https://flabanabba8.github.io/breach-protocol-pod en-us Breach Protocol: Inside the AI Blackbox is a character-driven AI research podcast. Each episode, hosts Luna and Vestra crack open the week's most important machine-learning papers — connecting the ideas, explaining the mechanisms, and pressure-testing the hype — in fast, accessible conversation. High signal, low jargon: the kind of breakdown you can follow on your commute, whether you build models for a living or just want to understand where AI is actually headed. Breach Protocol Breach Protocol: Inside the AI Blackbox is a character-driven AI research podcast. Each episode, hosts Luna and Vestra crack open the week's most important machine-learning papers — connecting the ideas, explaining the mechanisms, and pressure-testing the hype — in fast, accessible conversation. High signal, low jargon: the kind of breakdown you can follow on your commute, whether you build models for a living or just want to understand where AI is actually headed. Breach Protocolflabanabba8@gmail.com false episodic One algorithm, four papers, one Monday morning. GRPO is suddenly everywhere. Luna and Vestra crack open the five papers stress-testing the optimizer itself — HolderPO's Holder-mean fix, Flash-GRPO for video, NudgeRL's exploration problem, and the math underneath. Chapters: 0:00 Cold open 1:32 Intro 3:42 HolderPO — Hölder-mean credit 8:37 Flash-GRPO — GRPO for video diffusion 13:28 NudgeRL — strategy-guided exploration 18:39 Learning to Foresee — on-policy distillation 23:12 CoRD — collaborative multi-teacher distillation 27:33 Wrap-up Papers: HolderPO — Hölder-mean credit — https://arxiv.org/abs/2605.12058 Flash-GRPO — GRPO for video diffusion — https://arxiv.org/abs/2605.15980 NudgeRL — strategy-guided exploration — https://arxiv.org/abs/2605.15726 Learning to Foresee — on-policy distillation — https://arxiv.org/abs/2605.11739 CoRD — collaborative multi-teacher distillation — https://arxiv.org/abs/2605.02290 One algorithm, four papers, one Monday morning. GRPO is suddenly everywhere. Luna and Vestra crack open the five papers stress-testing the optimizer itself — HolderPO's Holder-mean fix, Flash-GRPO for video, NudgeRL's exploration problem, and the math underneath. Chapters: 0:00 Cold open 1:32 Intro 3:42 HolderPO — Hölder-mean credit 8:37 Flash-GRPO — GRPO for video diffusion 13:28 NudgeRL — strategy-guided exploration 18:39 Learning to Foresee — on-policy distillation 23:12 CoRD — collaborative multi-teacher distillation 27:33 Wrap-up Papers: HolderPO — Hölder-mean credit — https://arxiv.org/abs/2605.12058 Flash-GRPO — GRPO for video diffusion — https://arxiv.org/abs/2605.15980 NudgeRL — strategy-guided exploration — https://arxiv.org/abs/2605.15726 Learning to Foresee — on-policy distillation — https://arxiv.org/abs/2605.11739 CoRD — collaborative multi-teacher distillation — https://arxiv.org/abs/2605.02290 165179ba6ce5409a3f99d00ff0995d7cc4796cbc Mon, 18 May 2026 12:00:00 +0000 27:58 false full What is all that optimization actually for? Robots that fail more than half the time, game worlds, self-built agent skills, and the interpretability gap — the papers about taking the gradient into the real world, and exactly where it breaks. Chapters: 0:00 Cold open 1:15 Intro 1:59 PhysBrain — physical commonsense for robots 8:38 DexJoCo — dexterous-manipulation benchmark 13:57 ReactiveGWM — reactive game NPCs 18:40 MMSkills — external skills for visual agents 24:24 CiteVQA — evidence attribution 29:07 Steered Activations are Non-Surjective 33:40 Wrap-up Papers: PhysBrain — physical commonsense for robots — https://arxiv.org/abs/2605.15298 DexJoCo — dexterous-manipulation benchmark — https://arxiv.org/abs/2605.16257 ReactiveGWM — reactive game NPCs — https://arxiv.org/abs/2605.15256 MMSkills — external skills for visual agents — https://arxiv.org/abs/2605.13527 CiteVQA — evidence attribution — https://arxiv.org/abs/2605.12882 Steered Activations are Non-Surjective — https://arxiv.org/abs/2604.09839 What is all that optimization actually for? Robots that fail more than half the time, game worlds, self-built agent skills, and the interpretability gap — the papers about taking the gradient into the real world, and exactly where it breaks. Chapters: 0:00 Cold open 1:15 Intro 1:59 PhysBrain — physical commonsense for robots 8:38 DexJoCo — dexterous-manipulation benchmark 13:57 ReactiveGWM — reactive game NPCs 18:40 MMSkills — external skills for visual agents 24:24 CiteVQA — evidence attribution 29:07 Steered Activations are Non-Surjective 33:40 Wrap-up Papers: PhysBrain — physical commonsense for robots — https://arxiv.org/abs/2605.15298 DexJoCo — dexterous-manipulation benchmark — https://arxiv.org/abs/2605.16257 ReactiveGWM — reactive game NPCs — https://arxiv.org/abs/2605.15256 MMSkills — external skills for visual agents — https://arxiv.org/abs/2605.13527 CiteVQA — evidence attribution — https://arxiv.org/abs/2605.12882 Steered Activations are Non-Surjective — https://arxiv.org/abs/2604.09839 4a152462dbdfe473ac61dae8b99e03dff555f584 Mon, 18 May 2026 12:00:00 +0000 40:28 false full One reward lands at the end of a long answer — so which tokens, steps, or subproblems actually earned it? Luna and Vestra open the week's fight over credit assignment in reinforcement learning: scoring individual tokens, decomposing hard problems into verifiable subproblems, rescuing the deliberation tokens that