Subvolts take Learn how to build safer AI by controlling models from within, not just with external filters. Jan 20, 2026 - Developers
- Learn about advanced techniques for building safer AI systems using model-native controls.
- Business
- Understand how to implement more effective and efficient AI safety measures beyond basic filters.
This episode provides a solid interview with an expert on a critical topic: making advanced AI systems safer and more understandable.
Controlling AI Models from the Inside
Audio streams from Daniel Whitenack and Chris Benson. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
This Practical AI episode features Alizishaan Khatri discussing controlling AI models from the inside. They explore moving beyond traditional guardrails for generative AI, addressing limitations of black-box defenses. The conversation focuses on interpretability and using model-native, runtime signals to build safer AI systems.
Imagine AI models as complex machines. Instead of just putting fences around them, this episode talks about building better controls directly inside the machine. This helps make AI safer and easier to understand for everyone.
This episode offers a deep dive into AI safety and interpretability, crucial for deploying generative AI. It's for developers and researchers grappling with the limitations of current AI defenses. The discussion on model-native signals and runtime controls is relevant as frontier labs like OpenAI, Google Gemini, and Anthropic work to make their advanced models more robust and trustworthy for enterprise use.
- HostDaniel Whitenack
- HostChris Benson
- GuestAlizishaan KhatriWrynx
- Guest: Alizishaan Khatri of Wrynx
- Show: Practical AI
- Length: 44 minutes
- Topic: Controlling AI Models from the Inside
Chapters from the publisher ↗ · Transcript ↗ Key moments and transcript search come with a later piece.
As generative AI moves into production, traditional guardrails and input/output filters can prove too slow, too expensive, and/or too limited. In this episode, Alizishaan Khatri of Wrynx joins Daniel and Chris to explore a fundamentally different approach to AI safety and interpretability. They unpack the limits of today’s black-box defenses, the role of interpretability, and how model-native, runtime signals can enable safer AI systems.
Featuring:
- Alizishaan Khatri – LinkedIn
- Chris Benson – Website, LinkedIn, Bluesky, GitHub, X
- Daniel Whitenack – Website, GitHub, X
Upcoming Events:
- Register for upcoming webinars here!
- guardrails
- Alizishaan Khatri
- safety features or rules for AI
- interpretability
- Alizishaan Khatri
- understanding how an AI makes decisions
- black-box defenses
- Alizishaan Khatri
- safety measures that don't reveal the AI's inner workings
- runtime signals
- Alizishaan Khatri
- information generated by the AI while it is actively working
44 minNarrative intelligence and the human advantage
WithReed Frerichs
47 minOpen models and the future of Physical AI with NVIDIA
WithMing-Yu Liu
49 minFrom AGENTS.md to Enterprise Deployment
WithNick Kuhn
48 min
56 min
45 minHow Mintlify Is Rebuilding Documentation for Coding Agents
WithHan Wang, Hahnbee Lee
Questions people ask
- What are the limitations of current AI guardrails?
- Traditional guardrails and input/output filters can be too slow, too expensive, or too limited for generative AI in production.
- What is the proposed alternative for AI safety?
- The episode explores a fundamentally different approach using model-native, runtime signals to enable safer AI systems, focusing on interpretability.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Daniel Whitenack and Chris Benson.