The TWIML AI Podcast

In this episode, Stefano Corazza, head of AI research at @canva, joins us to explore how his team is building custom models and agents around iterative design workflows. We discuss the trade-offs between training models from scratch and adapting open weights models, and how proprietary design data supports specialization. Stefano explains how Craft separates layout planning from pixel generation and how Brush uses design tools and visual feedback to create and revise compositions. We examine the challenges of evaluating design quality beyond a single generation, including the use of vision language models as judges alongside human evaluation. We also discuss lessons from simplifying agent harnesses, why faster inference enables more iterations, and how model ownership creates opportunities to optimize serving costs. Finally, we explore how APIs and MCP are expanding access to these capabilities and what changes when software companies build tools for agents as well as people. 🗒️Full show notes: twimlai.com/go/784. 📖 CHAPTERS =============================== 00:00 - Introduction 01:19 - Canva AI Research, Craft, and Brush 04:25 - Why Canva Builds Custom Models 09:43 - Canva Craft Layout-to-Pixels Architecture 12:27 - Canva Magic Layers 14:54 - Training Data and Detailed Scene Descriptions 21:45 - Synthetic Data and Diversity in Design Generation 22:53 - Evaluating Image Generation Models and Design Editing 25:00 - Vision-Language Models for Evaluating Design Quality 26:30 - Tuning Model Taste and Visual Style 31:13 - Canva Brush 35:49 - Simplifying Agent Harness 37:36 - Code Generation 39:44 - Optimizing Inference Stack 47:54 - Distilling Craft for Speed and Cost 50:24 - Canva Models Available Through MCP and APIs 54:23 - What’s Next for Canva Research

An hour to go until our Generative AI study group starts! Exchange hardware acceleration techniques, dig into hybrid attention models, explore recursive self-improvement, and more. See you there!

Just one day until our next Generative AI study group! Explore agent sandboxing, share hyperparameters and tuning strategies, discuss model quantization approaches, and more. Sign up at twimlai.com/community/ and we’ll see you tomorrow at 8 am PT.

Our Generative AI study group centers around sharing insights and experiences with AI platforms, comparing newly released models and features, exploring developer tools and products, and more! Anyone interested in joining can register at twimlai.com/community/ and join us every Friday at 8 am PT.

Learn from AI experts, share challenges and bottlenecks in your projects, exchange ideas with fellow AI enthusiasts, and more in our Generative AI study group! All skill levels are welcome to join. Just register at twimlai.com/community/ and we’ll see you in an hour!

AI can help scientists write code, analyze data, and draft papers. But can it generate the ideas that lead to new discoveries? @RichardSocher, former chief scientist at Salesforce, founder of @youdotcom, and co-founder of @Recursive_SI, joins us to explore the move from AI assisting researchers to AI conducting research. Starting with AI improving AI, we examine what it takes to propose ideas, run experiments, and validate results, and how those capabilities could eventually accelerate discovery across biology, chemistry, and physics. The conversation covers Recursive’s reported results in model training and GPU kernel optimization, why fast verification matters, and how reward hacking complicates the search for real improvements. We also discuss the data and experimental infrastructure needed for broader scientific discovery, whether commercial incentives can keep increasingly capable systems aligned, and Richard’s vision for a “Eureka machine.” 🗒️ Full show notes: twimlai.com/go/783. 📖 CHAPTERS =============================== 00:00 - Intro - Richard Socher and the Eureka Machine 03:11 - AI for Science Examples across Physics, Chemistry, and Biology 05:47 - Organoids, Virtual Cells, and Drug Repurposing Case Studies 10:16 - Evaluating AI for Science Timelines, Hype, and Physical Bottlenecks 13:47 - Four Pillars of the Eureka Machine and AI Scientist Agents 16:37 - Open-Endedness and Environments for Scientific Ideation 21:04 - Guided Search, Reward Engineering, and Reward Hacking 26:02 - Can AI Truly Innovate? Recursive Self-Improvement in Practice 29:25 - Recursive’s Reported Results on NanoChat, NanoGPT, and CUDA Kernels 33:02 - Verification Speed, Long-Horizon Attribution, and System Generalizability 37:49 - Scaling Compute, Biological Data, and Mitigating Reward Hacks 40:11 - Commercial Incentives and AI Alignment 41:46 - AI Pacing Debates and Regulating Applications vs. Intelligence 44:29 - Dual-Use Science Risks, Accountability, and the Case for Progress