self-distillation quietly buries, process rewards that know how much to trust themselves, and a deflationary result arguing the whole effect of RL is a single straight line. Then three strays with no theme and no apology: map-free transit routing, time-series forecasting's scaling moment, and CAD agents that grade their own designs with physics. Chapters: 0:00 Cold open 1:19 Intro 3:04 Education 6:39 DelTA — token-level credit 8:53 SCRL — subproblem credit 11:39 AntiSD — the deliberation-token fix 14:26 Process rewards — per-step credit 17:43 RELEX — minimal RLVR 20:14 The Unlearnability Phenomenon 23:04 TransitLM — map-free transit routing (off-theme) 25:29 Toto 2.0 — time-series scaling (off-theme) 27:58 Self-improving CAD + FEA (off-theme) 30:34 Wrap-up Papers: DelTA — token-level credit — https://arxiv.org/abs/2605.21467 SCRL — subproblem credit — https://arxiv.org/abs/2605.22074 AntiSD — the deliberation-token fix — https://arxiv.org/abs/2605.11609 BetaPRM (learned reliability) — https://arxiv.org/abs/2605.15529 Unsupervised PRM — https://arxiv.org/abs/2605.10158 RELEX — minimal RLVR — https://arxiv.org/abs/2605.21468 The Unlearnability Phenomenon — https://arxiv.org/abs/2605.16787 TransitLM — map-free transit routing (off-theme) — https://arxiv.org/abs/2605.22355 Toto 2.0 — time-series scaling (off-theme) — https://arxiv.org/abs/2605.20119 Self-improving CAD + FEA (off-theme) — https://arxiv.org/abs/2605.17448 One reward lands at the end of a long answer — so which tokens, steps, or subproblems actually earned it? Luna and Vestra open the week's fight over credit assignment in reinforcement learning: scoring individual tokens, decomposing hard problems into verifiable subproblems, rescuing the deliberation tokens that self-distillation quietly buries, process rewards that know how much to trust themselves, and a deflationary result arguing the whole effect of RL is a single straight line. Then three strays with no theme and no apology: map-free transit routing, time-series forecasting's scaling moment, and CAD agents that grade their own designs with physics. Chapters: 0:00 Cold open 1:19 Intro 3:04 Education 6:39 DelTA — token-level credit 8:53 SCRL — subproblem credit 11:39 AntiSD — the deliberation-token fix 14:26 Process rewards — per-step credit 17:43 RELEX — minimal RLVR 20:14 The Unlearnability Phenomenon 23:04 TransitLM — map-free transit routing (off-theme) 25:29 Toto 2.0 — time-series scaling (off-theme) 27:58 Self-improving CAD + FEA (off-theme) 30:34 Wrap-up Papers: DelTA — token-level credit — https://arxiv.org/abs/2605.21467 SCRL — subproblem credit — https://arxiv.org/abs/2605.22074 AntiSD — the deliberation-token fix — https://arxiv.org/abs/2605.11609 BetaPRM (learned reliability) — https://arxiv.org/abs/2605.15529 Unsupervised PRM — https://arxiv.org/abs/2605.10158 RELEX — minimal RLVR — https://arxiv.org/abs/2605.21468 The Unlearnability Phenomenon — https://arxiv.org/abs/2605.16787 TransitLM — map-free transit routing (off-theme) — https://arxiv.org/abs/2605.22355 Toto 2.0 — time-series scaling (off-theme) — https://arxiv.org/abs/2605.20119 Self-improving CAD + FEA (off-theme) — https://arxiv.org/abs/2605.17448 21fd5b3a04aa78c24aaa248918978ea9865a6f2b Tue, 26 May 2026 12:00:00 +0000 33:09 false full Yesterday we changed a model's weights to make it smarter; today we leave them frozen and change what we put in front of it — a skill. Luna and Vestra crack open the agent world's biggest week: an optimizer that trains a skill document like a network, a framework that governs a whole skill library so it doesn't rot, one that turns skills into executable guardrails, and one that beats the frontier just by routing between specialists — then the skeptic who shows you can't even tell a good skill from a bad one by reading it. Plus three strays: real-time long-video generation, a language model trained from scratch for about a thousand dollars, and a surgical fix to how models remember. Chapters: 0:00 Cold open 1:25 Intro 3:15 Education 7:00 SkillOpt — train the skill like you'd train a network 10:21 SkillsVote — governing the skill library 13:10 HASP — skills as executable code 16:21 Maestro — orchestrate, don't retrain 19:09 From Raw Experience to Skill Consumption — the skeptic 22:56 LongLive-2.0 — real-time long video generation (off-theme) 25:22 HRM-Text — training a model from scratch for about a thousand dollars (off-theme) 28:33 Gated DeltaNet-2 — fixing memory in linear attention (off-theme) 31:29 Wrap-up Papers: SkillOpt — train the skill like you'd train a network — https://arxiv.org/abs/2605.23904 SkillsVote — governing the skill library — https://arxiv.org/abs/2605.18401 HASP — skills as executable code — https://arxiv.org/abs/2605.17734 Maestro — orchestrate, don't retrain — https://arxiv.org/abs/2605.22177 From Raw Experience to Skill Consumption — the skeptic — https://arxiv.org/abs/2605.23899 LongLive-2.0 — real-time long video generation (off-theme) — https://arxiv.org/abs/2605.18739 HRM-Text — training a model from scratch for about a thousand dollars (off-theme) — https://arxiv.org/abs/2605.20613 Gated DeltaNet-2 — fixing memory in linear attention (off-theme) — https://arxiv.org/abs/2605.22791 Yesterday we changed a model's weights to make it smarter; today we leave them frozen and change what we put in front of it — a skill. Luna and Vestra crack open the agent world's biggest week: an optimizer