gpt-oss

OpenAI just dropped gpt-oss, their first open-weight language models since GPT-2 — available in massive 120B and 20B sizes — and they’re already being called the frontier of open-weight reasoning models. Trained with a mix of reinforcement learning and techniques from OpenAI’s top internal models, gpt-oss excels at tool use, few-shot function calling, chain-of-thought reasoning, and more, all while matching proprietary models on safety benchmarks. In this live session, we’ll break down everything: training data (STEM, coding, general knowledge), architecture (MoE FFNs, RoPE, 128k context, alternating dense/sparse attention), post-training alignment methods, and their bold safety testing on biology and cybersecurity fine-tunes. We’ll also explore the open-source goodies — tokenizer, harmony prompt format, and CoT monitoring invites — and put the model to the test to see how it stacks up against today’s best. Don’t miss this deep dive into OpenAI’s biggest OSS drop in years. Join us every Wednesday at 1pm EST for our live events. SUBSCRIBE and RING THE BELL to get notified! Join our community to start building, shipping, and sharing with us today! https://discord.gg/RzhvYvAwzA Apply for The AI Engineering Bootcamp today! https://bit.ly/AIEbootcamp For team leaders https://bit.ly/41YtlA7 LLM Foundations, a free 5-day email-based course https://bit.ly/4izQVKk LLM Engineering - Open-Source! https://bit.ly/42ha7Xy Speakers: Dr. Greg, Co-Founder & CEO @ AI Makerspace https://www.linkedin.com/in/gregloughane The Wiz, Co-Founder & CTO @ AI Makerspace https://www.linkedin.com/in/csalexiuk/ 00:00 Intro: GPT OSS vs GPT-5 and Why Open Source 07:00 What’s New & What’s Not: Model Features & Research 13:30 Release Details: Safety, Model Card, and Benchmarks 21:00 The Harmony Response Format: Next-Gen Prompting 27:00 Demo Deep Dive: How Harmony Format Works 37:00 Best Practices, Tool Use & Real-World Application 43:00 Safety, Alignment, and the Agentic Core 54:00 Industry Impact & Closing Thoughts