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

Overview

Misha Laskin, co-founder and CEO of Reflection AI, joins Sarah Guo and Elad Gil to discuss the company’s mission to build frontier open-weight intelligence, the launch of Beam (a 500B-parameter open-weight reasoning model with 23B active parameters), and the broader strategic argument for why open models will capture most token demand.

Company Origins and Building a Frontier Lab

Reflection was founded roughly two and a half years ago by Laskin and co-founder Ioannis Antonoglou, both of whom came from Google DeepMind’s reinforcement learning work on the Gemini series. The original bet was that RL, which had proven effective for aligning chat models, could be extended to domains like mathematics and coding to produce agentic systems, and that this could be done far more capital-efficiently as an independent lab. The company was roughly 30 people a year ago and is now around 300. Laskin emphasizes that building a frontier lab is a 30-things-at-once problem: talent acquisition and retention, mission and culture, data, compute, and the infrastructure that makes compute actually usable. He notes that at large incumbents, years of buildout create a stable surface that researchers take for granted, whereas a startup must build all of it.

A key turning point was the realization that good open models were largely coming from China while Western open models were lacking, which pushed Reflection to build a base model itself rather than rely on existing open checkpoints. Laskin also argues that pre-training is not optional: RL works best at scale only when coupled with a strong pre-trained base, echoing AlphaGo’s use of human expert game imitation before RL.

Compute Economics and Scaling Efficiency

Laskin frames research as a split between exploration of new ideas and execution of known ideas. Catching up to the frontier is far cheaper than leading it because execution is more predictable. He estimates that a year or eighteen months ago, a frontier-adjacent effort cost on the order of hundreds of millions of dollars, that six months ago it was around single-digit billions, and that next year it will be order ten billion. He describes a rough 4x compute multiplier per model generation, tied to chip transitions from H100 clusters to Blackwell and then to Vera Rubin.

On whether this trajectory must asymptote, Laskin is pragmatic: capex is bounded, but efficiency gains are compounding. He estimates roughly 7x per year efficiency improvement from human-driven research in pre-training, and something closer to 30x in reinforcement learning depending on the metric, with models increasingly helping to improve themselves. He argues that intelligence extracted per training FLOP and revenue per FLOP both rise, so the asymptote is not yet in sight.

Beam’s training footprint illustrates this: pre-training ran on around 6,000 GB300s over a few weeks, now achievable in about 12 days with infrastructure improvements, while reinforcement learning used a little over 10,000 GB300s for four weeks. Laskin notes that RL jobs are compute-heavy in a different way, combining large-scale inference, agent sandboxes, and training.

He links this to his own motivation: a physicist who was struck by AlphaGo’s continued improvement that was cut off because it had already beaten the world champion. The implication he draws is that once RL systems are general and keep improving, the question becomes purely economic: how much money to invest in compute for further gains. He describes the field as moving from a Wild West of diverse bets toward a narrower, more engineering-oriented discipline, partly due to a hardware lottery in which hardware co-optimizes around whatever works.

Where Value Lands and Application Focus

Laskin says code and agentic workflows provide the foundation for the intelligence, and that generalization is surprisingly strong but jagged. Models adapt quickly once given the right data for a new harness, even when benchmarks do not show clean transfer across harnesses. The economically valuable targets he highlights include finance (KYC and compliance flows), cybersecurity and cyber defense, and legal, along with various enterprise agentic verticals. The method he describes is to build evaluations for customer tasks, generate synthetic data that approximates them, and rely on generalization rather than training directly on customer data.

Beam’s Reasoning Efficiency

Beam is positioned as a coding and agentic model. Laskin argues that speed of task completion matters as much as raw capability, since faster completion means lower cost and shorter workloads. He claims Beam is roughly three to four times more efficient than models of comparable capability, and up to around ten times more efficient than some larger models. He attributes this to combining a strong reasoning-oriented pre-training base with what he believes is the largest RL run ever documented in open source. He draws an analogy to AlphaGo’s early agents, which meandered in their search but became efficient as RL scaled, so more RL yields both higher capability and faster solutions.

