TWIML

Stanford professor and Big Spin co-founder Chris Potts joins Sam Charrington to discuss AI tokenomics and his research into “tokenflation”—the possibility that token usage is growing faster than the measurable value those tokens produce. They explore how to measure return on AI spending, why benchmarks alone give an incomplete picture, and what inference-time scaling means for the economics of increasingly capable models. Chris also shares why expert AI users get better results by challenging and iterating with models, and how AI fluency affects outcomes. The conversation covers DSPy, interpretability, the limits of today’s transformer architectures, and where Chris sees opportunities for more fundamental innovation in AI.

Justin Johnson, co-founder of World Labs, explores the emerging field of spatial AI and world models—systems that can understand, generate, and simulate 3D environments. He discusses the divide between explicit 3D representations and generative approaches, and why there is still no consensus on how to build these models. The conversation covers World Labs’ Marble system, which creates navigable 3D worlds from images, the open challenges of evaluating world models, and the role of simulation, planning, and action. Johnson also shares his vision for unified models that could power interactive virtual environments and agents that operate in the physical world.

Max Welling, co-founder and CTO of CuspAI and professor at the University of Amsterdam, discusses how physics could inspire the next generation of AI systems. He explains CuspAI's work using generative AI to design new materials for semiconductors, batteries, and clean energy. The conversation then explores how physics concepts like thermodynamics, waves, and symmetry breaking might lead to AI architectures beyond the current scaling paradigm.

Text-to-image models have become remarkably good at producing realistic images, but Fatih Porikli argues that realism is not the same as correctness — models still fail at several distinct people, precise compositions, and local high-resolution generation. The Qualcomm Vice President of Technology walks through approaches his team presented at CVPR: better training objectives for controllability, separating scene planning from rendering, and generating 16-megapixel images efficiently on edge devices. The conversation also covers eliminating visible artifacts in AI-powered editing, reinforcement learning for image generation, and agentic pipelines. Porikli maps what remains unsolved in generative vision and where the next phase of progress is likely to come from.

Damian Borth argues that as high-quality training data becomes scarcer and pretraining costs soar, AI research should treat trained neural networks themselves as training data. His work on weight space learning explores how foundation models can learn from the distilled optimization results of existing models rather than raw data, potentially reducing the cost of developing specialized models and reshaping how future AI systems are trained.

Alex Wiltschko, founder and CEO of Osmo, discusses building artificial olfactory intelligence—teaching computers to smell by mapping the relationship between molecular structure and human odor perception. The episode covers the neuroscience of smell, graph neural networks for encoding scent in multi-dimensional embedding spaces, Osmo's proprietary olfactory dataset, and near-term applications in fragrance, disease detection, and emotion sensing.

Dev Rishi, GM of AI at Rubrik, explains why traditional enterprise security approaches fail when AI agents move beyond answering questions to taking autonomous action across tools and systems. The episode explores how agent capabilities — planning, tool calling, code generation, and workflow automation — create new security vulnerabilities that static guardrails and human approval can't contain, and what defense-in-depth approaches enterprises need instead.

Alex Bowcut, Head of Engineering at Sphere, challenges the narrative that retrieval-augmented generation (RAG) is obsolete now that context windows accommodate millions of tokens. He explains why RAG remains critical in high-stakes domains like global tax compliance, where accuracy alone isn't enough—every AI conclusion must be backed by precise legal citations and withstand expert scrutiny. Sphere's TRAM system combines retrieval, reasoning models, legal review workflows, and reinforcement learning to help tax experts work nearly two orders of magnitude faster while maintaining trustworthiness and accuracy.

Jure Leskovec, co-founder of Kumo and Stanford CS professor, discusses two major research directions: AI Virtual Cell, a multiscale approach to learning representations from proteins to patients using single-cell RNA-seq and protein language models; and relational deep learning, which reframes enterprise databases as graphs for direct neural network training. He introduces Kumo's Relational Foundation Model (RFM2), demonstrating in-context learning over subgraphs for zero-shot predictions on new databases, with real-world deployments at Reddit, DoorDash, and Coinbase.

Scott Clark, co-founder and CEO of Distributional, discusses how teams can reliably operate and improve complex LLM systems and agents in production. The episode explores a Maslow's hierarchy of observability framework, real-world production failures that standard evals miss (like lazy tool-use hallucinations), and how vector fingerprinting of traces enables clustering to uncover emergent behaviors. Clark explains how online analytics feed a data flywheel to generate evals, guardrails, and training data, and why adaptive approaches are essential for non-stationary models.

Philip Kiely, head of AI education at Baseten, discusses inference engineering as the stickiest and most critical AI workload. The conversation covers the blend of GPU programming, applied research, and distributed systems that define inference, and how teams can now go from research to production in hours rather than months. Philip explains the key knobs—batching, quantization, speculative decoding, and KV cache reuse—that let teams balance latency, throughput, and cost. He traces the maturity journey from closed APIs to dedicated deployments and in-house platforms, surveys the runtime landscape (vLLM, SGLang, TensorRT LLM), and looks ahead to the need for specialized, workload-specific runtimes as agents and multimodality demand ever more efficient inference.

