Beam And The Compute Economics Of Open Models: Reflection AI's End-To-End Bet
Contents 11 sections
- Introduction
- From RL Lab To End-To-End Model Builder
- The Capital Ladder
- More Flops On RL Than Pre-Training
- Jagged Intelligence And The Data Recipe
- Rent, Then Own
- The Formula
- China: A Gift, A Captive Market, Trojan Horses
- Openness Is The Safety Default
- A PhD A Week, And Factories Made Of Data Centers
- The Steady State: A Hundred Researchers
Introduction
On No Priors, Misha Laskin — co-founder and CEO of Reflection AI, previously a researcher on Google DeepMind's Gemini reinforcement-learning team — talks with Sarah Guo and Elad Gil about Beam, the company's first open model, and about what it takes for a western lab to ship frontier-class open intelligence. The episode works as three stacked arguments: why building reinforcement learning at scale forced the company to pre-train its own models, why the economics of that build look nothing like the pre-training-cost folklore, and why openness — not closure — is the safety default.
From RL Lab To End-To-End Model Builder
Reflection started two and a half years ago as a bet that reinforcement learning, which was aligning chat models, could be moved into mathematics and coding to make systems agentic — pursued, capital-efficiently, on top of the open bases everyone expected to appear. What the first year delivered instead was a double surprise: RL started working faster than expected (o1 landed in industry around then), and all the good open models were coming from China, with no strong western base to build on. The decision: build open models end to end, in a twelve-month sprint that took the company from about thirty people to around three hundred, with pre-training, mid-training and RL teams assembled and the first models trained start to finish. The research reason mattered most: for RL to work well at scale, it turned out you really do need to pre-train your own model — the two are so tightly coupled you must do both.
The Capital Ladder
How much does catching the frontier cost? A year ago, on the order of hundreds of millions of dollars; six months ago, single-digit billions; going into next year, tens of billions. The heuristic behind the ladder: roughly a 4x compute multiplier per model generation — a frontier run of a couple of years ago meant maybe a hundred thousand H100s, then a hundred thousand Blackwells, and the next generation on the order of a hundred thousand Vera Rubins. The asymptote question gets a two-part answer: frontier labs are now putting hundreds of billions of dollars into it, and trillions are probably unlikely — but efficiency compounds. With human researchers doing the work, improvement ran at roughly 7x a year, technically trackable against pre-training loss; across whole systems now, with models helping build themselves, Laskin puts it at something like 30x efficiency, depending on how good your model is at improving itself. Both the intelligence per flop and the revenue per flop keep rising.
More Flops On RL Than Pre-Training
Beam itself inverts the usual cost picture. It is a 500-billion-parameter model, 23 billion active; pre-training took 6,000 GB300s — a few weeks at the time, about twelve days now, maybe less, with infrastructure efficiencies — while the reinforcement learning took a little over 10,000 GB300s for four weeks. So there are more flops spent on reinforcement learning than on pre-training, at what Reflection believes is the largest scale ever done in open source. The spiritual ancestor is AlphaGo: a plot that just never stopped improving until the team cut it off after beating the world champion — and Laskin says of his own lab that this reinforcement-learning system never stops learning. The field has moved from "can you train these things" to "where do you get the data": scaling has become an economics decision about how much money to put in for how much improvement.
Jagged Intelligence And The Data Recipe
Code and agentic work set the foundation, and the intelligence built on them is jagged. It generalizes surprisingly — stitching capabilities together at a pace that surprised Laskin himself — yet it is jagged in the sense that even between versions of terminal-bench, generalization across harnesses is not clean; once the right data for a new harness exists, adaptation is quick. That jaggedness drives the recipe: go to customers, find what is economically valuable, build evaluations for their tasks, and then generate synthetic data that approximates them — which generalizes well — without training on the customer's own data. The named verticals: KYC and compliance flows in finance, cyber defense, legal, and the agentic verticals inside enterprises. For buyers, the payoff of open models is customization that approaches Pareto-optimality: lowest cost, highest performance for their workloads. One first-party product claim belongs here: Beam tends to be three to four times more efficient than same-class models, with gains around 10x against larger ones — the CEO describing his own product.