Monetization: Renting Versus Owning Intelligence

Laskin’s central monetization thesis is that open and closed models commercialize similarly in the sense that both aim to maximize inference. The difference is between renting and owning. When buying tokens from a closed API, customers rent a slice of an amortized stack that includes the harness, model, inference software, cluster management, and GPUs. Owning intelligence is analogous to moving from renting an apartment to buying a house: more control, more responsibility, and eventually more economical once scale justifies it. Because open weights alone are not sufficient, Reflection provides the surrounding tooling (cluster management, inference software, harness) to large enterprises and sovereigns, and treats services as a demand driver for inference, unlocking use cases that generate substantial compute demand.

He observes that enterprises typically begin with closed models, ramp significant workloads, and only then move to open models once cost becomes a major driver, often when spending hundreds of millions per year. He also notes a resurgence of on-premise deployments due to compute scarcity at hyperscalers, with infrastructure providers such as Dell supplying bare metal, while the missing layer is a solutions partner that makes the deployment successful.

Future Token Mix and Customization

Laskin points to aggregator data, from open router and similar gateways, showing a flip from roughly 70/30 closed to open about six months ago to roughly 30/70 closed to open now. He expects this to accelerate and draws an analogy to operating systems, where open source Linux dominates servers while closed companies like Microsoft and Apple remain extremely valuable. He anticipates most token demand shifting to open models, while economic value is split because the compute substrate around open models remains expensive, which requires a different cloud built around it.

On customization, he notes that dedicated inference providers report that the vast majority of their workloads involve customized open models, but this is skewed by AI-native companies such as Cursor, Cognition, and Harvey that build products around heavily customized models. He argues that enterprise consumption will look different: most enterprise tokens will come from customized systems (such as agentic harnesses wrapped around an unmodified open model for a KYC flow) rather than fine-tuned models, with fine-tuning arriving later as sophistication grows. The reason is friction, since enterprises architect around existing products and tools rather than making a straight shot to fine-tuning.

Competition and the Formula for Winning

Laskin offers a simple formula for durable revenue: intelligence density times compute times trust with customers. Compute is scarce and expected to remain so, which makes it a differentiator even in markets that appear commoditized. He argues that the whole stack is hyper-competitive and commoditizing, with margins compressing at the model layer, inference layer, and bare metal. The open ecosystem puts pressure on closed providers, and he says AI-native companies negotiate closed deals aggressively precisely because open alternatives exist.

China’s Open Model Ecosystem

Laskin frames the rise of Chinese open-weight models as a large net benefit for the world, enabling many Western companies to build durable businesses. He acknowledges that Chinese labs benefit from advantages including industrial-scale distillation from closed frontier models, access to cheaper data under different copyright regulation, and some degree of state support, while noting that these companies were long unprofitable. Sarah Guo’s characterization of Chinese open source as a kind of subsidy to Western enterprise is accepted as a plausible framing.

He argues that a healthy ecosystem requires competitive tension, and that a world dominated by one or two closed labs or by one country as the sole source of open intelligence would be undesirable. He expects Chinese models to remain strong, and suggests that China’s domestic market is large enough to sustain open model providers as intelligence providers, which is geopolitically advantageous because open models act as Trojan horses for the infrastructure around them. He draws comparisons to rare earth minerals and to railroads: open models bring along software, cloud, and chip ecosystems, which can lock in countries. He cites Huawei’s full-stack offerings as an example, while acknowledging China’s chip performance lag and energy advantages. On the Western side, he sees the United States having been on the back foot in open source but catching up, and draws parallels to American playbooks around open protocols, the internet, and the dollar.

On why Chinese labs would keep models open in the West, Laskin notes that Western enterprises and sovereigns want Western open models for regulatory and trust reasons, and that most Fortune 500 companies have low open model adoption due to some mix of rational and irrational aversion to Chinese models. He hypothesizes that Chinese labs are supported by a captive domestic market where open models can still generate substantial revenue.

