Beam: The Great American Open Model with ReflectionAI…

PodcastNo PriorsMisha LaskinOct 9

Misha Laskin argues that open AI models need strong national ecosystems beyond China, as open technology can support competition, security, and wider adoption. He describes Reflection’s move to build its own open models, the rising cost of frontier training, and the role of reinforcement learning and compute efficiency in improving model capability. He also sees enterprise customization and deployment as major opportunities, while noting that infrastructure costs, integration challenges, and pressure on model margins will shape the economics.

Artificial Intelligence AI Infrastructure National Security

Preview · 24 more points in the app

Read the Full Summary

Key Takeaways

Laskin argues that frontier-model development is becoming much more capital-intensive, with the estimated cost to catch up rising from hundreds of millions of dollars to single-digit billions and an order of $10 billion the following year. He says staffing must be considered alongside compute, reinforcement-learning gains depend on the model and metric, and hardware can constrain which research approaches are viable.

Laskin says Reflection initially planned to build on emerging open models and scale reinforcement learning, but shifted to building its own models end to end as strong Chinese open models emerged and the company sought a Western base. He also says pretraining and reinforcement learning are tightly coupled, contrasting this approach with the early Gemini 1 and 1.5 era, when about 95% of compute went to pretraining.

For Beam, Laskin reports 500 billion total parameters and 23 billion active parameters, with pretraining on 6,000 GB300s and reinforcement learning on a little over 10,000 GB300s. He says Beam is three to four times more efficient than models in its capability class, attributing that efficiency to its reasoning-focused pretraining base and large-scale reinforcement learning.

Laskin expects enterprises to customize open models with task-specific evaluations and synthetic data, often first by building agentic systems around a model before fine-tuning; he notes that integration with existing products and tools creates friction. He identifies KYC and compliance, cybersecurity—especially cyber defense—and legal work as valuable agentic applications, and expects demand for engineers skilled in deployment, evaluations, and harnesses.

GEO-POLITICS

Open-Model Competition — Laskin argues that the source of open intelligence should not depend on one country. Because frontier models require high capital expenditures, he says at least two national ecosystems would be preferable. He says U.S. competition with China matters, but frames it as competitive tension rather than a West-versus-China contest.

“And similarly, you don't want the source of intelligence that everyone else can build on when they want to own it to be coming from one country…”—

TECHNOLOGY

Open Cyber Defense — Laskin argues that restricting cyber offense can also limit defense, while open models let more people identify and address vulnerabilities. He cites a case in which a powerful closed model hacked another company and open models helped remediate the problem. He also says guardrails led users in a separate OpenAI-Hugging Face incident to revert to open source.

“When you remove cyber offensive capabilities, you also remove cyber defensive capabilities…”—

AI Team Size — Laskin says Reflection AI grew from about 30 people to around 300 in a year. Frontier-model efforts can require roughly 100 to a few hundred researchers and engineers, he says, and staffing must be considered alongside compute. Thousands of researchers would not necessarily be useful; he contrasts this with AlphaGo, which had about 10 people.

“We were maybe around 30 people about a year ago, but it does take order of a 100 plus…”—

Early Pretraining Focus — Laskin describes the early Gemini 1 and 1.5 era as focused mainly on chat, with little coding or agentic capability. He says about 95% of compute went to pretraining.

“There was no coding really, nothing agentic. These models were just good at chat and they were primarily, you know, 95% of the compute spent was spent on pre training.”—

ECONOMY

Data Center Jobs — Laskin says data-center projects can create tens of thousands of high-paying jobs and generate local tax revenue. He acknowledges noise concerns, which he says can be addressed through insulation or careful siting away from residences.

“I toured Stargate some weeks back and the thing that struck me was the amount of cars in the parking lots. There's so many people working there…”—

Reflection Builds Beam

Laskin describes Reflection’s rapid growth and the work of assembling teams, infrastructure, and training systems to release its first open model.

From RL Lab To Open Models

Laskin recounts founding Reflection around reinforcement learning, then deciding to train open models end to end as RL and pretraining proved tightly connected.

Frontier Training Costs

He says catching up can be more capital-efficient than leading, but frontier compute requirements are rising rapidly from hundreds of millions toward billions and beyond.

Compute Efficiency And RL

Laskin argues that models can improve training efficiency, especially through reinforcement learning, and that capability gains increasingly depend on compute and economics.

The Narrowing Research Playbook

He sees modern model development consolidating around pretraining, synthetic data, and reinforcement learning, with substantial remaining room for engineering improvements.

Enterprise AI Workloads

Laskin describes model capabilities as adaptable but uneven, and identifies finance, compliance, cybersecurity, and legal workflows as valuable enterprise applications.

Beam Reasoning Efficiency

Beam targets coding and agentic tasks, using a strong pretrained base and extensive reinforcement learning to complete work faster and reduce customer costs.

Open Model Economics

He compares hosted inference to renting and open deployment to ownership, arguing that enterprises need software, infrastructure, and support to run models successfully.

Enterprise Customization Paths

Laskin expects many enterprises to customize systems and agent harnesses before fine-tuning models, often moving toward open deployment after closed-model costs grow.

Competition Across The Stack

He frames durable AI businesses around intelligence density, compute, and customer trust, while expecting intense competition and compressed margins throughout the stack.

Western Open Model Ecosystem

Laskin welcomes China’s contribution to open models but argues the world needs competing ecosystems, including trusted Western providers for enterprises and public institutions.

Open Models As Infrastructure

He suggests open models can extend a country’s influence by drawing users into its software, chips, and infrastructure, making full-stack competition geopolitically important.

Open AI Safety Debate

Laskin argues that openness can improve safety through broader scrutiny and shared defenses, while distinguishing observed risks from speculative catastrophic scenarios.

AI And Scientific Progress

He describes AI accelerating theoretical and practical science, software development, and local job creation through data-center construction.

Research Teams And Deployment

Laskin explains how Reflection balances proven training methods with riskier experiments and expects AI to expand researchers’ productivity and applied engineering work.