The TWIML AI Podcast

Ahmad Al-Dahle, Airbnb’s CTO, discusses Project Everest, an effort to capture institutional knowledge from employee workflows for use by AI agents. The conversation examines stale information, context poisoning and the costs of supplying excessive context, alongside evaluation of whether shared knowledge improves engineering work. It also considers productivity measures beyond token consumption and pull-request counts, limitations of chat interfaces, and the challenges of support agents mediating conflicting interests in a marketplace.

Zoe Weil discusses how recommendation and ranking systems can use in-session behavioral signals to model intent and optimize purchasing outcomes. The episode examines event-to-feature models, low-latency ranking infrastructure, evaluation against business outcomes, and the distinction between click relevance and useful recommendations. It also considers privacy tradeoffs in personalization and the relationship between advertising and organic revenue.

Hosted by Sam Charrington

Diogo Almeida joins Sam Charrington to discuss TypeSafe's Jev model and the use of calibrated decisions as a component of AI software. The conversation examines reinforcement learning from calibrated decisions, differences from classifiers and text-generating models, and the implications for agents, tool use and system architecture.

Hosted by Sam Charrington

Sam Charrington speaks with Greg Burnham, who leads AI capabilities research at Epoch AI, about the progression from elementary mathematical tasks to difficult research problems, including Navier–Stokes. They examine how advanced systems solve mathematical problems, the roles of persistence and existing human work, and the evidence for new ideas. The conversation also considers how to measure progress when traditional benchmarks become less informative, why improvements appear steady across successive model generations, and the remaining weaknesses in open-ended research, learning from experience and choosing productive scientific questions.

Hosted by Sam Charrington

Alex Smola discusses the development of voice agents, audiovisual systems and AI avatars. Topics include audio tokenization, inference cost, conversational latency, visual context and how artificial systems might learn from interaction while responding appropriately to human emotion.

Hosted by Sam Charrington

Christopher Potts examines whether increasing token consumption translates into useful improvements in AI performance. The discussion covers evaluation beyond benchmarks, inference-time scaling, DSPy, interpretability, user expertise and architectural changes that could improve the economics of capable AI systems.

Hosted by Sam Charrington

Justin Johnson joins Sam Charrington to discuss world models and spatial artificial intelligence. Topics include representations of three-dimensional environments, simulation and the role of such models in enabling systems to reason about and act within the physical world.

Hosted by Sam Charrington

Max Welling discusses how ideas from physics may inform the next generation of artificial intelligence. The conversation examines physical structure, generative modelling and the relationship between learning systems and scientific discovery.

Hosted by Sam Charrington

Fatih Porikli examines why scaling image models alone does not guarantee visually correct results. He discusses separating scene planning from rendering, improving controllability and editing, and bringing high-resolution generation to edge devices, including the role of training objectives and reinforcement learning.

Hosted by Sam Charrington

Damian Borth describes learning from trained neural networks themselves as a new source of training data. The conversation explores representations of model weights, transferring knowledge across architectures and tasks, and the possibility of specializing models with less computation by reusing what existing networks have learned.

Hosted by Sam Charrington

Alex Wiltschko explains efforts to predict smell from molecular structure and to give computers a useful representation of odors. He discusses graph neural networks, the difficulty of collecting olfactory data, and how learned embeddings connect chemistry to perception and potential practical applications.

Hosted by Sam Charrington

Devvret Rishi discusses how agents with access to tools and external systems change the security requirements of generative AI applications. He examines the limitations of static controls and the need for runtime policies, observability and recovery mechanisms as agent workflows become more capable.

Hosted by Sam Charrington

Alex Bowcut draws lessons from building an AI system for tax-law research. He explains why retrieval still matters with long-context models, how chunking and combined retrieval methods affect results, and why reliable citations and feedback from domain experts are central to evaluating a specialized application.

Hosted by Sam Charrington

Jure Leskovec connects data-driven biological modelling with relational deep learning. Topics include multiscale virtual-cell models, treating databases as graphs and using in-context prediction across multiple tables. The discussion also covers benchmark evaluation, explanations, agent integration and the practical limitations of relational foundation models.

Hosted by Sam Charrington

Scott Clark discusses operating language-model systems whose failures are difficult to anticipate with offline evaluations. He distinguishes telemetry, known-signal monitoring and production analytics, and describes using trace representations and clustering to discover emergent behaviour and feed improvements into evaluations and training.

Hosted by Sam Charrington

Philip Kiely surveys inference engineering at the intersection of GPU programming, applied research and distributed systems. Topics include batching, quantisation, speculative execution, cache reuse and workload-specific serving, alongside the choices between hosted APIs, dedicated deployments and internal platforms for multimodal and agent applications.

Hosted by Sam Charrington

Rashmi Shetty explains how Capital One develops multi-agent systems for regulated services. The discussion covers intent disambiguation, tool use and human handoffs, alongside platform design, runtime policies, observability and evaluation. It also examines model specialisation through fine-tuning and distillation, and learning from production feedback.

Hosted by Sam Charrington

Stefano Ermon discusses adapting diffusion models to language and code generation. Topics include discrete token spaces, parallel token generation, Mercury 2, and trade-offs with autoregressive models. The conversation examines training, serving and post-training challenges, controllable generation and possible directions for multimodal diffusion systems.

Hosted by Sam Charrington

Siddhant Pardeshi describes autonomous software-development systems that combine knowledge graphs and vector search to navigate large repositories. The conversation examines context engineering, agent orchestration, model-specific prompting and persistent feedback. It also considers how evaluation should reflect security, maintainability and acceptance of the resulting software.

Hosted by Sam Charrington

Sebastian Raschka reviews changes in language-model research, including reasoning-focused post-training, inference-time computation and reinforcement learning with verifiable rewards. The discussion connects these methods with self-consistency, model architectures, efficient attention and long contexts, while examining limits in continual learning and the reliability of agent workflows.

Hosted by Sam Charrington