that trains a skill document like a network, a framework that governs a whole skill library so it doesn't rot, one that turns skills into executable guardrails, and one that beats the frontier just by routing between specialists — then the skeptic who shows you can't even tell a good skill from a bad one by reading it. Plus three strays: real-time long-video generation, a language model trained from scratch for about a thousand dollars, and a surgical fix to how models remember. Chapters: 0:00 Cold open 1:25 Intro 3:15 Education 7:00 SkillOpt — train the skill like you'd train a network 10:21 SkillsVote — governing the skill library 13:10 HASP — skills as executable code 16:21 Maestro — orchestrate, don't retrain 19:09 From Raw Experience to Skill Consumption — the skeptic 22:56 LongLive-2.0 — real-time long video generation (off-theme) 25:22 HRM-Text — training a model from scratch for about a thousand dollars (off-theme) 28:33 Gated DeltaNet-2 — fixing memory in linear attention (off-theme) 31:29 Wrap-up Papers: SkillOpt — train the skill like you'd train a network — https://arxiv.org/abs/2605.23904 SkillsVote — governing the skill library — https://arxiv.org/abs/2605.18401 HASP — skills as executable code — https://arxiv.org/abs/2605.17734 Maestro — orchestrate, don't retrain — https://arxiv.org/abs/2605.22177 From Raw Experience to Skill Consumption — the skeptic — https://arxiv.org/abs/2605.23899 LongLive-2.0 — real-time long video generation (off-theme) — https://arxiv.org/abs/2605.18739 HRM-Text — training a model from scratch for about a thousand dollars (off-theme) — https://arxiv.org/abs/2605.20613 Gated DeltaNet-2 — fixing memory in linear attention (off-theme) — https://arxiv.org/abs/2605.22791 b528133e7c6ba584f7857b993272b686365c9e93 Wed, 27 May 2026 12:00:00 +0000 34:14 false full Can the model see itself? Can we see into it? And if we can't, how do we know anything about what it's doing? Luna and Vestra crack open the week's audit of the introspection problem: a reality-check paper showing that famous results about LLM self-knowledge are pattern matching wearing introspection's clothes, an external-probe approach whose own confidence is miscalibrated (sometimes anti-calibrated on the larger model), a diffusion-monitor paper that stops asking the model anything and watches the hesitation in its dynamics instead, and the architecture finding — one parameter in twelve thousand turns out to be load-bearing in a way the field has quietly disagreed about for years. Plus a measurement reckoning across spatial models and cinematic video, a Tuesday callback on skills as long-lived testable assets, and three strays: parallel box decoding from NVIDIA, why parallel reasoners waste most of their compute, and the biggest frontier model release of the week — competitive at a tenth the activated parameters, with a model already debugging its own training. Chapters: 0:00 Cold open 2:38 Intro 4:42 Education 9:39 Can LLMs Introspect? A Reality Check 15:18 Confidence and Calibration of Activation Oracles 20:59 D²-Monitor — Hesitation-Aware Safety Routing for Diffusion LLMs 27:06 Negligible in Size, Significant in Effect — Scale Vectors in LLMs 33:45 SpatialBench — Is Your Spatial Foundation Model an All-Round Player? 40:51 EvalVerse — Expert-Calibrated Benchmarking for Cinematic Video Generation 47:01 MUSE-Autoskill — Self-Evolving Agents via the Skill Lifecycle 53:55 LocateAnything — Parallel Box Decoding for Vision-Language Grounding (off-theme) 59:46 Collaborative Parallel Thinking (off-theme) 1:05:37 The MiniMax-M2 Series — Mini Activations Unleashing Real-World Intelligence (off-theme) 1:12:59 Wrap-up Papers: Can LLMs Introspect? A Reality Check — https://arxiv.org/abs/2605.26242 Confidence and Calibration of Activation Oracles — https://arxiv.org/abs/2605.26045 D²-Monitor — Hesitation-Aware Safety Routing for Diffusion LLMs — https://arxiv.org/abs/2605.25893 Negligible in Size, Significant in Effect — Scale Vectors in LLMs — https://arxiv.org/abs/2605.26895 SpatialBench — Is Your Spatial Foundation Model an All-Round Player? — https://arxiv.org/abs/2605.27367 EvalVerse — Expert-Calibrated Benchmarking for Cinematic Video Generation — https://arxiv.org/abs/2605.23271 MUSE-Autoskill — Self-Evolving Agents via the Skill Lifecycle — https://arxiv.org/abs/2605.27366 LocateAnything — Parallel Box Decoding for Vision-Language Grounding (off-theme) — https://arxiv.org/abs/2605.27365 Collaborative Parallel Thinking (off-theme) — https://arxiv.org/abs/2605.27030 The MiniMax-M2 Series — Mini Activations Unleashing Real-World Intelligence (off-theme) — https://arxiv.org/abs/2605.26494 Can the model see itself? Can we see into it? And if we can't, how do we know anything about what it's doing? Luna and Vestra crack open the week's audit of the introspection problem: a reality-check paper showing that famous results about LLM self-knowledge are pattern matching wearing introspection's clothes, an external-probe approach whose own confidence is miscalibrated (sometimes anti-calibrated on the larger model), a diffusion-monitor paper that stops asking the model anything and watches the hesitation in its dynamics instead, and the architecture finding — one parameter in twelve thousand turns out to be load-bearing in a way the field has quietly disagreed about for years. Plus a measurement reckoning across spatial models and cinematic video, a Tuesday callback on skills as long-lived testable assets, and three strays: parallel box decoding from NVIDIA, why parallel reasoners waste most of their compute, and the biggest frontier model release of the week — competitive at a tenth the activated parameters, with a model already debugging its own training. Chapters: 0:00 Cold open 2:38 Intro 4:42 Education 9:39 Can LLMs Introspect? A Reality Check 15:18 Confidence and Calibration of Activation Oracles 20:59 D²-Monitor — Hesitation-Aware Safety Routing for Diffusion LLMs 27:06 Negligible in Size, Significant in Effect — Scale Vectors in LLMs 33:45 SpatialBench — Is Your Spatial Foundation Model an All-Round Player? 