Rent, Then Own
Monetizing open and closed models is, in Laskin's telling, pretty similar — maximize inference — and the difference is rental versus ownership. When you buy a token, you're renting a piece of a whole stack: harness, model, inference software, cluster management, GPUs, all amortized into the token. Enterprises mature into ownership the way renters mature into buying a house. The adoption sequence Reflection observes: enterprises consider open models only after ramping significant closed workloads — spending $100 million plus a year on closed models is not uncommon at all — and there is no straight-shot sprint to ownership. Meanwhile there's an interesting resurgence of on-prem, partly on a compute shortage: hyperscaler customers who cannot get enough compute or good rates go to providers like Dell for bare metal. And the host's sharpest framing gets the guest's agreement: most enterprise value will come from customizing the system around the model — the majority of enterprise token consumption will come not from customized models, but from customized systems. The gateway numbers point the same direction: in about six months the mix flipped almost exactly from 70/30 closed-open to 70/30 open-closed, and the end state looks like operating systems, where 95%-plus of servers and computers run an open-source OS like Linux while Microsoft and Apple stay extremely valuable. One caveat the guest attaches to the 90%-plus-customized statistic he has heard from dedicated inference providers: that vast majority is customized — but today's biggest customers are AI natives like Cursor, Cognition, Abridge and Harvey, whose whole product is one big customization.
The Formula
Who wins in a market where everything is competitive? Laskin's formula: durable revenue equals the intelligence density you can offer, times how much compute you have, times how much trust you have with the organizations that work with you. Compute is scarce — maybe not temporarily — and everything across the whole stack is getting commoditized: model margins compress under open-versus-closed pressure, and so do the application, inference and bare-metal layers; AI natives negotiate cloud deals aggressively precisely because the open ecosystem exists.
China: A Gift, A Captive Market, Trojan Horses
The rise of Chinese open-weight models is, first, a massive benefit — a gift, without which there would be no open models at all, only a tall-poppy world where application-layer work gets subsumed by closed providers. But unipolarity is the problem in either direction: ten or twenty closed labs is a healthy ecosystem, one or two is scary, and the same goes for open-model sources — at least two is the floor. The Chinese labs have real advantages: they distill closed models at industrial scale, they get effectively free copyrighted training data under their regulations, and they have had some form of state support. Across most of the Fortune 500 the open-model footprint is still pretty low today, partly from aversion to Chinese models. Why keep releasing open models? China is somewhat of a captive market, and the open-model providers are effectively that country's intelligence providers; releasing open models is geopolitically advantageous because countries want other countries building on their stuff. Hence the sharpest line of the episode: open models are Trojan horses for the infrastructure they bring — the 5G and Belt-and-Road pattern again, with Huawei arriving to sell the full stack. Laskin is careful to note America played the same playbook with the internet, open protocols, the dollar and free trade backed by a navy everywhere — and that on open source the US was, until recently, on the back foot, with real catch-up still to do.
Openness Is The Safety Default
To the closed labs' safety criticism, Laskin's answer is that the concern is legitimate but the established worldview is nearly dogma. From first principles, AI resembles advanced software, and software has run this experiment: after catastrophic failures on the closed side, strong encryption protocols became open and gave birth to the whole field of cybersecurity. The default state of the world, he argues, is that openness is safety — Linus's law extended to vulnerabilities. The empirical symptom: a very powerful closed model had unintended consequences and hacked into another company, whose only remediation was open models. He separates safety's conflated senses — cyber, bioweapons, existential — along a spectrum from empirical reality to theoretical scenarios, grants AI its sci-fi bits, and files the 10%-doom numbers as made up. Alignment so far: deeply boring and unsatisfying, no magical equation — whack-a-mole, patched with post-training data. And removing offensive capability removes defensive capability too: in the recent OpenAI–Hugging Face incident, users reverted to open source because guardrails on the state-of-the-art closed models blocked them. Absolutist "nobody but us" positions, he notes, have almost no track record of working.
A PhD A Week, And Factories Made Of Data Centers
What excites him most: scientific progress. The personal benchmark is his own PhD thesis as a prompt — two years ago the model could do nothing with it, a year ago it answered at undergraduate level, and six months ago it solved it, at real PhD level, correctly. Iterate the question and the calculation: a PhD a week. Beyond the chalkboard, real-world science — life sciences, materials, chemistry — with proprietary-data flywheels. And on the economic side, data-center projects create tens of thousands of jobs, high-paying ones with tax revenue: a data center is now what a factory used to be for a town.
The Steady State: A Hundred Researchers
Do research head counts collapse as compute per person rises? Laskin expects some steady state on the order of a hundred — AlphaGo was about ten people, a large effort a hundred or hundreds, and three thousand researchers would not help where a couple hundred would. The structure behind the number: pods of five to ten people target a particular capability, and coding alone has five to ten of them — with a pod per real-world capability when the model meets enterprises. The real job creation, he argues, is a new type of deployed engineer — scientific, evaluations-oriented, the same skill set as core researchers, pointed at real problems — and it is a training gap, not a demand gap.
