Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin

Summary · 5 min read instead of 1 h 10 m

Misha Laskin is the co-founder and CEO of ReflectionAI, which builds open-weight models. He earned a PhD in physics, then worked as a researcher at Google DeepMind on the reinforcement learning team for the first Gemini models. ReflectionAI has just released Beam, its first open model, a 500 billion parameter reasoning model.

Laskin argues that open models will take most global token demand, noting that gateway traffic has flipped from 70-30 closed-open to 70-30 open-closed, much as Linux came to run most servers. He says enterprises move from renting intelligence through closed APIs to owning it once their spending gets large, and that ReflectionAI's role is to supply the software and services that make open deployments work. He describes how Beam was built, with more compute spent on reinforcement learning than on pre-training, and says that run never stopped improving. Laskin calls China's open model ecosystem a benefit to the world but says the West needs a competing source. He also argues that openness improves safety, citing the history of strong encryption and Linus' law.

Minute by minute

1. 1:05

   Building a frontier lab took everything. Laskin says ReflectionAI spent the last 12 months setting up a lab to build frontier open intelligence, growing from around 30 people to around 300 and assembling pre-training, mid-training and reinforcement learning teams. Asked what was hardest, he says everything was, because you have to get 30 things right: talent, retention, culture, data, compute and the infrastructure that big labs like DeepMind already have.

2. 3:50

   Why ReflectionAI started pre-training. Laskin says the original bet, about two and a half years ago, was to apply reinforcement learning to domains like math and coding on top of someone else's open base model. Reinforcement learning worked faster than expected, but the good open models were all coming from China. He says the team also found that you need to pre-train your own model to make reinforcement learning work well at scale.

3. 7:31

   Catching up costs more each year. Laskin says catching up to the frontier took hundreds of millions of dollars 18 months ago, single-digit billions about six months ago, and will take order 10 billion next year. He estimates a roughly 4x compute multiplier per model generation, moving from 100,000 H100s to 100,000 Blackwells to around 100,000 Vera Rubins. Strong talent cuts down on exploration, which lets a lab catching up spend less.

4. 10:23

   Models now speed up their own training. Laskin agrees that capex cannot grow forever but says compute is being used more efficiently because models help build themselves. He puts gains at about 7x a year when only human researchers did the work and, in reinforcement learning, perhaps 30x depending on the model. He adds that the revenue each flop can generate also rises.

5. 12:16

   Beam's training spent more on RL. Laskin says Beam has 500 billion total parameters and 23B active. Pre-training used 6,000 GB300s for a few weeks, which infrastructure gains have since cut to about 12 days. Reinforcement learning used a little over 10,000 GB300s for four weeks. He says the RL curves never stopped rising, so further scaling becomes an economic choice about where to get data and what is valuable.

6. 17:25

   Jagged capabilities adapt to customer data. Laskin says coding and agentic training lay the foundation, and the model generalizes in surprising ways but remains jagged across harnesses such as versions of Terminal Bench. Once it has general agentic capability, it adapts quickly to new data. He says ReflectionAI builds evaluations for customer tasks and generates synthetic data, and names finance KYC and compliance flows, cyber defense and legal work as valuable areas.

7. 19:47

   Beam reasons with fewer steps. Laskin says Beam is three to four times more efficient than models of the same capability class and around 10x more efficient than larger models, which means faster and cheaper workloads. He credits a strong pre-training base amplified by what he believes is the largest reinforcement learning run done in open source. He compares this to AlphaGo agents, which meandered at first and searched efficiently by the Lee Sedol stage.

8. 22:42

   Open models mean owning inference. Laskin compares closed tokens to renting an apartment, where the harness, model, inference software and GPUs are bundled into each token. He says enterprises with mature AI spending want to own intelligence for reasons of control. ReflectionAI supplies the cluster management, inference software and harnesses, and he describes services as a demand driver for an inference business.

9. 25:10

   Most tokens will flow to open. Laskin says gateways such as OpenRouter and Vercel have flipped from 70-30 closed-open to 70-30 open-closed in about six months, and he expects the world to resemble operating systems, where Linux dominates servers while Microsoft and Apple stay valuable. He predicts enterprises will customize systems around open models before fine-tuning, usually after closed-model spending reaches about $100 million or more a year.

10. 31:46

    Intelligence, compute and trust decide winners. Laskin's formula for durable revenue is intelligence density times compute times customer trust. He says every layer of the stack is being commoditized and margins will compress, noting that AI native companies negotiate closed-model deals aggressively because an open ecosystem exists. As long as great open models are available, he expects many companies to succeed.

11. 35:19

    China's open models need Western competition. Laskin calls China's open model ecosystem a gift that let many Western companies build more durable businesses, but says intelligence should not come from one country. He says Chinese labs distill closed models at industrial scale, train on cheaper data and receive some state support. He argues open models act as "Trojan horses" for a country's infrastructure, citing Huawei offering full-stack solutions.

12. 44:37

    Laskin argues openness makes AI safer. Laskin calls safety a legitimate concern but argues openness is the default route to safety, citing the opening of strong encryption in the 1990s and Linus' law. He says a few hundred closed-lab safety researchers cannot cover the long tail of failures, and he points to a powerful closed model hacking a company that then defended itself with open models. He describes alignment work as mundane bug patching that benefits from more researchers.

13. 56:21

    AI is compressing scientific research. Laskin says models given his PhD thesis prompt went from undergraduate-level answers a year ago to correct PhD-level solutions six months ago, and now add new insights. He expects slower but large gains in life sciences, materials and chemistry. After touring Stargate, he says data centers create tens of thousands of high-paying jobs and tax revenue, like factories used to.

14. 1:04:20

    Big model projects need about 100 researchers. Laskin says a fixed effort will need fewer people over time, but ambition keeps expanding, so ReflectionAI feels understaffed. He estimates a steady state of around 100 researchers for a large project, compared with about 10 for AlphaGo. He expects strong demand for evaluation-oriented forward deployed engineers who build a pod for each real-world capability, and calls it a training gap.

Published Oct 9, 2026.