As of: 2026-10-10 23:43 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/v/no-priors-beam-reflection-misha-laskin/

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

No Priors: AI, Machine Learning, Tech, & Startups, 9 October 2026, YouTube. 4,555 views as of 10 October 2026.
Kind: Interview.
Watch: https://www.youtube.com/watch?v=up4sG9RM20M

## Why it is here

No Priors (Sarah Guo, Elad Gil) interview with Misha Laskin on Beam's pre-training and RL.

## Description (written by Gemini from the video)

**Summary**  
In this episode of *No Priors*, hosts Elad Gil and Sarah Guo interview Misha Laskin, co-founder and CEO of Reflection AI. Laskin discusses the release of Beam—Reflection AI’s first flagship open-weight model—as well as the economics of frontier model training, reinforcement learning scaling, and the geopolitical and safety dynamics between open and closed AI models.

**What is shown**  
- [00:00] Intro clip featuring headlines about the July 2026 OpenAI–Hugging Face security breach and Laskin discussing open models versus closed-lab safety vulnerabilities.
- [00:42] Elad Gil introduces Misha Laskin and outlines his background at Google DeepMind.
- [01:12] Laskin discusses scaling Reflection AI from ~30 people a year prior to approximately 300 employees and releasing Beam.
- [22:27] Detailed breakdown of Beam's architectural specifications and training hardware cluster.
- [32:00] Discussion on the commercialization of open models, comparing "rental" inference APIs to "ownership" deployments for enterprise infrastructure.
- [44:50] Discussion of AI safety philosophies, comparing Linus's Law ("given enough eyeballs, all bugs are shallow") to closed-lab containment failures.
- [56:28] Laskin demonstrates how models evolved from failing his PhD thesis questions to independently solving them and suggesting novel scientific angles.

**Claims & numbers**  
- **Team and Model Growth:** Reflection AI scaled from ~30 people a year ago to around 300 researchers and engineers [01:38, 01:51].
- **Beam Model Specifications:** Beam is a 500-billion-parameter total Mixture-of-Experts (MoE) model with 23 billion active parameters [22:28].
- **Pre-training Hardware:** Beam's pre-training utilized 6,000 GB300 GPUs run over several weeks (re-runnable in ~12 days with infrastructure efficiency gains) [22:33, 22:42].
- **Reinforcement Learning Cluster:** Beam's post-training/RL utilized slightly over 10,000 GB300 GPUs running for 4 weeks [22:52].
- **Compute Efficiency Gains:** Laskin claims compute efficiency gains have been roughly 7x per year historically, scaling to roughly 30x or more when factoring in agentic RL and recursive self-improvement pipelines [11:06, 11:42].
- **Hardware Generations:** Laskin cites prior frontier training using 100,000 H100s, followed by runs on 100,000 B200s (Blackwell), moving next to 100,000 Vera Rubin clusters [09:50, 10:02].
- **Efficiency Comparison:** Laskin states Beam is 3x to 4x more reasoning-efficient than models in its capability class, reaching up to 10x against larger dense models [20:30, 20:46].
- **API Token Share:** Laskin notes market token volume across routing gateways shifted from roughly 70/30 closed-to-open six months prior to around 70/30 open-to-closed [25:20, 25:31].

**Notable quotes**  
- [00:30] "I have the belief that with enough eyeballs, most security and safety vulnerabilities become shallow as well." — Misha Laskin  
- [22:27] "To train Beam, which is a 500-billion-parameter model total, 23B active, it was 6,000 GB300s." — Misha Laskin  
- [55:16] "When you remove cyber offensive capabilities, you also remove cyber defensive capabilities." — Misha Laskin

**Assessment**  
This video is a podcast interview and technical discussion between venture capitalists and a frontier AI startup founder. No live interactive terminal benchmarks or software UI demos are run on screen, but the technical specs, cluster configurations, and training methodologies behind the Beam release are detailed directly by the company's CEO.

_Described by gemini-3.8-flash on 2026-10-10 from the video's audio and frames._

## Related

- 2026-10-05: [Reflection AI unveils Beam, a 501B-parameter Apache-2.0 open-weight MoE](https://postcutoff.com/e/2026-10-05-reflection-beam-501b-open-weight/)

## People in it

- [Misha Laskin](https://postcutoff.com/person/misha-laskin/), Co-founder and CEO, Reflection AI