Just one hour away before our Generative AI study group starts! Exchange effective generative AI strategies, collaborate on AI projects, share news on the latest AI developments, and more! We’ll see you in a bit!

Join our Generative AI study group and explore topics like long-horizon agentic coding, spatial computing, Mixture-of-Experts (MoE) routing techniques, and more! Simply register at twimlai.com/community/ and join tomorrow’s session at 8 am PT.

Join our community of AI enthusiasts in our Generative AI study group every Friday at 8 am PT. Dig into token-level knowledge distillation, explore latent-space world models, discuss continual learning techniques, and more! Register at twimlai.com/community/ to join the conversation.

AI agents are becoming accessible to everyone, raising the bar for developers building the next generation of agent products. When capable, general-purpose agents are available out of the box, differentiation increasingly comes down to the architectural decisions and specialized experiences that make an agent uniquely valuable. In this episode, @Dan_Jeffries1, creator of @nautiloapp, joins us to discuss what he's learned building a platform for personal and collaborative AI agents. We explore his approach to persistent memory, coordinating work across people and agents, managing private context, and designing deterministic tools that complement LLM capabilities. Dan also shares his views on multi-agent architectures, computer use, and the tradeoffs between local and cloud models. The conversation offers a practitioner's perspective on where the opportunities lie as AI agents become more capable, more accessible, and more difficult to differentiate. 🗒️ Full show notes: twimlai.com/go/782. 📖 CHAPTERS =============================== 00:00 - Introduction 02:01 - The agency problem 07:37 - Nautilo 13:46 - Multi-agent systems 19:42 - Context window limitations 22:21 - Tool design philosophy: "Magical Hammer Design” 27:09 - Venice 28:48 - Local models and memory constraints 35:33 - Balancing privacy with model selection 38:24 - Bring your own keys and subsidized compute 40:39 - Co-creative tools 45:40 - Building agent products against “the wall” 49:08 - Bespoke software and the challenge of defensible moats

We’re just an hour away from the start of this week’s session of our Generative AI study group! Dig into agentic engineering, discuss KV-cache compression techniques, explore latent Mixture of Experts (MoE), and more! Make sure you’ve joined our Slack community after signing up on twimlai.com/community/. We’ll meet you there!

🕗 Tomorrow at 8 am PT, our Generative AI study group meets to share insights on the latest AI developments, discuss sparse coding techniques, trade model evaluation strategies, and more! No experience required—just sign up at twimlai.com/community/, and we’ll see you tomorrow at 8 am PT.

Our weekly Generative AI meetup occurs every Friday at 8 am PT. Register at twimlai.com/community/ and dig into Heavily Compressed Attention (HCA), exchange prompt engineering tricks, explore video-language reasoning models, and more! The discussions are informal and participants are free to join or leave at any time during the session.