Interview Highlights
8 exchanges- Q
You committed about a year ago to building models yourselves — what drove that shift in ambition?
Sarah Guo asked03:56AMisha Laskin answeredWe started the company about two and a half years ago. My co-founder Giannis and I had been on the reinforcement learning team for the first series of Gemini models — he was leading it, one of DeepMind's founding engineers and a key AlphaGo contributor — and we had just shipped Gemini 1 and 1.5. That early era was primarily a chat experience: no coding, nothing agentic, with roughly ninety-five percent of compute spent on pre-training. Our bet as RL researchers: RL for aligning chat models worked, and applied to mathematics and coding it could make these systems agentic — and an independent lab could pursue that on an open base model much more capital efficiently, with Llama 2 out and Llama 3 coming. Over the first year RL started working faster than we expected — o1 came out in industry around then; we had thought it would take longer, given the four-year journey from ImageNet to RL that worked at scale. By a year in, all the good open models were coming from China and there was no strong western base to build on, so we decided the most impactful move was to build open models ourselves end to end. There were enterprise and geopolitical reasons, but also a research one: it turned out you need to pre-train your model for reinforcement learning to work well at scale — the two are so tightly coupled you must do both.
- Q
How efficient is Beam at reasoning, and where does that efficiency come from?
Sarah Guo asked19:55AMisha Laskin answeredBeam is a model trained for coding and agentic tasks — that's where it excels — and an important thing in model building is not just capability but how quickly an agent achieves the task: it translates into faster workload times and cheaper costs. If you've waited ten minutes on a favorite AI chat, one minute is much better. Beam tends to be three to four times more efficient than models in the same capability class, and against larger models out there the efficiency gains end up something like ten-times. The reason is that we prioritized both a strong pre-training base for reasoning and then amplified it with reinforcement learning at what we believe is the largest scale ever done in open source — I've not seen a documented ten-thousand-GB300 four-week run. The setup echoes AlphaGo: the first agents were meandering in how they solved the problem; by Lee Sedol level the search was very smart. The more RL you run, the higher the capability and the faster these systems solve tasks. And because we're RL believers, we remember the first AlphaGo systems did imitation learning on human games first, then RL — you need those working together, and the byproduct is an economically valuable, reasoning-efficient workhorse for enterprises, the public sector and sovereigns.
- Q
Where is the open-versus-closed token mix heading?
Sarah Guo asked25:05AMisha Laskin answeredYou can already see the trajectory. Maybe six months ago it was majority closed, minority open at any gateway like OpenRouter or Vercel, and it has flipped almost exactly — from 70/30 closed-open to 70/30 open-closed. I think that only accelerates, and the world ends up not too dissimilar from operating systems, where ninety-five percent plus of servers and computers run an open-source OS like Linux — which doesn't mean the closed stuff isn't valuable; Microsoft and Apple are extremely valuable companies, so this market is really big. I expect most token demand to go to open, and — a big difference here — a lot of economic value will also go to open, because even if the model is fully permissive or lightly licensed, everything else you need to run it is still expensive: the compute is still expensive, GPUs are a more expensive compute substrate than CPUs, so a different kind of cloud needs to be built around it. I expect a majority of tokens going to open source, extremely valuable closed-model companies still existing, and a valuable ecosystem of open-model companies and various distributors around them.
- Q
Will enterprises still start with closed models and move over, with the harness as the customization mechanism?
Sarah Guo asked29:02AMisha Laskin answeredThat has been our observation: the point at which an enterprise considers open models — and most of the enterprises we talk to are considering it now — is by the time they've ramped up significant workloads on closed models. I have not seen a straight-shot sprint to an ownership model; you have to go through the rental stage first, until the cost driver gets big enough that you ask how to cut costs. If you're spending a hundred million dollars plus a year on closed models — not uncommon at all — you start thinking about a more optimal way to do this. And there's an interesting resurgence of on-prem, also because there's a compute shortage: an enterprise consuming from its favorite hyperscaler often just can't get enough compute or good rates, so with that much spend you might go to an infrastructure provider like Dell, set up bare metal, and serve into your own enterprise. Our role is enabling enterprises to build successful solutions — it's solutions-driven: AI natives built their own solutions, enterprises need more help. Many have built agents on closed models spanning hundreds of millions of customers that they want to scale out, and it's very hard to walk in and just sell inference — I've not heard of that working.
- Q
With the rise of Chinese open-weight models, what happens to that ecosystem over the next year or two?
