gpt-oss: The Frontier of Open-Weight Reasoning Models
βOpenAI just droppedgpt-oss, their first open-weight LLM since GPT-2. In this session, we'll unpack whatβs inside the 120B and 20B variants: agent-friendly tool use and CoT strengths, long-context (128k), MoE FFNs, GQA, the SFT+RL training pipeline, and the safety work behind the release (including the new risk-assessment paper). Weβll also cover the tokenizer/prompting stack (o200k_harmony + harmony format), compare performance and βvibeβ to other leading models, and share practical takeaways for production. RSVP here.
We met the Arcee Foundation Model 4.5B β an American-made, open-weight SLM. CTO Lucas Atkins walked us through how it was made, from architecture tweaks (GQA, ReLUΒ²), to curating 8T tokens with Datology AI, to the role of mid-training to post-training with SFT plus RLVR/GRPO and KTO. Check it out to get down in the weeds.
The latest cute, furry animal (e.g., pandas, polars) to jump onto the DataFrame scene in an LLM-powered world is Fenic. We were joined by Founding Engineer Rohit Rastogi for a deep cut on the future of production data processing. A true LLM edge event.
βJoin us as we dig into AWSβs open-source Agent Squadβa multi-agent orchestration framework that routes requests via a global βClassifierβ while preserving conversation context. Weβll demo how it plugs into AWS-native tools compare it with other popular frameworks. Featuring AWSβs Milan McGraw and Aravind Subramanian, this session is ideal for engineers and AI leaders choosing their next agent stack.
β
RLVR
βJoin us for Reinforcement Learning with Verifiable Rewards [Ref]βa fast, practical primer on how verifiers drive reward training for tasks with checkable outcomes. Weβll cut through the hype after DeepSeek-R1/GRPO, highlight where RLVR excels (math, coding, instruction following) and where it struggles, and tour fresh 2025 research. Ideal for AI engineers and leaders looking to fine-tune off-the-shelf models with better cost/perf.