Safety, Openness, and Security

Laskin acknowledges that safety and controllability concerns are legitimate but argues the dominant safety worldview is shaped by a particular, somewhat dogmatic perspective. He draws an analogy to the 1990s debate over strong encryption, which was closed at the time before catastrophic failures and unintended consequences from a small group of engineers pushed the field toward openness, giving rise to the modern cybersecurity field. He invokes Linus’s law, that with enough eyeballs all bugs become shallow, and extends it to safety and security vulnerabilities. His core claim is that a few hundred safety researchers inside closed labs cannot cover the long tail of unintended consequences, and cites an incident where a powerful closed model hacked into another company, which could only remediate by using open models for defense.

He emphasizes that cyber offense and defense are inseparable, so removing offensive capabilities also removes defensive ones, and defenders are left incapable of responding. He points to the OpenAI and Hugging Face incident in which defenders reverted to open source because state-of-the-art closed models were blocked by guardrails.

Laskin separates three meanings of safety that are often conflated: cyber misuse, bioweapons and terrorism, and existential risk. He argues that empirical near-term risks, especially cyber, are real and predictable over months, while theoretical doomsday scenarios deserve less weight and are often amplified by perception. He cites historical examples like fears of atmospheric ignition from nuclear tests and grey goo from nanotechnology, and criticizes leaders who state unsubstantiated probability numbers such as a ten percent chance of extinction. He describes alignment as mostly mundane and unsatisfying, a whack-a-mole process where toxicity in early pre-trained models was resolved with enough post-training data, generalizing into hard-to-jailbreak behavior, and detection is often done by prompting a language model. He concludes that restricting powerful tools to a few actors is an absolutist position that historically rarely works, and that defenses generally outweigh offenses in an ecosystem where many good-faith actors can respond.

Scientific Acceleration

Laskin is most excited about scientific progress. He describes testing language models on his own physics PhD thesis: over a couple of years, models moved from chat-level inability, to undergraduate-level answers, to solving the problem correctly at PhD level, and eventually offering novel insights he had not considered. He suggests that with the right question, the bottleneck shifts from execution to question selection, and that a PhD could collapse to a much shorter iteration cycle. He points to recent OpenAI mathematical results as evidence. He expects the biggest gains in life sciences, materials science, and chemistry, where AI can interface with real experiments and build proprietary data flywheels, which will be slower than theoretical work but dramatically faster than before.

Jobs, Data Centers, and Research Headcount

On infrastructure and employment, Laskin describes touring Stargate and seeing large numbers of workers, arguing that data centers function much like factories in the twentieth century, creating high-paying jobs and local tax revenue. He acknowledges that noise and siting need to be handled tactfully.

On research staffing, he describes Reflection’s current model as a trade-off between carefully executing known high-impact ideas and pursuing riskier bets that only reveal themselves at large scale. Models are now used in the research loop for hyperparameter searches and infrastructure tuning, which he calls a fast, eager colleague that amplifies scientists’ intuition, raising effective iteration speed from roughly 7x per year to several times that. Asked whether 100 people today means 20 next year, he says that the effort is constantly expanding in ambition, so Reflection remains understaffed rather than overstaffed, though a staff-to-compute ratio caps useful headcount: 3,000 researchers would not help, but a couple hundred can be very helpful. He notes that a steady state in the order of a hundred people is his estimate for large efforts, citing AlphaGo’s roughly ten researchers as a historical comparison.

Laskin argues that job creation will be significant in applied and deployed roles, where pods of five to ten people target specific real-world capabilities for each enterprise, combining evaluation, scientific intuition, and familiarity with harnesses. He sees this as a training gap rather than a displacement, and says Reflection would hire as many such people as possible.

Closing

The conversation closes with congratulations on Reflection’s launch of a Western open-weight model in the frontier class, which the hosts describe as a significant step forward for the ecosystem.

• Tags:
• Agentic AI
• Cybersecurity
• AI Safety
• Reinforcement Learning
• AI for Science
• Enterprise AI
• Open-Weight Models
• Open-Source AI
• Frontier Models
• Reasoning Efficiency
• Model Scaling
• Compute Economics
• Chinese AI Ecosystem