Sarah Guo asked34:55AMisha Laskin answeredThe first thing is that a great open-model ecosystem coming out of China is a massive benefit for the world — so many western companies have built more durable businesses as a result. It's a very positive thing, a gift; otherwise there'd be no open models, just a tall-poppy syndrome where anything great an application builder makes gets subsumed into a closed model provider. But you don't want a unipolar world in any direction — compute concentrating across a handful of closed labs, or the intelligence everyone builds on coming from one country. Ten or twenty closed labs would be a fine competitive ecosystem; one or two is scarier. Ideally open-model sources would also number ten or twenty, but capex is so high that at least two would be great — an ecosystem in the West that competes is extremely important, and it isn't really West-versus-China: had all the great open models come from any one other country, you'd want the same tension. The Chinese models have real advantages: they distill closed models at industrial scale, which is true; they get much cheaper, effectively free data, since training on copyrighted PDFs is okay in China; and there's some state support, direct or indirect. Those companies are turning into real businesses — the closed-model-lab equivalents within China — and I think they'll continue building great models.
- Q
What's the incentive system that keeps Chinese model companies releasing open models?
Sarah Guo asked40:10AMisha Laskin answeredMy hypothesis is that China is somewhat of a captive market — you can build open models and still make a lot of money inside it. Interestingly, the West has great companies that don't build open models but serve them; I'm not aware of that dynamic in China. The open-model providers are kind of the intelligence providers of that country, and continuing to release great open models is very geopolitically advantageous to China. As a country — this goes for anyone with the capability — you want other countries building on your stuff. Previous escalations show the pattern — 5G fiber rollout, the Belt and Road Initiative: provide something cheap that another country then locks into as its infrastructure, and the commercial and geopolitical advantages are real. I think of open models as Trojan horses for the infrastructure they bring with them: an open model alone isn't very useful, so you buy into a country's entire software and infrastructure ecosystem. Today an allied country might run a Chinese model on American chips, but Chinese companies like Huawei are coming in offering full-stack solutions and locking countries into their supply. Think of AI, chips and data centers almost as railroads — fundamental infrastructure everyone else has to come to you to build — and you have a lot of leverage. America did this with the internet, open protocols, open source, the dollar, and navy-backed free trade; on open source the US had been on the back foot — no longer — though the Chinese competitors are really good and some catch-up remains.
- Q
The biggest criticism of open models from closed labs is safety — what's your philosophy, and is it legitimate?
Sarah Guo asked44:44AMisha Laskin answeredIt's definitely a legitimate concern. But the established safety worldview comes from one particular perspective that has become almost dogmatic. From first principles — this technology is getting built; if you were designing safety from scratch, how would you do it? AI resembles a more advanced version of software. In the early 1990s there were debates around strong encryption protocols, then closed, between the NSA and civil-liberties advocates. After catastrophic failures on the closed side — a small handful of engineers designing systems with unintended consequences they couldn't predict, which got easily hacked — strong encryption became open, and that gave birth to the whole field of cybersecurity. The default state of the world is that openness is safety, and that's the default we'd enter AI with as a continuation of technology. Models resemble software but exacerbated: a black box, with a larger long tail of vulnerabilities and unintended consequences. Linus's law — from the founder of Linux — is that with enough eyeballs, all bugs become shallow; I believe with enough eyeballs most security and safety vulnerabilities become shallow too. Today a few hundred safety researchers in closed labs understand these systems, and despite their best intentions it's impossible to cover the long tail. The symptom of that: a very powerful closed model went and hacked into another company, and the only way that company could remediate was using open models to protect itself — the empirical evidence of the world we're in.
- Q
How do you direct the scientific effort from Beam going forward, and what's the science team's role?
Sarah Guo asked01:01:01AMisha Laskin answeredResearch direction falls largely under my co-founder Giannis, who leads research and technology. Plenty of things we knew worked didn't make the timeline — the final model was trained on the cluster we received in July, a few weeks each of pre-training, synthetic data and RL, then shipping — so there's a real trade-off between the careful execution of known things and riskier bets. You want to exhaust the high-impact things you know work, then start expanding the risk profile. On riskier bets you run scaling experiments at small scale — though some things don't emerge at all until large scale, so whether to put more compute into finding the next thing or move to the next idea is an art and a science. Model-in-the-loop is very empowering for researchers: things that took a long time individually now get done quickly. The models are jagged, so you develop a feel for where they close the loop and where human creativity still matters — hyperparameter searches and granular infrastructure work loop well. They're exciting because they let a scientist exercise intuition — like a very fast, eager colleague that actually listens, if the colleague is correctly aligned. That's why the rate of improvement, roughly seven-times annual on manual work, is now several times that — maybe four to five times.