ThursdAI 100th Episode: Meta Llama 4, Google AI Updates, and ChatGPT Memory

ThursdAI 100th Episode Highlights

• 👏 The 100th episode of ThursdAI celebrates two years of documenting incredible AI progress and community engagement.
• 💡 The show covers major announcements including Meta Llama 4, Google's extensive updates, and OpenAI's new ChatGPT memory feature.

Meta Llama 4 Release Analysis

• 🚀 Meta released Llama 4 Scout (109B total, 17B active parameters, 10M context) and Llama 4 Maverick (400B total, 17B active parameters, 1M context).
• 🧠 These models feature multimodal, multilingual Mixture of Experts (MoE) architecture, trained with FP8 on trillions of tokens.
• ⚠️ The release was chaotic, with a Saturday drop, LMArena score discrepancies, and initial inference issues where local quantized versions sometimes outperformed hosted full-precision models.
• 💬 Community criticism focused on the large model sizes making local execution difficult, questioning the open-source spirit for local AI enthusiasts.

Notable Open-Source AI Models

• 🛠️ DeepCoder 14B from Agentica (Berkeley) and Together AI achieved an 8% improvement on LiveCodeBench using reinforced learning for better generalization.
• ⚡ Nvidia NeMo Guardrails Ultra, a distilled and pruned Llama 3.1 model, demonstrated superior performance over Llama 4 Behemoth (unreleased) in complex math and science questions.
• 🔬 Kim VL and VL Thinking models (3B active parameters) showed impressive results, beating larger models like Gemma 2 and Qwen 2.5 VL on math vision tests.

Google AI Announcements

• 🌐 Google Next brought a wave of updates, including the upcoming Gemini 2.5 Flash model and V2 video editing capabilities.
• 💻 Firebase Studio (rebranded Project IDX) was launched as a cloud-based agentic development environment for full-stack applications.
• ✅ Google announced official support for the Multi-Agent Communication Protocol (MCP) and introduced the new Agent-to-Agent (A2A) communication protocol.
• 📊 Deep Research with Gemini 2.5 Pro was highlighted, with internal comparisons showing user preference over OpenAI's research tools.

ChatGPT Memory & Agent Protocols

• 💡 OpenAI introduced ChatGPT Memory, allowing the model to reference all past chats for more personalized responses with no storage limit.
• 💬 Users expressed a desire for configurable options to include or exclude specific chats from memory.
• 🚀 Git MCP, a community-built tool, enables turning any GitHub repository into an MCP server for documentation, supporting LLMs.txt files and gaining significant viral traction.

Community Innovations & Observability

• 📈 Weights & Biases launched Observable Tools, proposing an RFC for OpenTelemetry trace support in MCP to enhance full-stack agent observability.
• 🤝 The initiative aims to make MCP tools transparent and observable, ensuring developers can trace every step of agent execution regardless of tool usage.
• ✨ Other innovations include Gina's rearranker M0, Deep Cogito models, Amazon's Novas speech-to-speech model, and Hydream, a new state-of-the-art image diffusion model.
• 🎯 The community's active participation in open-source development and protocol standardization is crucial for AI's future and is strongly encouraged. This includes the 1b.me/mcpspec link for community feedback. This episode also celebrated passing 1,000 YouTube followers.