High Agency: The Podcast for AI Builders

Merrill Lutsky, co-founder and CEO of Graphite, discusses their evolution from stack diff workflows to Diamond, an AI code review agent that just helped secure their $50M Series B. He shares insights on building reliable AI review systems, why over-generating and pruning comments works better than single responses, and the shift from RAG to agentic code browsing. Merrill offers a provocative vision where developers define requirements and AI agents build the code, potentially eliminating traditional IDE coding. This episode provides valuable perspectives on how AI is fundamentally reshaping software development workflows and engineering roles. Chapters: 00:00 - Introduction and Graphite overview 01:58 - Evolution from stack diffs to AI review 07:39 - Diamond: The AI code reviewer explained 10:13 - Human vs AI review: Finding the balance 11:44 - Engineering challenges of reliable AI review 17:38 - Over-generate and prune: A winning strategy 24:49 - From RAG to code browser agents 28:12 - The bitter lesson of AI engineering 30:48 - The future of software engineering 37:33 - Is AI over or under-hyped?

This week Raza is joined by Amit Jain, CEO and co-founder of Luma AI, to explore why the future of artificial intelligence lies beyond language. Amit shares Luma’s bold mission to build world models through multimodal training and why video is the most overlooked and critical data source in AI today. Chapters: 00:00 - Introduction 03:40 - Competing with Big AI Labs: Language vs. Multimodality 08:09 - Joint Training and Why Current Multimodal Models Fall Short 11:01 - Language is Discrete, the World is Continuous 14:36 - Do These Models Have World Models? 18:18 - Planning, Counterfactuals, and Causal Reasoning in AI 22:08 - Capabilities of Ray 2 and Real-World Use Cases 26:14 - Rethinking Video Length and Creative Workflows 29:18 - Solving Coherence Across Shots and Characters 30:00 - When Will AI Create a Feature-Length Film? 31:27 - What You Can Build with Luma’s API Today 35:49 - Overlooked Ideas and Noise in the AI Industry 38:34 - Why Video is the Missing Link in AI

Eric Simons discusses the meteoric rise of Bolt.new, an AI-powered web app builder that went from zero to $40 million ARR in just five months. He shares insights on how they built an AI agent capable of creating full-stack web applications from simple prompts, the challenges of rapid growth, and the future of AI in software development. From nearly shutting down the company to becoming one of the fastest-growing AI products in history, Eric offers valuable lessons for anyone building in the AI space. Chapters: 00:00 - Introduction and Bolt.new overview 06:05 - The journey from near-shutdown to rapid growth 13:28 - Challenges of explosive growth and scaling 18:50 - Technical deep dive: Building Bolt.new 26:37 - Debugging and improving AI-generated code 32:09 - Future directions and enterprise adoption 34:11 - Advice for building AI applications 37:03 - The concept of "vibe revenue" in AI startups 39:39 - Is AI over or under-hyped? ------------------------------------------------------------------------------------------------------------------------------------------------ Humanloop is the LLM evals platform for enterprises. We give you the tools that top teams use to ship and scale AI with confidence. To find out more go to humanloop.com

In this episode of High Agency, Patrick Leung from Faro Health explains how they're using AI to revolutionize clinical trial design by both generating regulatory documents and extracting insights from thousands of existing trials. Patrick emphasises the essential collaboration between clinical experts and AI engineers when building reliable systems in healthcare's high-stakes environment.  Chapters: 00:00 - Introduction 04:26 - Clinical trials before: Microsoft Word Documents 08:17 - Document generation using AI 12:26 - What makes clinical trials so expensive 16:26 - Parsing and processing clinical trial data 18:04 - Challenges with traditional evaluation metrics 21:28 - Importance of domain experts in the evaluation process 24:35 - Collaboration between domain experts and engineering 31:26 - Building a graph-based knowledge system 34:27 - Roles and skillsets required 38:06 - Lessons learned building LLM products 40:56 - Discussion on AI capabilities and limitations 46:07 - Is AI overhyped or underhyped ------------------------------------------------------------------------------------------------------------------------------------------------ Humanloop is the LLM evals platform for enterprises. We give you the tools that top teams use to ship and scale AI with confidence. To find out more go to humanloop.com

In this episode, Raza is joined by Shahriar Tajbakhsh, the co-founder of Metaview. They discuss how Metaview’s AI scribe automates interview note-taking, how AI agents can surface top candidates from thousands of resumes, and why hiring managers should think of AI as a co-worker, not just a tool. Raza's recomended reading: Creating a LLM-as-a-Judge That Drives Business Results [https://hamel.dev/blog/posts/llm-judge/]. Chapters: 00:00 - Introduction 03:32 - How AI Co-Workers Are Transforming Recruiting 06:21 - Inside MetaView: AI Scribe and Workflow Automation 09:11 - Unlocking Hiring Insights with AI-Driven Conversations 11:30 - Balancing AI Innovation and User Adoption 14:05 - Metaview’s Tech Stack and the Role of LLMs 18:29 - How MetaView Generates Superhuman Interview Notes 23:18 - The Challenges of Building Reliable AI Hiring Agents 32:40 - The Future of AI in Hiring: Automating Job Descriptions 40:26 - AI Co-Workers That Work While You Sleep 47:08 - Why Vertical AI Will Win Over General AI Agents 50:24 - The Underrated Power of Graph-Based AI ------------------------------------------------------------------------------------------------------------------------------------------------ Humanloop is the LLM evals platform for enterprises. We give you the tools that top teams use to ship and scale AI with confidence. To find out more go to humanloop.com