No Priors

Misha Laskin on why open intelligence wins

Misha Laskin· Co-founder & CEO of Reflection AI at Reflection AI
·~70 min·English·No Priors
Open SourceLLMTrainingReasoningAI SafetyBusiness Strategy
TL;DR

Reflection AI's CEO argues that open, ground-up intelligence puts real capability in more hands, tying together reinforcement learning, enterprise ownership, infrastructure strategy, safety, and scientific progress.

01Core Mental Model

The RL curve that never bends back

Laskin's team are reinforcement-learning believers, and their system never stopped improving — so capability is now set by how much compute you choose to spend, not by the method.

this reinforcement learning system never stopped learning. And if you look at our plots, they just keep going up and it's just a matter of compute

— Misha Laskin, No Priors
Key Insight
When the ceiling on capability stops being scientific and becomes financial, model building quietly turns from a research question into an investment decision — you no longer ask whether you can improve the system, only how much you are willing to spend to keep it climbing.
02The Key Decision

No suitable Western base, so they built the whole thing

With the good open models all coming out of China and no suitable Western base, Reflection decided to build its own end to end — because pre-training and reinforcement learning turned out to be too tightly coupled to split apart.

it turned out you actually do need to pre-train your model in order to make reinforcement learning work very well at scale

— Misha Laskin, No Priors
Key Insight
The company set out to do RL on top of someone else's open base, but in Laskin's assessment no suitable Western one was available — and the discovery that RL at scale leans on how the model was pre-trained collapsed two separate plans into one far more expensive commitment: own every stage of training.
03The Product Edge

Beam optimizes for how fast, not just how smart

Beam, Reflection's first open model, was tuned so an agent reaches the same answer in a fraction of the compute and time — three to four times more efficient than its class, and up to ten times versus larger models.

an important thing in model building is not just the capability, but how quickly an agent achieves a thing

— Misha Laskin, No Priors
Key Insight
Framing efficiency as a first-class goal, not a nice-to-have, is what makes an open model a practical workhorse for enterprises: a model that is as capable but far cheaper and faster to run changes the economics of deploying it, which is exactly the lever an open-weight business needs to pull.
04The Business Model

Renting tokens versus owning intelligence

Buying a token means renting a thin slice of someone else's full stack; as enterprises mature, they want to own their intelligence the way a grown-up buys a house instead of renting an apartment.

when you're buying a token, you're renting like a piece of a whole stack

— Misha Laskin, No Priors
Key Insight
This reframes an open-model company as a landlord-to-homebuilder: the money is not only in the free weights but in selling everything a customer needs to actually run them — the cluster management, inference and harness that a closed provider hides inside the token price.
05The Market Shift

The Linux moment for AI models

Laskin reports that at model gateways like OpenRouter and Vercel, the token mix flipped from roughly 70% closed to roughly 70% open in about six months, and he expects AI to follow operating systems: open takes the volume while a few closed players keep an enormously valuable prize.

the world is going to look not too dissimilar from operating systems where 95% plus of servers computers in the world run on an open source operating system like Linux

— Misha Laskin, No Priors
Key Insight
The operating-system analogy is a quiet argument against the fear that open models commoditize everything to zero: Linux took almost all server share without erasing Microsoft or Apple, which suggests open AI can dominate usage while leaving room for a handful of extremely valuable closed labs.
06Geopolitics

Open models are Trojan horses for infrastructure

A free, permissive model can pull a whole country's inference software, cluster tooling and chips along behind it — which, Laskin argues, is part of why releasing open models is geopolitically advantageous.

open models are Trojan horses for the infrastructure that they bring with them

— Misha Laskin, No Priors
Key Insight
This recasts open-sourcing from an act of generosity into an act of strategy: giving the model away cheaply is how you get the rest of the world standardizing on your stack, the same railroad-and-standards playbook the US once ran — and the reason, Laskin argues, that China's open releases function as a subsidy to Western enterprise.
07The Safety Argument

Openness is the default state of safety

Laskin argues cyber offense and defense cannot be cleanly separated, so banning offensive capability also strips the defense that fights back — and a few hundred closed-lab researchers can never cover the long tail of bugs that an open ecosystem can.

When you remove cyber offensive capabilities, you also remove cyber defensive capabilities

— Misha Laskin, No Priors
Key Insight
By grounding the safety case in the 1990s fight over strong encryption — which opened up and gave birth to the whole field of cybersecurity — Laskin turns the usual framing on its head: openness is presented not as the risky choice to be justified, but as the historical default, with closed secrecy as the exception that has to earn its keep.
08The Human Stakes

The model caught up to his PhD

Feeding models his own physics PhD thesis year after year, Laskin watched the answers climb from useless, to undergraduate, to correct PhD-level work, to genuinely new ideas — which is why the scientist in him is most excited about what this does for science.

I'm personally very excited about scientific progress

— Misha Laskin, No Priors
Key Insight
Using his own thesis as a fixed yardstick makes the capability jump personal and concrete in a way benchmarks never are: the work that once took him years to execute now returns in a chat box — and that compression of the discovery loop is what the scientist in him keeps coming back to, even as he points to gains in software engineering and the real-world sciences as well.