Reflection AI Ships First Model Beam; Employee #1 Recalls a Year Building From 1k H100s to GB300 Clusters
bigblueboo · x · 2026-11-09
Reflection AI's first employee (also its first Chinese employee) marked his one-year anniversary by announcing the company's first model, Beam, since pivoting to open-source frontier models.
He recapped his year on the infra team:
- Moved from RL infra to pre-training infra, building almost everything from scratch.
- Early on, even a 1k-GPU H100 run needed constant babysitting and still failed regularly; they then built the Kubernetes team and cluster foundation, engineered fault tolerance and observability from scratch, turning fragile runs into self-healing systems.
- Onboarded GB300 at scale and relentlessly improved MFU and goodput.
- Sat in on the transition from dense models to large-scale MoE, seeing firsthand how science scales from small experiments to large training runs.
More from Companies & People
- TAMU Assistant Prof. Kuan-Hao Huang shares PhD survival and application blog series — cneuralnetwork · 2026-11-09
- Ruth AI and Google host invite-only AI for Social Good women's hackathon in LA — jocarrasqueira · 2026-11-09
- Madison Huang's planned Nvidia visit to South Korea canceled over family issues — ShakeelHashim · 2026-11-09
- Jensen Huang accused of lying to Bloomberg reporters, 'everything is securities fraud' — ShakeelHashim · 2026-11-09
- Polymarket opens book on Jensen Huang exiting Nvidia CEO role by year-end at 10% odds — Polymarket · 2026-11-09
- Early Anthropic employee memoir: equity pledges, FBI briefings, traumatized bureaucrats — LessWrong 精选 · 2026-11-09