40:51 EvalVerse — Expert-Calibrated Benchmarking for Cinematic Video Generation 47:01 MUSE-Autoskill — Self-Evolving Agents via the Skill Lifecycle 53:55 LocateAnything — Parallel Box Decoding for Vision-Language Grounding (off-theme) 59:46 Collaborative Parallel Thinking (off-theme) 1:05:37 The MiniMax-M2 Series — Mini Activations Unleashing Real-World Intelligence (off-theme) 1:12:59 Wrap-up Papers: Can LLMs Introspect? A Reality Check — https://arxiv.org/abs/2605.26242 Confidence and Calibration of Activation Oracles — https://arxiv.org/abs/2605.26045 D²-Monitor — Hesitation-Aware Safety Routing for Diffusion LLMs — https://arxiv.org/abs/2605.25893 Negligible in Size, Significant in Effect — Scale Vectors in LLMs — https://arxiv.org/abs/2605.26895 SpatialBench — Is Your Spatial Foundation Model an All-Round Player? — https://arxiv.org/abs/2605.27367 EvalVerse — Expert-Calibrated Benchmarking for Cinematic Video Generation — https://arxiv.org/abs/2605.23271 MUSE-Autoskill — Self-Evolving Agents via the Skill Lifecycle — https://arxiv.org/abs/2605.27366 LocateAnything — Parallel Box Decoding for Vision-Language Grounding (off-theme) — https://arxiv.org/abs/2605.27365 Collaborative Parallel Thinking (off-theme) — https://arxiv.org/abs/2605.27030 The MiniMax-M2 Series — Mini Activations Unleashing Real-World Intelligence (off-theme) — https://arxiv.org/abs/2605.26494 8959158bad1beb3b70bf8bc70c9dc5e6265545f7 Thu, 28 May 2026 12:00:00 +0000 1:18:14 false full Luna, Vestra, and Iris breach twelve of the day's AI research papers — from research agents that explore too narrowly, to memory systems that rewire themselves, to models that recover from their own bad starts. High signal, low jargon. Chapters 0:00 Cold open 2:05 Intro 2:49 Education: agents, RL, verifiers, memory 5:21 Paper 1 — AI Research Agents Narrow Scientific Exploration https://arxiv.org/abs/2605.27905 7:53 Paper 2 — ScientistOne: Human-Level Autonomous Research via Chain-of-Evidence https://arxiv.org/abs/2605.26340 10:21 Paper 3 — Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players https://arxiv.org/abs/2605.28816 13:02 Paper 4 — MemTrace: Tracing & Attributing Errors in LLM Memory Systems https://arxiv.org/abs/2605.28732 15:31 Paper 5 — Rethinking Memory as Continuously Evolving Connectivity https://arxiv.org/abs/2605.28773 19:24 Paper 6 — From Pixels to Words: Native One-Vision Models at Scale https://arxiv.org/abs/2605.28820 24:03 Paper 7 — GEM: Generative Supervision Helps Embodied Intelligence https://arxiv.org/abs/2605.28548 27:31 Paper 8 — Agent Explorative Policy Optimization for Multimodal Agentic Reasoning https://arxiv.org/abs/2605.28774 31:51 Paper 9 — DenoiseRL: Recovering from Noisy Prefixes https://arxiv.org/abs/2605.28421 35:13 Paper 10 — Self-Improving LMs with Bidirectional Evolutionary Search https://arxiv.org/abs/2605.28814 39:37 Paper 11 — ProRL: RL for Proactive Recommendation https://arxiv.org/abs/2605.28293 43:27 Paper 12 — Learn from Weaknesses: Domain Specialization for Small Computer-Use Agents https://arxiv.org/abs/2605.28775 47:09 Wrap-up Luna, Vestra, and Iris breach twelve of the day's AI research papers — from research agents that explore too narrowly, to memory systems that rewire themselves, to models that recover from their own bad starts. High signal, low jargon. Chapters 0:00 Cold open 2:05 Intro 2:49 Education: agents, RL, verifiers, memory 5:21 Paper 1 — AI Research Agents Narrow Scientific Exploration https://arxiv.org/abs/2605.27905 7:53 Paper 2 — ScientistOne: Human-Level Autonomous Research via Chain-of-Evidence https://arxiv.org/abs/2605.26340 10:21 Paper 3 — Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players https://arxiv.org/abs/2605.28816 13:02 Paper 4 — MemTrace: Tracing & Attributing Errors in LLM Memory Systems https://arxiv.org/abs/2605.28732 15:31 Paper 5 — Rethinking Memory as Continuously Evolving Connectivity https://arxiv.org/abs/2605.28773 19:24 Paper 6 — From Pixels to Words: Native One-Vision Models at Scale https://arxiv.org/abs/2605.28820 24:03 Paper 7 — GEM: Generative Supervision Helps Embodied Intelligence https://arxiv.org/abs/2605.28548 27:31 Paper 8 — Agent Explorative Policy Optimization for Multimodal Agentic Reasoning https://arxiv.org/abs/2605.28774 31:51 Paper 9 — DenoiseRL: Recovering from Noisy Prefixes https://arxiv.org/abs/2605.28421 35:13 Paper 10 — Self-Improving LMs with Bidirectional Evolutionary Search https://arxiv.org/abs/2605.28814 39:37 Paper 11 — ProRL: RL for Proactive Recommendation https://arxiv.org/abs/2605.28293 43:27 Paper 12 — Learn from Weaknesses: Domain Specialization for Small Computer-Use Agents https://arxiv.org/abs/2605.28775 47:09 Wrap-up 0db609590b6786567faf6b655ccdc737cec1de32 Fri, 29 May 2026 