Hey everyone, Alex here 👋Welcome back to ThursdAI! After what felt like ages of non-stop, massive model drops (looking at you, O3 and GPT-4!), we finally got that "chill week" we've been dreaming of since maybe... forever? It seems the big labs are taking a breather, probably gearing up for even bigger things next week (maybe some open source 👀).But "chill" doesn't mean empty! This week was packed with fascinating developments, especially in the open source world and with long-awaited API releases. We actually had time to dive deeper into things, which was a refreshing change. We had a fantastic lineup of guests joining us too: Kwindla Kramer (@kwindla), our resident voice expert, dropped in to talk about some mind-blowing TTS and her own open-source VAD release. Maziyar Panahi (@MaziyarPanahi) gave us the inside scoop on OpenAI's recent meeting with the open source community. And Dex Horthy (@dexhorthy) from HumanLayer shared some invaluable insights on building robust AI agents that actually work in the real world. It was great having them alongside the usual ThursdAI crew: LDJ, Yam, Wolfram, and Nisten!So, instead of rushing through a million headlines, we took a more relaxed pace. We explored NVIDIA's cool new Describe Anything model, dug into Google's Quantization Aware Training for Gemma, celebrated the much-anticipated API release for OpenAI's GPT Image generation (finally!), checked out the new Grok API, got absolutely blown away by a tiny, open-source TTS model from Korea called Dia, and debated the principles of building better AI agents. Plus, a surprise drop from Send AI with a powerful video model!Let's dive in!Open Source AI Highlights: Community, Vision, and EfficiencyEven with the big players quieter on the model release front, the open source scene was buzzing. It feels like this "chill" period gave everyone a chance to focus on refining tools, releasing datasets, and engaging with the community.OpenAI Inches Closer to Open Source? Insights from the Community MeetingPerhaps the biggest non-release news of the week was OpenAI actively engaging with the open source community. Friend of the show Maziyar Panahi was actually in the room (well, the Zoom room) and joined us to share what went down It sounds like OpenAI came prepared, with Sam Altman himself spending significant time answering questions . Maziyar gave us the inside scoop, mentioning that OpenAI's looking to offload some GPU pressure by embracing open source – a win-win where they help the community, and the community helps lighten their load. He painted a picture of a company genuinely trying to listen and figure out how to best contribute. It felt less like a checkbox exercise and more like genuine engagement, which is awesome to see.What did the community ask for? Based on Maziyar's recap, there was a strong consensus on several key points:* Model Size: The sweet spot seemed to be not tiny, but not astronomically huge either. Something in the 70B-200B parameter range that could run reasonably on, say, 4 GPUs, leaving room for other models. People want power they can actually use without needing a supercomputer.* Capabilities: A strong desire for reliable structured output. Surprisingly, there was less emphasis on complex, built-in reasoning, or at least the ability to toggle reasoning off. This likely stems from practical concerns about cost and latency in production environments. The community seems to value control and efficiency for specific tasks.* Multilingual: Good support for European languages (at least 20) was a major request, reflecting the global nature of the open source community. Needs to be as good as English support.* Base Models: A huge ask was for OpenAI to release base models. The reasoning? Empower the community to handle fine-tuning for specific tasks like coding, roleplay, or supporting underrepresented languages . Let the experts in those niches build on a solid foundation.* Focus: Usefulness over chasing leaderboard glory. The community urged OpenAI to provide a solid, practical model rather than aiming for a temporary #1 spot that gets outdated in days or weeks . Stability, reliability, and long-term utility were prized over fleeting benchmark wins.* Safety: A preference for separate guardrail models (similar to LlamaGuard or GemmaGuard) rather than overly aligning the main model, which often hurts performance and flexibility . Give users the tools to implement safety layers as needed, rather than baking in limitations that might stifle creativity or utility.Perhaps most excitingly, Maziyar mentioned OpenAI seemed committed to regular open model releases, not just a one-off thin=! This, combined with recent moves like approving a community Pull Request to make their open-source Codex agent work with non-OpenAI models (as Yam Peleg excitedly pointed out!), suggests a potentially significant shift. Remember, it's been a long time since GPT-2 and Whisper were OpenAI's main open contributions! ...
