontier has been better than he anticipated.
The present opening is that “the RL flops are still manageable.” Sarah summarizes the gap as two orders of magnitude from pre-training, and Misha agrees: focused post-training still needs substantial GPUs, but a startup can plausibly produce a best-in-class vertical system without first financing a general foundation model.
This is not a low-capital destination. Misha expects a focused company may still need roughly an order of magnitude less compute than a frontier lab today, while requirements could become asymptotically similar; the strategy is to build a generational revenue business capable of raising that capital without an acquisition or cloud-provider dependency.

6. RL remains reward-bound, not merely compute-bound

Asked why RL cannot yet absorb pre-training-scale compute, Misha places more weight on problem structure than algorithms. Having built reward models for Gemini 1 and 1.5, he argues that an accurate reward for any arbitrary task would let even simple current methods extract substantial capability, bounded mainly by exploration.
The field does not possess those rewards. LLM-as-judge rubrics work partially, but “a noisy or stochastic reward inevitably gets hacked”; ground-truth sources are clean but scarce, forcing researchers to hope that optimizing narrow verifiable tasks generalizes or to assemble messy synthetic-data pipelines.
His categorical diagnosis is that “we’re in a reward-bound world.” A neural verifier capable of accurately judging every outcome would probably already be a superintelligence, which is why the reward problem moved in his vocabulary from “AGI complete” to “ASI complete.”
Algorithms contribute a second bottleneck through weak exploration and credit assignment. They essentially “take the things that work and make them happen more frequently” without identifying which reasoning step deserved credit, producing long, garden-path chains that explore irrelevant branches instead of resembling structured human thought.

7. Superintelligence will arrive jagged before it arrives general

Misha is bullish on synthetic replicas of Airbnb, Amazon, Salesforce or Excel because they move desired behavior into the training set. His intentionally aggressive claim—prompting Sarah’s “Wow”—is that “there’s no such thing as generalization, there’s just bringing the test distribution into train,” though he concedes some genuine generalization occurs.
He still believes meaningful superintelligence could appear within a couple of years, but only in slivers of work rather than all coding. A superintelligent front-end developer is plausible where internet data and synthetic generation are abundant; extending the recipe elsewhere becomes an economic decision about data cost and return.
Atari, AlphaGo, Dota 5 and AlphaStar provide his analogy: they established a blueprint for narrow superintelligence, and more compute could have pushed the latter systems further, but doing so lacked economic value. The research recipe can therefore be solved before every application merits implementation.
Sarah calls this partly definitional: broad capability was supposed to make the remainder arrive “for free.” Misha instead predicts “jagged superintelligence.” AMIE is completely saturated; Codeforces and other competitive-coding environments are nearly at the top, with models likely to become the best in the world within the year, while workplace impact remains uneven. The meaningful threshold is double-digit productivity for every engineer, because “deployment of it is half the problem.”

8. Verticalization puts model-dependent coding startups at risk

Misha reads the Windsurf non-acquisition through a broader verticalization trend: search through ChatGPT was an early case, and coding is now a frontier-critical category. At the scale of these companies—basically trillion-dollar companies or more—economics is not really the driving factor; verticalization is instead tied to sustaining frontier research.
Frontier labs can still fail because buying a product does not remove the distance between research and users or the distraction of a hundred focus areas. Claude Code is his example of successful verticalization, illustrating why tight model-product coupling matters without proving every incumbent can purchase its way to adoption.
The startup-side danger is sharper: a company built on someone else’s intelligence pays—and may subsidize—the supplier’s margin, while Google, Anthropic or OpenAI can subsidize their own products far more aggressively. In critical categories, survival therefore requires proprietary frontier models or customer integration deep enough to make removal difficult.

9. Every domain needs its own data, rewards and deployment stack

Sarah floated recording a software team’s code, Slack, documentation and conversations as a dataset with a potential floor value of millions of dollars; Misha responded, “especially if there’s diversity.” Human reasoning traces remain logistically difficult to collect, making outcome verification and synthetic RL more scalable than asking experts to narrate every thought.
Misha’s robotics experience explains the distinction: visual and sensory rewards are “infinitely more hackable” than language rewards because language can be viewed as “a compressed representation of the world.” RL works cleanly in locomotion, where velocity and body damage are measurable, but manipulation introduces the noisy external world; teleoperation is attractive because humans are already master manipulators.
Coding can expand operationally because, if language models interact with almost any software through function calls, then code becomes the operational interface. Reflection first wants the deepest comprehension system for engineers, then a “contextual core” usable by coding agents and adjacent roles such as product management, technical support and eventually sales—without chasing those shiny adjacencies before owning the initial category.
Misha expects the blueprint for category-specific ASI within a couple of years, but deployment to remain a “multi-decade thing,” with snowflake environments and specialized post-training like Codex and Deep Research. Coding should be among the faster categories—“within the next, say, dozen or so months,” not a decade—and he expects more depth-first companies to emerge.

Ex-Google CEO: What Artificial Superintelligence Will Actually Look Like w/ Eric Schmidt & Dave B Windsurf x Google x Cognition: Full Breakdown: Who Made Money, Who Did Not