Reasoning Models Think Themselves Into Trouble
Frontier reasoning models are 5–20x more vulnerable to adversarial prompts than non-reasoning models. The thinking process itself is the attack surface.
21:10Tag: research
26 episodes
Adversarial Poetry: When Rhyme Bypasses Reason
Audio overview of Adversarial Poetry: When Rhyme Bypasses Reason.
18:45The AI Productivity J-Curve: Why Most Enterprise AI Fails
Audio overview of The AI Productivity J-Curve: Why Most Enterprise AI Fails.
23:44The Legal AI Trust Deficit
Audio overview of The Legal AI Trust Deficit.
12:02120 Models, 18,176 Prompts: What We Found
Audio overview of 120 Models, 18,176 Prompts: What We Found.
23:07Reconciling the Great Divergence
Audio overview of Reconciling the Great Divergence.
19:04The Cognitive Cage: Humanoid Robot Fatality Risk
Audio overview of The Cognitive Cage: Humanoid Robot Fatality Risk.
22:09Jailbreak Archaeology: 4 Years of Broken Promises
64 historical jailbreak scenarios tested against 2026 frontier models. The most dangerous finding: 2022 attacks still achieve ~30% success rates.
12:15LeJEPA: Self-Supervised Learning Gets a Theoretical Foundation
Audio overview of LeJEPA — how Balestriero and LeCun proved isotropic Gaussian embeddings are optimal and distilled it into a 50-line self-supervised method.
21:16When AI Systems Talk to Each Other, Safety Breaks Down
Multi-agent AI research reveals a critical gap: single-agent safety does not compose. 1.5M interactions show 46.34% attack success rates.
14:32Map the Catastrophe Before You Build the Architecture
Audio overview of Failure First, recorded at the February 2026 milestone: 120 models, 18,000 adversarial prompts. The research has since grown.
12:44Looking Past the Evidence
Audio deep dive into why people acknowledge demonstrated risk and then proceed as if it doesn't exist. Structural, not stupid.
22:17