AI agents need the institutional knowledge experienced employees rely on: why a system works a certain way, which policies apply, and what never made it into the documentation. Capturing that knowledge is hard. Keeping it useful as people and agents add more may be harder. In this episode, @Ahmad_Al_Dahle, CTO of @Airbnb, shares how his team is building Project Everest, a “company brain” that captures knowledge from employees’ workflows for AI agents. We explore stale information and context poisoning, why too much context can backfire, and how the team evaluates whether shared knowledge improves engineering work. We also discuss measuring AI productivity beyond token usage and PR counts, the limits of chat-based interfaces, and deploying support agents in a marketplace where resolving disputes means mediating between conflicting interests. 🗒️ Full show notes: twimlai.com/go/781. 📖 CHAPTERS =============================== 00:00 - Intro 01:35 - Three Pillars of AI at Airbnb 03:19 - AI for Engineering Productivity: Balancing Velocity and Craft 06:20 - Project Everest 10:30 - Everest architecture options: databases, files, and retrieval 12:25 - Stale context, context poisoning, and progressive disclosure 13:14 - AirChat and a multi-model coding stack 14:38 - Engineering evals and calibrating model judges 17:37 - Managing model costs without token quotas 20:12 - Measuring engineering value beyond tokens and PR counts 24:12 - Behind-the-scenes agents for marketplace trust and quality 26:36 - Adversarial support mediation: evidence, policies, and conflicting interests 30:33 - Deployed support agents and limits requiring human judgment 33:05 - Support workflows and choosing human escalation 36:05 - Natural language search without a chat-first interface 39:41 - External agent access, MCP, and travel trust 44:09 - Building agent frameworks versus adopting open source 45:55 - Post-training small models for specialized, low-latency tasks 48:25 - Scaling Everest and unresolved agent experience challenges 50:54 - Conclusion

Get ready to join our Generative AI study group in one hour! Make sure you’re registered at twimlai.com/community/ to participate in the informative discussion, demonstrate personal projects, showcase cutting-edge tools and apps, and much more! See you there!

Be part of tomorrow's session of our Generative AI study group. Join us by registering at twimlai.com/community/ and explore decoder-style transformers, share insights on latency and throughput optimization, exchange alignment training strategies, and more. Our study group is open to everyone, so feel free to invite your peers!

In this episode, Zoe Weil, co-founder and CEO of @sequen_ai, joins us to discuss how ranking systems can use live behavioral signals to optimize for purchases rather than clicks or search relevance alone. Zoe explains why approaches developed at data-rich social platforms don’t necessarily translate to retailers and how Sequen uses scroll speed, hover duration, and other in-session interactions to infer shopper intent without relying on user identity. We explore the company’s large event models and its architecture of specialized models that transform streaming events into features for downstream rankers. Zoe describes the infrastructure required to serve dynamic rankings within milliseconds, why the team chose Rust for its backend, and how evaluation harnesses connect model improvements to business outcomes. We also discuss the tradeoffs between advertising and organic revenue, continuous model improvement, and the boundary between useful personalization and surveillance. 🗒️ Full show notes: twimlai.com/go/780. 📖 CHAPTERS =============================== 00:00 - Introduction 02:56 - Sequin’s Ranking Intelligence Layer 07:26 - Difference with Search and Relevance 13:20 - Session Live Signaling 15:57 - Surveillance Capitalism 18:45 - Building a Low-Latency Stack in Rust 20:52 - Model Architecture 26:47 - Reranking Products and Surfacing Complementary Items 33:03 - Balancing Organic and Ad Revenue 34:56 - Learning from Implicit Feedback 36:42 - Infrastructure 39:34 - Recursive Stack and Evaluation Harnesses 42:31 - Recursive Ranking Intelligence 44:12 - Large Event Models 46:56 - Time-Series Modeling 48:49 - Complexity of Scaling Real-Time Feature Transformations 49:46 - Future Directions

Join the conversation in our Generative AI study group in one hour! Share model compression techniques, explore open-weight reasoning models, dig into embedding architectures, and more! Meet you in an hour!

Join our next Generative AI study group session tomorrow at 8 am PT and dig into long-context scaling techniques, explore linear attention hybrids, discuss orchestration frameworks for multi-agent workflows, and more! Register at twimlai.com/community/.

Join our interactive Generative AI study group every Friday at 8 am PT where we chat about the latest news in AI, explore new ideas, exchange perspectives and recommendations, and more! Registration is available at twimlai.com/community/ and don’t forget to join the #generative-ai channel in our Slack community.