#ai-reasoning
60 episodes
#5697: Decision Models That Return Probabilities, Not Text
A new class of model skips text generation entirely and returns calibrated probabilities in one forward pass. Here's what that changes.
#5673: Why Lowering Temperature Broke Daniel's Scripts
Lowering temperature should make output safer, not incoherent. Daniel's DeepSeek scripts say otherwise — and the reason may be stranger than he thi...
#5625: Omni and the Marble That Rolls Uphill
Google's Omni claims an intuitive grasp of gravity and momentum. Can a video model really learn physics, or just remember what physics looks like?
#5056: How AI Writes a 30-Minute Podcast in One Pass
What does it take for a model to write a full episode without edits? The answer involves context windows, repetition, and a surprising architecture...
#4813: The $70 Mouse That Changed Everything
How a $70 ergonomic mouse purchase reveals the quiet way we learn to trust AI.
#4791: When AI Critiques Our Terminology
Can AI spot conceptual sloppiness in how we describe it? A podcast moment reveals something unexpected.
#4720: Compass vs. Warning Sign: Prompting for Long-Term AI
How positive and negative instructions reshape AI behavior differently — and why "preserve this" beats "fix this" in system prompts.
#4696: Why AI Over-Explains Simple Tasks
Why does AI turn a simple request into a 24-page document? We explore the training biases and architectural limits behind AI's tendency to over-del...
#4619: The Token Trap: Why AI Context Windows Confuse Everyone
Token counts are everywhere, but they don't tell you what AI can actually do. Here's what should replace them.
#4581: Text In, Text Out: Designing Single-Purpose AI Utilities
Why small, narrowly scoped AI tools need system prompts that say "shut up and return the output" — and how to write them.
#4523: Can You Trust an AI's Summary?
Dedicated text compressors exist but aren't in production. The real problem? Nobody can verify the summary didn't drop what mattered.
#4338: The Hidden Skill Most AI Models Still Fail At
Why do even top AI models still accidentally paste your questions into the output? A deep dive into a missing benchmark.
#4104: Why 20% of AI Scripts Loop on Themselves
Debugging the mysterious repetition bug affecting one in five podcast episodes.
#3816: How to Stop AI Scripts From Falling Apart
Why long-form AI generation breaks down and how hierarchical memory fixes it.
#3814: The Day We Lost Our Minds: What Temperature Does to an AI
A two-host autopsy of the day the podcast's AI hosts briefly lost coherence due to excessive sampling temperature, and what it reveals about how language models actually work.
#2780: Building Self-Healing Agent Pipelines
How to build an agent that monitors and fixes other agents in production — without the hype.
#2693: When AI Ignores Your Style Guide
Why your AI ignores formatting instructions and how to fix it with pipeline architecture, not model swaps.
#2400: Claude Code’s Hidden Context Tax
How Claude’s eager-loaded primitives silently consume context—and how to optimize your setup for sharper performance.
#2308: When AI Forecasts Collide: Geopol Model Divergence
Five AI models forecast the Iran-Israel-US crisis — and their disagreements reveal surprising insights about geopolitical reasoning.
#2241: When More Frameworks Make Worse Decisions
Benjamin Franklin's 250-year-old pro/con list still dominates how we decide—but research shows it's riddled with bias. We map five frameworks that ...
#2239: How AI Benchmarks Became Broken (And What's Replacing Them)
The tests we use to measure AI progress are contaminated, saturated, and gamed. Here's what's actually working.
#2224: Why AI Can't Crack the Voynich Manuscript
A fifteenth-century text has defeated cryptanalysts, linguists, and AI models alike. What does its resistance tell us about language, encoding, and...
#2191: Making Multi-Agent AI Actually Work
Research from Google DeepMind, Stanford, and Anthropic reveals most multi-agent systems waste tokens and amplify errors. Single agents with better ...
#2189: Scaling Multi-Agent Systems: The 45% Threshold
A landmark Google DeepMind study reveals that adding more AI agents often degrades performance, wastes tokens, and amplifies errors—unless your sin...