Anthropic Code Leak: A Rare Look Inside Frontier AI | EP.52

AI Is Designing the Next Cancer Fighter | EP.53

What if AI could design proteins to help your immune system find and kill cancer cells? That's not a hypothetical — it's what 28 teams across 40 countries attempted in the Bits-to-Binders Challenge, an open-science competition organized by PhD students at the University of Texas at Austin. In this episode, Ron sits down with three of the organizers — Clay Kosonocky, Daryl Barth, and Aaron Feller — to unpack how they pulled off one of the most ambitious student-led experiments at the intersection of AI and biology. Together, they submitted 12,000 AI-designed protein sequences to bind to a cancer target called CD20, then validated the results in real biological assays. The conversation covers the 100-year history of protein folding, how AlphaFold changed everything, why AI biology can't just rely on benchmarks, what a CAR-T cell actually does, and what a 7% hit rate tells us about where the field really stands. Plus: open source science, the verification gap between digital predictions and wet lab reality, and why a global team of strangers working together might be the most hopeful signal of all. 00:00 Intro & Why AI Protein Design Matters 02:38 Why Protein Folding Is So Important 04:47 What AlphaFold Changed 07:59 From Predicting Proteins to Designing Them 11:10 The Rise of AI Protein Design 13:27 AI Skepticism in Biology 15:34 Why Wet Lab Validation Still Matters 20:36 Inside the Bits-to-Binders Challenge 22:05 Designing CAR-T Cell Proteins 26:57 Why Most Designs Failed 31:12 Open Source Biology & Global Collaboration 33:37 Competition Winners & Best Results 35:39 Final Takeaways on AI + Biology

Anthropic Code Leak: A Rare Look Inside Frontier AI | EP.52

What can we actually learn from the recent Anthropic code leak? In this episode of Hidden Layers, Ron Green, Michael Wharton, and Dr. ZZ Si unpack what the leak reveals about how a frontier AI company may be building agentic systems in practice. They explore Anthropic’s apparent approach to memory, skills, and context compaction, and why the biggest takeaway is not model weights, but the harness around the model. The conversation also gets into why simple, human-readable systems may be outperforming more complex architectures, and what these design choices could mean for the next generation of domain-specific AI agents. 00:00 Intro and why the leak matters 00:43 What leaked and what it reveals 03:50 Memory systems and context management 07:20 Skills, extensibility, and simple design 11:39 Compaction and the limits of context windows 17:23 Why the harness matters so much 18:36 A blueprint for building agentic systems

The "AI Bubble" Bubble | EP.51

Is the AI bubble narrative itself a bubble? Billions of dollars are flowing into chips, data centers, and frontier models. From the outside, it can look speculative. But from inside the industry, the signal looks very different. In this episode of Hidden Layers, Ron Green is joined by Michael Wharton and Dr. ZZ Si to discuss what it actually feels like to build with AI today. They explore rapid advances in model capabilities, the growing power of coding agents, and why many organizations are still struggling to absorb the productivity gains AI already enables. They also examine the massive capital investment in AI infrastructure and debate what signals would actually indicate the industry has hit a plateau. 00:00 – Is the AI Bubble Narrative Itself a Bubble? 03:00 – Rapid Advances in AI Model Capabilities 05:35 – Coding Agents and the Changing Development Workflow 09:30 – Benchmarks Showing AI Capability Acceleration 16:20 – Verifying AI Outputs and the Limits of Evaluation 18:20 – CAPEX, Chips, and the Dot-Com Bubble Comparison 21:50 – What Would Actually Signal an AI Bubble 26:30 – Why AI May Become a Utility

Did AI Kill Programming? | EP. 50

Are AI coding tools actually replacing programmers, or just changing how software gets built? In this episode of Hidden Layers, Ron Green sits down with Dr. ZZ Si and Michael Wharton to unpack what has shifted with modern coding agents, what has not, and where the hype breaks down. They share concrete examples from their own workflows, including how coding tools have moved from autocomplete to handling larger chunks of work, and why the real bottleneck is no longer writing syntax, but defining intent, architecture, and product direction. The conversation also explores how these tools are reshaping team velocity, why senior engineers tend to get more leverage from AI than junior developers, and the risks of weakening the talent pipeline if companies stop investing in early-career engineers. The episode closes with a candid look at what skills will matter most in an AI-assisted world, how abstraction layers are changing the role of programmers, and whether we may already be near peak computer science graduates. 00:00 – The rise of AI coding tools 03:07 – How workflows are changing 06:27 – Team velocity and delivery speed 08:19 – Product thinking vs. engineering execution 09:46 – Is programming actually dying? 11:41 – What “programming” means now 15:23 – Senior vs. junior developer leverage 16:33 – The developer talent pipeline 18:21 – Ego, identity, and automation 19:08 – Before vs. after: building with AI 22:30 – Debugging and fixing issues with AI 24:42 – Spec-writing and product shaping with AI 26:49 – The future of computer science grads 29:20 – Closing reflections

Your AI Is Too Big, Too Expensive, and Probably Wrong | EP. 49

What if the most powerful AI in your organization isn’t the biggest model you can buy, but the one trained on data only you own? In this episode of Hidden Layers, Ron Green is joined by Dr. ZZ Si and Michael Wharton to break down why domain-specific AI models consistently outperform general-purpose systems in real enterprise environments. They explore how narrowly scoped models deliver higher accuracy, lower costs, better reliability, and stronger governance, especially when built on proprietary data. Through real-world examples spanning finance, industrial systems, healthcare, and document understanding, the conversation tackles when to build custom models, when to rely on APIs, and how to identify AI initiatives that actually make it into production. The takeaway is clear: focus beats scale, and specificity is often the fastest path to durable competitive advantage. CHAPTERS * 00:00:00 What Is Domain-Specific AI * 00:01:15 General Models vs. Focused Systems * 00:02:48 Performance, Cost, and Model Size * 00:04:13 Proprietary Data as Advantage * 00:07:58 Why AI Fails in Production * 00:08:42 Real-World Domain-Specific Examples * 00:10:54 How to Decide What to Build * 00:14:53 Scale, Accuracy, and Uncertainty * 00:18:49 The Spectrum of Domain-Specific AI * 00:27:01 What We’d Build Differently Today