Yejin Choi discusses her research on making small language models reason more effectively, challenging the assumption that scale is the only path to intelligence.

Nikita Rudin discusses the gap between current robotic capabilities and what is required to deploy fully autonomous robots in the real world.

Aakanksha Chowdhery from Reflection explores the fundamental shifts required to build true agentic AI that can plan, act, and learn autonomously.

Munawar Hayat from Qualcomm AI Research discusses key papers from NeurIPS 2025, including advances in multimodal AI and solutions for object hallucination in vision-language models.

Zain Asgar from Gimlet Labs discusses running AI inference across heterogeneous hardware — mixing GPUs, CPUs, and custom accelerators for optimal performance.

Devi Parikh, co-founder of Yutori, discusses browser use models and how autonomous AI agents can navigate and act within web environments.

Robin Braun from HPE and Luke Norris from Kamiwaza discuss enterprise AI infrastructure challenges and how organizations can scale AI deployment effectively.

A deep dive into reasoning architectures: how chain-of-thought, tree-of-thought, and other inference-time compute strategies work and when to use each.

Advances in robotic manipulation: how AI systems learn to grasp, move, and manipulate objects in unstructured environments.

Research on LLM security: prompt injection attacks, jailbreaking techniques, data extraction, and emerging defense strategies.

Insights from training large foundation models: data curation, compute allocation, and the increasingly important role of data quality over data quantity.

Moving from responsible AI principles (fairness, transparency, accountability) to practical implementation — the organizational structures, evaluation pipelines, and cultural changes required to build AI responsibly.

How diffusion models are being applied beyond image generation: video, 3D modeling, protein design, and audio synthesis.

How organizations must rethink data governance for AI: new challenges around training data provenance, model outputs, and the blurred line between data and code.

Runway CEO Cristobal Valenzuela discusses whether video generation represents a path toward artificial general intelligence — the argument that world models trained on video data must develop physical reasoning.

The state of reinforcement learning: from game-playing to robotics to RLHF. How RL has evolved from a research curiosity to a core component of modern AI systems.

The emerging field of edge AI deployment — running machine learning models on devices (phones, cameras, sensors) rather than in the cloud. The challenges of model compression, latency, and power efficiency.

Bridging the gap between AI ethics theory and practice. How organizations can move from principled statements to concrete implementation of ethical AI.

Alexandre Pesant from Lovable discusses the evolution and practice of vibe coding — building software through natural language descriptions rather than manual code.

Atlassian co-CEO Mike Cannon-Brookes shares internal data on AI tool adoption across 300,000+ enterprise customers — what AI features are actually used, which are ignored, and what this reveals about enterprise AI readiness.

A clinical informatics researcher separates the hype from reality in healthcare AI: which applications are actually deployed in clinical settings, which remain research demonstrations, and why the gap exists.

A critical examination of AI safety evaluation methodologies: red-teaming, benchmarks, and stress testing — what they actually measure, what they miss, and why the field lacks agreed-upon standards.

Arvind Jain, CEO of Glean, discusses the evolution from enterprise search to agentic AI tools for the workplace — how Glean went from finding documents to acting on them.

The convergence of vision, language, and audio in multimodal generative AI systems — how models that can see, speak, and write simultaneously are creating new capabilities and new challenges.

GitHub CEO Thomas Dohmke discusses how AI is transforming software engineering through Copilot and beyond — the shift from AI as autocomplete to AI as a development partner.

How AI is transforming scientific discovery beyond the headline achievements of AlphaFold — applications in materials science, drug discovery, climate modeling, and mathematics.

A deep dive into synthetic data: how AI-generated training data is replacing real data in many applications, the quality challenges, and the risk of model collapse when models are trained on outputs of other models.

Christopher Manning from Stanford discusses the intersection of linguistics and large language models — whether LLMs truly understand language or process it as statistical patterns, and what this distinction means for the field.

A nuanced discussion of the risks and benefits of releasing open-weight AI models, challenging both the AI safety maximalist view (all models should be closed) and the open-source maximalist view (all models should be open).

The AI2 team discusses OLMo, their fully open-source large language model that releases not just weights but training data, training code, and evaluation frameworks — everything needed to reproduce and study the model.

An exploration of how training data composition affects chain-of-thought reasoning capabilities in LLMs, and whether prompting techniques actually elicit genuine reasoning or sophisticated pattern matching.

JPMorgan AI Research discusses DocLLM, their approach to reasoning over complex business documents that combine text, tables, and figures — a fundamental challenge for enterprise AI deployment.

A critical examination of whether emergent abilities in large language models are genuine phase transitions or measurement artifacts caused by nonlinear evaluation metrics.

Kamyar Azizzadenesheli from NVIDIA discusses how the reinforcement learning community is adapting to the dominance of large language models — where RL techniques enhance LLMs and where LLMs change RL.

Naila Murray discusses 2024 trends in computer vision: the integration of vision with language models, the challenge of 3D understanding, and how video generation is reshaping the field.

Thomas Dietterich discusses 2024 AI trends in machine learning and deep learning, including the tension between monolithic LLMs and modular architectures, hallucination challenges, and the role of uncertainty quantification.