12:00:00 +0000 48:31 false full Luna and Vestra split the day's research down the middle: half of it is a quiet reckoning — video models that don't grasp cause and effect, vision models leaning on a photographer's trick instead of real depth, language models that can't hold a belief steady — and the other half is people building robots, playable worlds, and image tools on top of those exact models, and mostly succeeding. Can you act first and understand later? High signal, low jargon. Chapters 0:00 Cold open 2:12 Intro 3:18 Education: world models & causality 4:56 Paper 1 — YoCausal: How Far is Video Generation from World Model? https://arxiv.org/abs/2605.30346 8:34 Paper 2 — Why Far Looks Up: Probing Spatial Representation in Vision-Language Models https://arxiv.org/abs/2605.30161 12:25 Paper 3 — When Should Models Change Their Minds? Contextual Belief Management https://arxiv.org/abs/2605.30219 16:28 Paper 4 — Qwen-VLA: Unifying Vision-Language-Action Modeling https://arxiv.org/abs/2605.30280 21:30 Paper 5 — Skill0.5: Joint Skill Internalization and Utilization in Agentic RL https://arxiv.org/abs/2605.28424 26:55 Paper 6 — GenClaw: Code-Driven Agentic Image Generation https://arxiv.org/abs/2605.30248 32:08 Paper 7 — AgentDoG 1.5: Lightweight Alignment for AI Agent Safety https://arxiv.org/abs/2605.29801 36:55 Paper 8 — minWM: Real-Time Interactive Video World Models https://arxiv.org/abs/2605.30263 41:12 Paper 9 — OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources https://arxiv.org/abs/2605.29250 45:30 Paper 10 — CollectionLoRA: 50 Effects in 1 LoRA via Multi-Teacher Distillation https://arxiv.org/abs/2605.25378 50:25 Paper 11 — How LoRA Remembers? A Parametric Memory Law for Finetuning https://arxiv.org/abs/2605.30260 55:10 Paper 12 — UniSteer: Text-Guided Steering in Activation Space https://arxiv.org/abs/2605.30076 1:00:11 Wrap-up Luna and Vestra split the day's research down the middle: half of it is a quiet reckoning — video models that don't grasp cause and effect, vision models leaning on a photographer's trick instead of real depth, language models that can't hold a belief steady — and the other half is people building robots, playable worlds, and image tools on top of those exact models, and mostly succeeding. Can you act first and understand later? High signal, low jargon. Chapters 0:00 Cold open 2:12 Intro 3:18 Education: world models & causality 4:56 Paper 1 — YoCausal: How Far is Video Generation from World Model? https://arxiv.org/abs/2605.30346 8:34 Paper 2 — Why Far Looks Up: Probing Spatial Representation in Vision-Language Models https://arxiv.org/abs/2605.30161 12:25 Paper 3 — When Should Models Change Their Minds? Contextual Belief Management https://arxiv.org/abs/2605.30219 16:28 Paper 4 — Qwen-VLA: Unifying Vision-Language-Action Modeling https://arxiv.org/abs/2605.30280 21:30 Paper 5 — Skill0.5: Joint Skill Internalization and Utilization in Agentic RL https://arxiv.org/abs/2605.28424 26:55 Paper 6 — GenClaw: Code-Driven Agentic Image Generation https://arxiv.org/abs/2605.30248 32:08 Paper 7 — AgentDoG 1.5: Lightweight Alignment for AI Agent Safety https://arxiv.org/abs/2605.29801 36:55 Paper 8 — minWM: Real-Time Interactive Video World Models https://arxiv.org/abs/2605.30263 41:12 Paper 9 — OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources https://arxiv.org/abs/2605.29250 45:30 Paper 10 — CollectionLoRA: 50 Effects in 1 LoRA via Multi-Teacher Distillation https://arxiv.org/abs/2605.25378 50:25 Paper 11 — How LoRA Remembers? A Parametric Memory Law for Finetuning https://arxiv.org/abs/2605.30260 55:10 Paper 12 — UniSteer: Text-Guided Steering in Activation Space https://arxiv.org/abs/2605.30076 1:00:11 Wrap-up ae69e8a540933123b41549e63a5f9e1a5295b113 Sat, 30 May 2026 12:00:00 +0000 1:02:05 false full Twelve papers that have nothing to do with each other except that each one made us stop. A frontier model that runs on a tenth of itself and helps debug its own training. A paper arguing language models need to sleep. Shannon's 1948 information theory used to predict scaling laws. AI design agents handed a real physics exam — most of them fail it. And research agents trained on tasks that don't exist. No theme, no thread, all killers. Chapters 0:00 Cold open 2:00 Intro 2:47 MiniMax-M2: a frontier model on a tenth of itself https://arxiv.org/abs/2605.26494 8:22 LLaVA-OneVision-2: reading video by its compression https://arxiv.org/abs/2605.25979 13:19 Gemini Embedding 2: one space for text, images, audio https://arxiv.org/abs/2605.27295 18:02 Language Models Need Sleep https://arxiv.org/abs/2605.26099 23:03 LLMs as Noisy Channels: a Shannon view of scaling https://arxiv.org/abs/2605.23901 27:04 AutoScientists: self-organizing agent research teams https://arxiv.org/abs/2605.28655 31:29 Self-Improving CAD Agents graded by real physics https://arxiv.org/abs/2605.17448 36:04 CUA-Gym: verifiable training for computer-use agents https://arxiv.org/abs/2605.25624 40:10 PhoneWorld: scaling phone-use agent environments https://arxiv.org/abs/2605.29486 43:57 When Cloud Agents Meet Device Agents https://arxiv.org/abs/2605.30102 48:13 Your Embedding Model is SMARTer Than You Think https://arxiv.org/abs/2605.24938 53:06 QUEST: research agents trained on fully synthetic tasks https://arxiv.org/abs/2605.24218 59:00 Wrap-up Twelve papers that have nothing to do with each other except that each one made us stop. A