Transformer Metamorphosis: From Caterpillar to Butterfly @LatentSpacePod

Uploaded October 2025 | Updated September 2026, 2 weeks ago

#aiengineering #JacobBuckman

Transformer Metamorphosis: From Caterpillar to Butterfly

The Blueprint for Autonomous Work Agents | Gavriel Cohen, NanoClaw

AI Security After Codex and Claude Code — Zico Kolter & Matt Fredrikson, Gray Swan

Inside xAI: Building Grok Imagine in 3 Months, Videogen vs World Models, and Video Agents— Ethan He

Why Your AI Agents Don’t Work with Dex Horthy of HumanLayer | In-Context Cooking

The AI Memory Problem: Why Long Context Isn’t Enough — Dan Biderman, Engram Co-founder & CEO

The Tension Between AI and Biology Thats Actually Driving Breakthroughs

$Building Jamba 3B: the tiny Hybrid Transformer State Space Reasoning Model - Barak Lenz, CTO of AI21 Join us for an exclusive conversation with *Barak Lenz,* CTO of AI21, as he unveils their groundbreaking *Jamba 3B* model - a tiny yet powerful hybrid transformer-state space model designed to bring long context capabilities to edge devices. Barak shares the fascinating journey of how AI21 became pioneers in hybrid architectures, combining attention mechanisms with Mambas state space models to achieve unprecedented efficiency without sacrificing performance. In this deep technical discussion, Barak reveals how the 1:8 ratio of attention to Mamba layers emerged from extensive ablations, why hybrid models are essential for the future of long-context AI, and how Jamba 3B can fit the same context length as much larger models while using a fraction of the memory. He explains why images quickly become long context problems (with just 4 images requiring thousands of tokens) and how this makes hybrid architectures crucial for on-device AI applications. The conversation explores AI21s broader vision for *AI systems over standalone models,* with Barak making a compelling case for why enterprises need model-agnostic orchestration layers that can balance cost, latency, and accuracy. He introduces *Maestro,* AI21s enterprise AI system that treats models as actions with statistical properties rather than monolithic solutions, enabling continuous learning and adaptation without being locked into any single model provider. Drawing from his unique background in algorithmic trading, Barak shares insights on developing frontier models with thousands of GPUs, the importance of world-class engineering in AI research, and why the industry needs to move beyond brute force approaches to reasoning. He candidly discusses the challenges of training at scale, the persistence of optimization issues even with modern architectures, and why good engineering is just as crucial as algorithmic innovation.$