/ About this event
Intelligence in the Open: A MiniMax Research Night
For our SF Tech Week edition, we’re bringing AI researchers, engineers, and founders together to explore a question: What happens when AI systems begin contributing to their own improvement?
Hosted at UpScaleX, the evening will focus on recursive self-improvement and the research leading toward it: autonomous experimentation, post-training, synthetic data, verifiers, and self-improving agents. We’ll also discuss MiniMax’s PostTrainBench, where M3 ran cycles of data synthesis, training, evaluation, and iteration on pretrained models.
Expect short research talks, an open discussion, food, and time to connect with speakers and other builders.
Speakers: More To be announced.
Yi Su — Member of Technical Staff, Research, Fireworks AI
Yi works on specialized intelligence, model routing, and orchestration. He’ll discuss Fireworks’ Gen-1, trained on top of MiniMax-M3, and how the team used autonomous agents to design rewards and training recipes.
Khannan S. — Principal ML Engineer, DigitalOcean
Personalized Intelligence & why Evals are all you need Walk through how DigitalOcean enabled customers to set up their own continuous learning flywheel: structuring Evals as goals for Agentic optimization, how automated Evals can recursively help in the discovery of new Evals and finally how DO's managed infrastructure makes personalized intelligence more accessible.
Shi Dong — Core Contributor, Miles · RadixArk / SGLang
Title: Miles: RL at Scale
Abstract: RL post-training has demonstrated remarkable successes. However, scaling RL remains challenging due to limited observability, long-term instability, complex coordination, and susceptibility to bugs. We present Miles, an RL framework designed to scale post-training to thousands of GPUs and trillion-parameter models while preserving robustness, reliability, and cost efficiency.
Tyler Kenney — AI Performance Engineer, Modular
Self-tuning ML Inference Systems
LLM serving configurations interact, and the best settings change as traffic shifts. Tyler will share how Modular builds self-tuning inference systems that adjust performance parameters based on live workload behavior.
Yujie Zheng (Jade) — DevRel Engineer, EverMind
Talk: The Harness of Harnesses for Composable Agentic Intelligence
Jade will introduce Raven, EverMind’s open-source multi-agent system that builds and improves specialized agent workflows, coordinates agents across complex tasks, and turns past experience into reusable skills.
Schedule
5:30–6:00 PM — Arrival and check-in
6:00–6:10 PM — Welcome
6:10–7:10 PM — Research talks
7:10–7:30 PM — Discussion and Q&A
7:30–8:30 PM — Food and networking
MiniMax is a global AI company developing multimodal foundation models and AI-native products. Founded in 2022, our models—including MiniMax M3, H3, Speech 2.6 and Music 3.0—support advanced coding, agentic performance and long-context processing across text, audio, images, video and music. They power MiniMax Code, MiniMax Hub, MiniMax Audio, Talkie and our Open API Platform, serving 300M+ users in 200+ countries and 1M+ enterprises and developers in 100+ countries.
UpScaleX is a pre-seed to seed fund co-building with technically strong founders across the AI stack, from AI infrastructure to applications. We support founders from day one with capital, hands-on product and GTM work, and a network of 10,000+ innovators.
EverOS is plug-and-play memory infrastructure that gives AI agents persistent, portable, low-latency memory and turns every run into experience, so they keep improving on their own.
Presented by MiniMax and hosted at UpScaleX, 465 California St, 7th floor, San Francisco. This event is part of #SFTechWeek.