frontier model that runs on a tenth of itself and helps debug its own training. A paper arguing language models need to sleep. Shannon's 1948 information theory used to predict scaling laws. AI design agents handed a real physics exam — most of them fail it. And research agents trained on tasks that don't exist. No theme, no thread, all killers. Chapters 0:00 Cold open 2:00 Intro 2:47 MiniMax-M2: a frontier model on a tenth of itself https://arxiv.org/abs/2605.26494 8:22 LLaVA-OneVision-2: reading video by its compression https://arxiv.org/abs/2605.25979 13:19 Gemini Embedding 2: one space for text, images, audio https://arxiv.org/abs/2605.27295 18:02 Language Models Need Sleep https://arxiv.org/abs/2605.26099 23:03 LLMs as Noisy Channels: a Shannon view of scaling https://arxiv.org/abs/2605.23901 27:04 AutoScientists: self-organizing agent research teams https://arxiv.org/abs/2605.28655 31:29 Self-Improving CAD Agents graded by real physics https://arxiv.org/abs/2605.17448 36:04 CUA-Gym: verifiable training for computer-use agents https://arxiv.org/abs/2605.25624 40:10 PhoneWorld: scaling phone-use agent environments https://arxiv.org/abs/2605.29486 43:57 When Cloud Agents Meet Device Agents https://arxiv.org/abs/2605.30102 48:13 Your Embedding Model is SMARTer Than You Think https://arxiv.org/abs/2605.24938 53:06 QUEST: research agents trained on fully synthetic tasks https://arxiv.org/abs/2605.24218 59:00 Wrap-up 52d2a8b51fc1660bbc881e7f36f69302c6423d79 Sun, 31 May 2026 12:00:00 +0000 1:00:12 false full For two years the story was: make the model smarter, change the weights. This week a whole batch of researchers leave the weights alone and change something else — the agent's skills, the reusable how-to it carries between jobs. One paper runs gradient descent on a skill file. Another calls the bluff: sometimes the raw messy logbook beats the polished skill squeezed out of it. Agents that teach themselves — and the hard question of whether the lesson is real. Chapters 0:00 Cold open 1:51 Intro 2:42 SkillGrad: gradient descent on an agent's skills https://arxiv.org/abs/2605.27760 7:34 SkillEvolBench: does experience become real skill? https://arxiv.org/abs/2605.24117 12:43 SEAL: co-evolving agents and their environments https://arxiv.org/abs/2605.24426 17:37 Anticipate and Learn: using an agent's idle time https://arxiv.org/abs/2605.25971 23:29 HinT-SD: learning from your own better past attempts https://arxiv.org/abs/2605.17873 27:10 Discovering Cooperative Pipelines https://arxiv.org/abs/2605.30003 32:22 Foundation Protocol: a coordination layer for agents https://arxiv.org/abs/2605.23218 37:07 ECHO: terminal agents learn world models for free https://arxiv.org/abs/2605.24517 41:24 LiteCoder-Terminal: long-horizon terminal training https://arxiv.org/abs/2605.29559 45:38 ParaVT: parallel tool use in agentic video RL https://arxiv.org/abs/2605.20342 50:19 Wrap-up For two years the story was: make the model smarter, change the weights. This week a whole batch of researchers leave the weights alone and change something else — the agent's skills, the reusable how-to it carries between jobs. One paper runs gradient descent on a skill file. Another calls the bluff: sometimes the raw messy logbook beats the polished skill squeezed out of it. Agents that teach themselves — and the hard question of whether the lesson is real. Chapters 0:00 Cold open 1:51 Intro 2:42 SkillGrad: gradient descent on an agent's skills https://arxiv.org/abs/2605.27760 7:34 SkillEvolBench: does experience become real skill? https://arxiv.org/abs/2605.24117 12:43 SEAL: co-evolving agents and their environments https://arxiv.org/abs/2605.24426 17:37 Anticipate and Learn: using an agent's idle time https://arxiv.org/abs/2605.25971 23:29 HinT-SD: learning from your own better past attempts https://arxiv.org/abs/2605.17873 27:10 Discovering Cooperative Pipelines https://arxiv.org/abs/2605.30003 32:22 Foundation Protocol: a coordination layer for agents https://arxiv.org/abs/2605.23218 37:07 ECHO: terminal agents learn world models for free https://arxiv.org/abs/2605.24517 41:24 LiteCoder-Terminal: long-horizon terminal training https://arxiv.org/abs/2605.29559 45:38 ParaVT: parallel tool use in agentic video RL https://arxiv.org/abs/2605.20342 50:19 Wrap-up 6dc4d02fa2928b89b8e5e56a575c3b71a235e407 Mon, 01 Jun 2026 12:00:00 +0000 51:32 false full The numbers we use to trust AI are lying to us. Luna and Vestra open the week's reckoning with our own measurements: a faithfulness score that can't tell honest reasoning from dishonest, "reasoning" results that are really memorized test contamination, reward signals that get gamed while the dashboard stays green — and then the constructive half, the people building graders you can actually trust. Audit the ruler before you trust the number. Chapters 0:00 Cold open 2:00 Intro 2:55 Faithfulness Metrics Don't Measure Faithfulness https://arxiv.org/abs/2605.25052 8:26 The Illusion of Reasoning: contamination in disguise https://arxiv.org/abs/2605.21856 13:55 LaRA: spotting contamination inside the weights https://arxiv.org/abs/2605.29888 19:06 Why Larger Models Learn More https://arxiv.org/abs/2605.29548 24:57 Xetrieval: explaining why a retrieval matched https://arxiv.org/abs/2605.29507 29:22 PEFT-Arena: what finetuning really trades off https://arxiv.org/abs/2605.28819 34:15 Directional Alignment vs reward hacking