Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
Misha Laskin argues that open frontier models, built with large-scale reinforcement learning, will take most token demand, and that openness improves safety, with Reflection's Beam as the first Western open model at that level.
Watch on YouTube ↗Charlie's summary
- ▶ 2:31
Building a frontier lab means getting 30 things right at once; there is no single hard part
Asked what was hardest, Misha says “everything is extremely hard.” Talent, retention, mission and culture, data, compute and the infrastructure to make compute usable all have to work together. At DeepMind he took for granted tools that just worked, and at a startup you must build them. Reflection grew from about 30 to about 300 people in a year.
- ▶ 4:05
Pre-training and reinforcement learning are tightly coupled, so Reflection decided to build open models end to end
Misha and his co-founder, both RL researchers, first planned to build RL on top of someone else's open base model. RL worked faster than expected, and the good open models were all coming from China. They also found you actually need to pre-train a model to make RL work well at scale. That pushed them to train models themselves.
- ▶ 8:44
Frontier compute costs rise about 4x per generation, but efficiency gains will blunt the climb
Misha puts catching the frontier at hundreds of millions of dollars about 18 months ago, billions recently, and on the order of tens of billions next. Each model generation needs roughly 4x the compute. He expects an asymptote because capex is limited. Human researchers gain about 7x compute efficiency a year, and he thinks models helping build themselves make that several times faster, so more intelligence comes out per training flop.
- ▶ 12:17
RL systems that never stop improving turn frontier AI into an economics decision
Misha got into AI after seeing an AlphaGo plot that never stopped improving until it was cut off. Beam's RL run used over 10,000 GB300s for four weeks, more flops than pre-training, and its curves keep rising. The question has moved from whether you can train these systems to where the economically valuable data is and how much money to put in. He says the field now feels more like engineering than the Wild West.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.