Distillation: Making AI Models Better With Less

AI Is Better Than Us At So Much, So Why Don't We Still Trust It?

We are standing on the edge of a cliff in 2025, facing a "Great Disconnect": machines that are empirically better at staying alive than we are, yet a public that would rather trust a tired, distracted human driver than a computer that never blinks. This episode explores the "proficiency gap" in industries from autonomous trucking in Texas to AI-driven healthcare and legal arbitration. We dive into the staggering data from Aurora Innovation, where simulations proved that AI could have avoided every single fatal human collision on Interstate 45 between 2018 and 2022. If the machine is now the expert, why does the 2024 Edelman Trust Barometer show that the public is still recoiling from AI innovation? The answer lies not in math, but in the "Perfection Penalty" and "Algorithm Aversion". We explore why we forgive a "hangry" human judge but demand impossible transparency from a "Black Box" neural network. We analyze high-stakes "socially illiterate" failures—from Waymo vehicles maneuvering through active police gunfights to AI-driven buses illegally passing stopped school buses. As we examine the "Superposition Hypothesis" and the struggle to untangle millions of AI parameters, we ask the ultimate question: can we ever trust a partner that can calculate orbital mechanics but cannot "read the room" in a simple construction zone?

How to Design a Flying Taxi

Imagine bypassing a 90-minute Manhattan gridlock for a smooth, seven-minute electric flight to the airport. This isn't a far-fetched dream; it’s the reality of Urban Air Mobility, powered by electric vertical takeoff and landing aircraft (eVTOLs). This episode goes behind the low hum of electric motors to reveal a hidden world of equations and algorithms balanced on a knife's edge. We deconstruct the "tyranny of the vertical"—the immense power required to lift an object straight up—and explain the critical physics of disk loading. Using the "snowshoe versus stiletto" analogy, we explore how rotor size dictates everything from hover efficiency to the noise levels essential for a license to operate in dense cities. We profile the radical blueprints currently racing toward a 2025 commercial launch, each offering a unique solution to the disk loading problem. Discover the mechanical genius of Joby Aviation’s tilt-rotor S4, the pragmatic "lift-plus-cruise" approach of Archer’s Midnight, and the high-speed efficiency of the Lilium Jet, which uses 30 small, ducted electric fans embedded directly in its wings. From managing the delicate transition between vertical and forward flight to minimizing aerodynamic drag, we uncover the unseen web of math and engineering that makes the "vertical dream" possible. Join us as we prepare to navigate the "urban jungle" of the very near future.

Distillation: Making AI Models Better With Less

In early 2025, a small Chinese startup called DeepSeek sent a shockwave through the technology sector. Their R1 chatbot didn't just rival the performance of the world’s most advanced reasoning models—it did so at a fraction of the cost, wiping $600 billion off Nvidia's market cap in a single day. This episode deconstructs the controversial "computer science trick" at the heart of this disruption: Knowledge Distillation. We explore how the industry is moving away from "Foundational Giants" that cost hundreds of millions to train, toward a more democratized era where startups and independent researchers can finally compete with the tech titans. We dive into the "Teacher-Student" process that makes this possible. Imagine a master chef teaching a student not just to follow a recipe, but to understand the "intuition" behind every ingredient. We explain the difference between White-Box Distillation, which accesses a model's internal "soft targets" or "dark knowledge," and the more elusive Black-Box Distillation used by DeepSeek to mimic a teacher's behavior through millions of targeted questions. As we look toward 2026, we tackle the growing legal and ethical gray areas of this "AI shrinking ray": what happens to intellectual property when a company’s public-facing product effectively becomes the training data for its competitors?

Physics That Gave Birth to AI

While we often credit Silicon Valley for the rise of artificial intelligence, the true seeds of modern AI were planted decades ago in the obscure world of theoretical physics. This episode reveals how the "physics of disorder"—specifically the study of random metal alloys called spin glasses—provided the radical blueprint for machines that can learn and remember. You’ll discover how physicists like John Hopfield realized that a network of firing neurons behaves exactly like a grid of atomic magnets, allowing him to model human memory not as a digital warehouse, but as a physical "energy landscape" where memories are deep, stable valleys. The journey continues with the introduction of "temperature" into these digital minds. By repurposing the laws of thermodynamics, researchers like Geoffrey Hinton created the Boltzmann machine, using controlled randomness to help AI escape "local minimums"—shallow, incorrect valleys of thought. This thermal jiggling allowed networks to move beyond simple mimicry to grasp the underlying statistical patterns of data, sparking the first "generative" stirrings of AI. Join us as we trace this incredible evolution from frustrated atoms to the deep learning explosion, proving that the universal drive toward minimum energy is the hidden engine behind the world's most advanced thinking machines

Mathematics of VR and AR

When Apple released the Vision Pro, it wasn't just a new gadget; it marked the official entry into the era of spatial computing. This episode peels back the curtain on the "science of the impossible," exploring how these devices act as translators between messy human biology and rigid silicon logic. We break down the five layers of technology required to build a virtual world—from the "Senses In" army of cameras and LiDAR scanners to the "Life Support" systems that process over one billion pixels per second. We dive deep into the central mathematical challenge of SLAM (Simultaneous Localization and Mapping), revealing how your glasses "know" where you are by constantly blending internal motion data with external visual landmarks to lay 3D "graph paper" over your real-world environment. We also explore the "Illusion Out" stack, comparing the "pancake lenses" that fold light paths to keep headsets thin with the "waveguides" that use microscopic nano-gratings to plumb virtual light directly into your pupils. Finally, we uncover the ultimate computational "cheat" that makes high-resolution VR feasible: foveated rendering. By using infrared cameras to track exactly where your eye is focusing, headsets only render a tiny circle of the screen at full detail, leaving your peripheral vision blurry to save the chip from overheating and the battery from draining. This mathematical triage is the keystone that allows a 23-million-pixel display to function in a wearable device.