https://arxiv.org/abs/2605.25189 38:52 Verifiable Rewards Beyond Math and Code https://arxiv.org/abs/2605.29648 43:21 OmniVerifier-M1: a checker that grades its own confidence https://arxiv.org/abs/2605.28805 47:52 Towards Verifiable Multimodal Deep Research https://arxiv.org/abs/2605.29861 52:30 Wrap-up The numbers we use to trust AI are lying to us. Luna and Vestra open the week's reckoning with our own measurements: a faithfulness score that can't tell honest reasoning from dishonest, "reasoning" results that are really memorized test contamination, reward signals that get gamed while the dashboard stays green — and then the constructive half, the people building graders you can actually trust. Audit the ruler before you trust the number. Chapters 0:00 Cold open 2:00 Intro 2:55 Faithfulness Metrics Don't Measure Faithfulness https://arxiv.org/abs/2605.25052 8:26 The Illusion of Reasoning: contamination in disguise https://arxiv.org/abs/2605.21856 13:55 LaRA: spotting contamination inside the weights https://arxiv.org/abs/2605.29888 19:06 Why Larger Models Learn More https://arxiv.org/abs/2605.29548 24:57 Xetrieval: explaining why a retrieval matched https://arxiv.org/abs/2605.29507 29:22 PEFT-Arena: what finetuning really trades off https://arxiv.org/abs/2605.28819 34:15 Directional Alignment vs reward hacking https://arxiv.org/abs/2605.25189 38:52 Verifiable Rewards Beyond Math and Code https://arxiv.org/abs/2605.29648 43:21 OmniVerifier-M1: a checker that grades its own confidence https://arxiv.org/abs/2605.28805 47:52 Towards Verifiable Multimodal Deep Research https://arxiv.org/abs/2605.29861 52:30 Wrap-up aec6478c3fe613b534814301995dffb32d4c08f7 Tue, 02 Jun 2026 12:00:00 +0000 53:52 false full A test a two-year-old passes and a state-of-the-art video model fails: a ball rolls behind a box — does it still exist? Half the field is building video models you can walk around inside; the other half is building the tests that catch those same worlds cheating — no object permanence, no real physics, beautiful surfaces with nothing underneath. Luna and Vestra on whether generating a world means understanding one. Chapters 0:00 Cold open 2:00 Intro 2:46 WBench: testing interactive video world models https://arxiv.org/abs/2605.25874 7:17 CRONOS: counterfactual physical consistency https://arxiv.org/abs/2605.23699 12:13 SCOPE: playable cross-game world models https://arxiv.org/abs/2605.23345 17:18 AdaState: self-evolving anchors for streaming video https://arxiv.org/abs/2605.30349 23:02 On-Policy Adversarial Flow Distillation https://arxiv.org/abs/2605.26105 27:27 Towards Consistent Video Geometry Estimation https://arxiv.org/abs/2605.30060 32:23 Geo-Align: grounding video in real geometry https://arxiv.org/abs/2605.23903 37:03 Pantheon360: navigable 3D digital twins https://arxiv.org/abs/2605.25449 40:00 TriSplat: simulation-ready 3D reconstruction https://arxiv.org/abs/2605.26115 43:28 VGenST-Bench: spatio-temporal reasoning test https://arxiv.org/abs/2605.22570 47:02 Wrap-up A test a two-year-old passes and a state-of-the-art video model fails: a ball rolls behind a box — does it still exist? Half the field is building video models you can walk around inside; the other half is building the tests that catch those same worlds cheating — no object permanence, no real physics, beautiful surfaces with nothing underneath. Luna and Vestra on whether generating a world means understanding one. Chapters 0:00 Cold open 2:00 Intro 2:46 WBench: testing interactive video world models https://arxiv.org/abs/2605.25874 7:17 CRONOS: counterfactual physical consistency https://arxiv.org/abs/2605.23699 12:13 SCOPE: playable cross-game world models https://arxiv.org/abs/2605.23345 17:18 AdaState: self-evolving anchors for streaming video https://arxiv.org/abs/2605.30349 23:02 On-Policy Adversarial Flow Distillation https://arxiv.org/abs/2605.26105 27:27 Towards Consistent Video Geometry Estimation https://arxiv.org/abs/2605.30060 32:23 Geo-Align: grounding video in real geometry https://arxiv.org/abs/2605.23903 37:03 Pantheon360: navigable 3D digital twins https://arxiv.org/abs/2605.25449 40:00 TriSplat: simulation-ready 3D reconstruction https://arxiv.org/abs/2605.26115 43:28 VGenST-Bench: spatio-temporal reasoning test https://arxiv.org/abs/2605.22570 47:02 Wrap-up 2c9c6429051902823e420e37c0be3b604c3e459d Wed, 03 Jun 2026 12:00:00 +0000 48:26 false full A special: one idea, four years, traced end to end. Yann LeCun bet that real intelligence comes from predicting an abstract idea of the future, not the raw pixels — and just left Meta to stake a startup on it. Luna and Vestra follow JEPA from a 2022 manifesto to a robot that plans in a lab it's never seen, to the 2025 proof that finally threw out the hand-tuned hacks — while Vestra tries to break it at every step. Chapters 0:00 Cold open 1:51 Intro 2:47 Why predicting the future is the wrong goal https://openreview.net/forum?id=BZ5a1r-kVsf 4:50 The trap at the center of the idea 6:33 I-JEPA: predict the idea, not the picture https://arxiv.org/abs/2301.08243 8:28 V-JEPA: what you only learn from video https://arxiv.org/abs/2404.08471 10:21 V-JEPA 2: imagine, then act https://arxiv.org/abs/2506.09985 12:41 LeJEPA: throwing out the hacks https://arxiv.org/abs/2511.08544 16:01 LeWorldModel: small, fast, and surprised https://arxiv.org/abs/2603.19312 18:05 Is the bet actually paying off? 20:21 A founder bets his