Topics / Multimodal AI
Multimodal AI Podcasts
4 shows and 1,808 episodes sit under Multimodal AI.
Curated podcast guide to multimodal AI — Fei-Fei Li on world models, SAM 3, Genie 3, Sora 2, and the VLM research that goes beyond the demo reel
Photo: Brecht Corbeel on Unsplash · Unsplash License
Swipe · Multimodal AI only
Swipe through Multimodal AI
Save time discovering your next favourite podcast by listening to just the best minute of an episode.
The multimodal ai slate · every cover a link
All 4 multimodal ai shows →
Next to this on the shelf
Subtopics to explore
12 subtopics
A field guide
Finding the right Multimodal AI show for you
When these conversations work, you can hear the model being pulled out of the chat box. The strongest episodes stay close to what a system can see, generate, or connect to the physical world—and where those connections break. Audio makes that harder. There is no demo to hide behind.
Good hosts turn the missing screen into an advantage by slowing down the architecture, failure mode, or product decision that would otherwise flash past. Go up to Artificial Intelligence for the broader model and industry picture. AI Reasoning follows how models plan and solve problems; AI Agents is about systems taking action, while AI Infrastructure stays with the machinery required to run them.
When seeing goes wrong
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) offers the clearest route into visual systems. Why Vision Language Models Ignore What They See with Munawar Hayat puts the central failure plainly. High-Efficiency Diffusion Models for On-Device Image Generation and Editing with Hung Bui moves toward generation under practical constraints. Then Intelligent Robots in 2026: Are We There Yet? with Nikita Rudin carries the question into embodied machines.
Those titles mark three distinct doorways: perception, image creation, and action in the world. TWIML is the place to start when “multimodal” feels too broad and you want one concrete technical problem at a time.
From research system to product
No Priors: Artificial Intelligence | Technology | Startups takes a measured, peer-briefing approach. Its interviews make room for long technical answers, with questions aimed at architecture and strategy rather than basic orientation. From Restoring Sight to Reimagining the Brain, with Max Hodak sits at an ambitious boundary between sensing and biology. Building an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang points toward machines operating beyond a browser window.
The appeal is the change in altitude. A technical idea is allowed to remain technical, but the conversation keeps asking what kind of company, system, or deployment could carry it into ordinary life.
Engineers at the hinge
Latent Space: The AI Engineer Podcast is brisker and messier in a useful way. It assumes the listener can handle dense vocabulary, then keeps tugging the discussion back toward mechanisms and developer choices. 🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI is a sharp declaration of scope. 🔬 The Self-Driving Lab — Joseph Krause, Radical AI pushes that interest toward automated physical research.
Machine Learning Street Talk changes the cadence again. It is fast solo commentary, comfortable turning a product announcement into a wider market thesis and then questioning the hype around it. Multimodal coverage is narrower here; examples include AI Cloning Tech in Machine Learning and Meta's Gemini 4 and Future Directions. Treat it as the side conversation: quick, speculative, and useful after the engineering-heavy interviews.
Where to start
Begin with The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) on vision-language failure if you want the cleanest technical entry. Choose Latent Space: The AI Engineer Podcast for diffusion research and builder-oriented conversation. No Priors: Artificial Intelligence | Technology | Startups is the better on-ramp when product strategy matters as much as model design. Save Machine Learning Street Talk for a shorter burst of market-minded commentary.
Featured shows below are part of the PlayNext catalog. Updated weekly.
Featured shows
4 shows, described rather than listed
Each one comes with a short write-up and the role it plays in this guide.
The deepest interview slot for the people building multimodal frontier models. Fei-Fei Li on world models, SAM 3 with Meta researchers, Khosla on the largest non-LLM bet of the cycle.
Sam Charrington's academic-to-applied bridge for multimodal research. VLM architecture, world models, on-device inference — the substance, not the demo reel.
Tim Scarfe's methodology-first take on the consumer multimodal rollout. Sora 2, image gen at scale, OpenAI's audio bet — the public-facing layer, examined.
Sarah Guo and Elad Gil's investor framing. Bridge cite — useful for how the world-models thesis evolved across 2025 from VC and operator angles.
Where to start
3 routes in, one per reader
Fei-Fei Li argues the world-models thesis directly in the After LLMs episode — the cleanest single statement of where the frontier is moving beyond pure text.
Inside Nano Banana goes into the actual VLM architecture; Why Vision Language Models Ignore What They See covers the systematic failure modes.
AI Video at Scale: OpenAI's Sora 2 is the cleanest read on how the Sora launch actually played out in late 2025.
Glossary
Jargon you'll hear on Multimodal AI podcasts
AI system that simulates an environment rather than just describing it. Argued in 2025-2026 as the next foundation layer beyond text — Genie 3 from DeepMind, Fei-Fei Li's World Labs, Khosla's largest non-LLM bet.
Model that processes images and text together. The architecture under Nano Banana, GPT-4o vision, Claude vision; the layer where research and consumer demos diverge most sharply.
Fei-Fei Li's frame for the kind of 3D scene understanding that text-only models can't represent. The core argument behind World Labs and the world-models thesis.
Meta's third-generation Segment Anything Model. The perception substrate that many world-model and robotics arguments assume underneath them.
OpenAI's late-2025 consumer video-generation model. The most visible rollout in this beat; both technically meaningful and a product story about distribution.