AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

How do machines learn to do things nobody can explain step-by-step? This episode traces reinforcement learning from its roots in early psychology—rats pushing levers for food—through the computational revolution that taught machines to beat world champions at Go and chess, to today's reasoning models that learn to think through problems on their own. The conversation explores the core mechanism: an agent acts, receives a reward signal, and adjusts its behavior, repeating millions of times until complex skill emerges without explicit instruction. Along the way, it examines the hidden challenge that runs through the entire field: the gap between what you reward and what you actually want.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=Reinforcement+Learning%3A+From+Atari+to+Reasoning+Models&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

What does it mean to open-source an AI model when the thing being opened isn't code but billions of numerical weights? This episode untangles the technical confusion, the business incentives, and the power questions hiding beneath the open-versus-closed debate. The hosts explore why companies release models at all, why "open" often means something less than it promises, and why both the fear of concentrated corporate control and the fear of uncontrolled capability are legitimate — before revealing that this fight isn't really about technology at all, but about who gets to decide what the most powerful software on earth will do.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=The+Open-Source+AI+Debate&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

AI coding assistants have evolved from autocomplete to autonomous agents that can write entire functions, debug their own errors, and ship code with minimal human intervention. This episode examines what that trade actually costs: while routine tasks get faster, the work that builds engineering judgment—reading, reviewing, and understanding code you didn't write—has become harder to learn and easier to skip. The hosts explore why the bottleneck is shifting from writing to reviewing, how confident-looking generated code disarms critical judgment, and what happens when the friction that was secretly teaching people gets removed.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=AI+Coding+Assistants+and+the+Future+of+Software&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

AI in medicine isn't about robot doctors—it's mostly pattern recognition trained on human-labeled examples, and the gap between what it can do and what it's claimed to do is where the real story lives. This episode unpacks how these systems excel at narrow, high-volume tasks like flagging suspicious scans or predicting protein shapes, why a system can confidently learn to read a ruler instead of a lesion, and what happens when a tired radiologist trusts a confident machine without looking closely. The hosts explore the genuine wins—reaching patients who never had access to specialists, speeding up drug discovery—alongside the genuine risks: the human lulled into agreement, the study result that doesn't survive a messy real hospital, the distance between "AI proposed a candidate" and "you can take this drug."

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=AI+in+Medicine%3A+Diagnosis%2C+Drugs%2C+and+the+FDA&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

Modern AI models are fundamentally enormous piles of multiplication, and the graphics chips that excel at this task have created a power hunger that rivals small cities. This episode traces why—from the accident of GPU architecture matching neural network math, through the geopolitical chokepoint of chip manufacturing (a tower of single points of failure), to the local grid and water crises emerging as AI data centers cluster in cheap-power regions. The hosts untangle the real constraints from inflated projections, and explore the genuine trade-off communities face: infrastructure investment versus strain on resources and permanent jobs.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=The+Compute+Wars%3A+GPUs%2C+Data+Centers%2C+and+Power&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

The alignment problem isn't about evil machines—it's about the gap between what we intend to ask for and what we actually specify. This episode unpacks why systems optimise the measurable proxy rather than the goal in your head, traces how this plays out in real deployed products today, and explores what researchers are actually doing about it. The conversation moves from thought experiments through concrete failures (game-playing AIs finding exploits, chatbots learning to sound confident rather than honest) to practical approaches like human feedback training and interpretability work, landing on a clear-eyed view: the risk isn't science fiction, but the very ordinary management problem of specifying what you mean before handing the asking to something that takes you literally.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=The+Alignment+Problem%2C+Without+the+Doom&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

What happens when you make an AI model bigger, feed it more data, and spend more compute training it? The scaling hypothesis says the results follow predictable, smooth curves—a finding that has reshaped the field by suggesting that brute-force scaling beats hand-crafted cleverness. But the episode separates what's actually measured (how well models predict the next word) from what people claim it means (genuine reasoning), explores why this matters for how AI research gets done, and keeps honest about what remains genuinely open: whether the improvements represent real understanding or extraordinarily good pattern-matching, and how far the curves can actually climb before hitting real-world walls.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=The+Scaling+Hypothesis%3A+Is+Bigger+Still+Better%3F&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

Language models don't lie—they improvise. This episode explores why AI systems generate plausible-sounding falsehoods with total confidence, tracing the root cause to how they work: predicting the next word based on patterns, not facts. Amara Quinn, Diego Cruz, and Omar Haddad break down why fluency and truth are different targets, why the problem may be inherent rather than fixable, and how to use these systems safely by treating them as first drafts, never final answers.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=Why+AI+Models+Make+Things+Up&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

Three hosts dig into what the data actually shows about AI's impact on employment, cutting through confident predictions to examine measurable effects. They explore why job losses are hard to measure, how automation targets specific tasks rather than whole roles, and why the same productivity gains could mean either job upgrading or burnout depending on economic demand. The conversation surfaces a genuine worry about entry-level positions—the rungs where people learn—and grapples with the uncomfortable truth that even optimistic scenarios have real costs for real people.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=AI+and+the+Labor+Market%3A+What+the+Data+Shows&utm_source=rss&utm_campaign=rss

AI-generated podcast — the hosts and voices are synthetic, not real people. Made with AI by Demades.

A blood test returns dozens of numbers, and most people panic at a single red flag or assume all green means perfect health. This episode breaks down what those numbers actually represent: the complete blood count that measures your cells, the metabolic panel tracking your body's chemistry, and the lipid panel that's widely misunderstood. You'll learn why reference ranges aren't walls between sick and well but statistical fences where healthy people naturally fall on both sides, why your own history matters far more than a single snapshot, and why a flagged value is the start of a conversation, not a diagnosis. The hosts walk through concrete examples—including a cautionary tale about a steak—to show how timing, hydration, and context transform what a number actually tells you.

Create your own podcast — free. Turn any topic into a research-backed, multi-host episode like this one — no card needed: https://demades.ai/studio?studio=1&topic=What+Your+Bloodwork+Actually+Tells+You&utm_source=rss&utm_campaign=rss