AI Engineer Conference: Gemini 2.5 Pro, AI Thinking, and Verified Superintelligence

AI Engineer Worlds Fair Highlights

• 📌 The AI Engineer Worlds Fair in San Francisco hosted 3,000 attendees, featuring live streams and a vibrant community atmosphere.
• 💡 The conference emphasized community building with 18 tracks and a dedicated "hallway track" for informal presentations and networking.
• ✅ Notable discussions included OpenAI's enterprise features like ChatGPT for Zoom meetings and the release of Codex for Plus users, which saw over 111,000 pull requests opened in 24 hours.

Google's Gemini 2.5 Pro Innovations

• 🚀 Google DeepMind announced an updated Gemini 2.5 Pro model, hailed as the "final update" to 2.5 Pro, achieving state-of-the-art (SOTA) benchmarks on Aider and HLE.
• 📈 This progress follows Google's organizational shift, integrating research and product teams under DeepMind, leading to a 50x increase in AI inference over the past year.
• 🧠 Gemini now features a "thinking stage" where the model generates internal text for reasoning, hypothesizing, and self-correction before providing a final answer.
• ⚙️ Developers gain a "thinking budget" slider in the Gemini 2.5 Flash and Pro APIs, allowing a continuous trade-off between cost and performance.
• 🔬 The DeepThink mode for extremely hard problems uses deeper, parallel chains of thought, pushing performance to the 65th percentile on the USA Math Olympiad.

Future of Gemini and AI Development

• 💡 Gemini is evolving into a universal assistant, unifying Google's products and moving towards proactive AI systems that anticipate user needs.
• 🌐 The vision includes an omnimodal model with native audio capabilities (TTS, audio interaction) and future integration of video from models like V3.
• 🛠️ Google is developing agentic developer offerings and enhancing AI Studio into a dedicated developer platform, with upcoming features like Gemini embeddings and a deep research API.
• ⚠️ Discussions highlighted the importance of traceability and full thoughts for agentic systems, acknowledging developer demand for debugging model reasoning.

MorphLabs and Verified Superintelligence

• ✨ Jesse Han of MorphLabs unveiled a vision for "verified superintelligence", aiming to build infrastructure for the Singularity.
• 🔑 Christian Szegedy, inventor of batch norm and adversarial examples, joined MorphLabs as Chief Scientist to lead this effort.
• 🔬 MorphLabs' Infinibanch virtualization stack enables scaling test-time search and reinforcement learning R&D by checkpointing and replicating agent states in software environments.
• 🤔 The company's approach emphasizes "trust but verify" and views AI development through the lens of "parenthood," focusing on algorithmic guarantees for aligned outputs.

Industry Trends and Bottlenecks

• 💬 The concept of "Big Model Smell" describes the qualitative, often unquantifiable, capabilities of large models beyond benchmarks.
• 📊 Key bottlenecks identified for AI progress include compute (for accelerating new algorithmic ideas) and organizational coordination (for effectively leveraging resources).
• 🧩 The debate around Multi-Agent Communication Protocol (MCP) highlighted its utility for tool integration but also its hype, with a call for open standards.