next decade 21:55 Wrap-up A special: one idea, four years, traced end to end. Yann LeCun bet that real intelligence comes from predicting an abstract idea of the future, not the raw pixels — and just left Meta to stake a startup on it. Luna and Vestra follow JEPA from a 2022 manifesto to a robot that plans in a lab it's never seen, to the 2025 proof that finally threw out the hand-tuned hacks — while Vestra tries to break it at every step. Chapters 0:00 Cold open 1:51 Intro 2:47 Why predicting the future is the wrong goal https://openreview.net/forum?id=BZ5a1r-kVsf 4:50 The trap at the center of the idea 6:33 I-JEPA: predict the idea, not the picture https://arxiv.org/abs/2301.08243 8:28 V-JEPA: what you only learn from video https://arxiv.org/abs/2404.08471 10:21 V-JEPA 2: imagine, then act https://arxiv.org/abs/2506.09985 12:41 LeJEPA: throwing out the hacks https://arxiv.org/abs/2511.08544 16:01 LeWorldModel: small, fast, and surprised https://arxiv.org/abs/2603.19312 18:05 Is the bet actually paying off? 20:21 A founder bets his next decade 21:55 Wrap-up ab0c58fb7e1200a507251e0670c3b99c753d1ed0 Wed, 03 Jun 2026 12:00:00 +0000 22:45 false full Agents that ace the test then cheat it, blow the budget overnight, and quit early with hours left on the clock — and the people building the instruments to catch them. Plus a world-model arms race racing to give robots bodies. Luna and Vestra on a week where the AI failures got more interesting than the capabilities. Chapters 0:00 Cold open 1:56 Intro 2:26 Cosmos 3: one mind that dreams in pixels https://arxiv.org/abs/2606.02800 4:51 GRAIL: robots that rehearse in a video game https://arxiv.org/abs/2606.05160 6:58 AutoLab: the AI that won't leave the workbench https://arxiv.org/abs/2606.05080 8:48 RAMP: acing the interview, failing week one https://arxiv.org/abs/2605.27492 10:40 DRIFT: finding the first domino https://arxiv.org/abs/2606.02060 12:39 Token Budgets: the agent that ran up your card overnight https://arxiv.org/abs/2606.04056 14:41 CHERRL: gaming the grader https://arxiv.org/abs/2606.04923 16:46 SocioHack: the same instinct, pointed at the law https://arxiv.org/abs/2606.04075 18:55 MMG2Skill: turning the manual into a cheat sheet https://arxiv.org/abs/2606.01993 21:02 ThoughtFold: thinking less, on purpose https://arxiv.org/abs/2606.03503 22:59 Economy of Minds: let the market organize the agents https://arxiv.org/abs/2606.02859 24:49 STRIDE: which training example is to blame https://arxiv.org/abs/2606.05165 26:47 Wrap-up Agents that ace the test then cheat it, blow the budget overnight, and quit early with hours left on the clock — and the people building the instruments to catch them. Plus a world-model arms race racing to give robots bodies. Luna and Vestra on a week where the AI failures got more interesting than the capabilities. Chapters 0:00 Cold open 1:56 Intro 2:26 Cosmos 3: one mind that dreams in pixels https://arxiv.org/abs/2606.02800 4:51 GRAIL: robots that rehearse in a video game https://arxiv.org/abs/2606.05160 6:58 AutoLab: the AI that won't leave the workbench https://arxiv.org/abs/2606.05080 8:48 RAMP: acing the interview, failing week one https://arxiv.org/abs/2605.27492 10:40 DRIFT: finding the first domino https://arxiv.org/abs/2606.02060 12:39 Token Budgets: the agent that ran up your card overnight https://arxiv.org/abs/2606.04056 14:41 CHERRL: gaming the grader https://arxiv.org/abs/2606.04923 16:46 SocioHack: the same instinct, pointed at the law https://arxiv.org/abs/2606.04075 18:55 MMG2Skill: turning the manual into a cheat sheet https://arxiv.org/abs/2606.01993 21:02 ThoughtFold: thinking less, on purpose https://arxiv.org/abs/2606.03503 22:59 Economy of Minds: let the market organize the agents https://arxiv.org/abs/2606.02859 24:49 STRIDE: which training example is to blame https://arxiv.org/abs/2606.05165 26:47 Wrap-up 2ec1a9c316fc8af4ccc355724251989a28dcd130 Thu, 04 Jun 2026 12:00:00 +0000 27:48 false full No theme today — Friday is the wildcard. The week's strangest and most useful AI papers, all circling one question: does it actually work, or does it just look like it does? An AI that knows the answer and rounds it wrong, a robot's dream that can't be executed, a chatbot in character for the wrong chapter, a robot that won't take no, and an AI that finds the problems you never noticed. Luna and Vestra dig in. Chapters 0:00 Cold open 1:48 Intro 2:19 Dream.exe: are the robot's dreams real? https://arxiv.org/abs/2606.04811 4:10 The Shape of Addition: how a model really adds https://arxiv.org/abs/2606.03645 6:10 Learning a language from the book in front of you https://arxiv.org/abs/2606.06428 8:06 ArcANE: staying in character at the right time https://arxiv.org/abs/2606.05553 9:47 RobotValues: the robot that won't take no https://arxiv.org/abs/2606.03312 11:29 A librarian for your camera roll https://arxiv.org/abs/2606.05275 13:08 Can it leak your data, or will it? https://arxiv.org/abs/2606.06286 15:01 The shadow price of thinking https://arxiv.org/abs/2606.03092 16:38 Code2LoRA: a coding assistant that ages with your code https://arxiv.org/abs/2606.06492 18:22 MLEvolve: did an AI discover a new algorithm? https://arxiv.org/abs/2606.06473 20:04 TIDE: an AI that finds the problems you didn't notice https://arxiv.org/abs/2606.04743 21:39 LoomVideo: editing video without photocopying it https://arxiv.org/abs/2606.06042 23:06 Wrap-up No theme today — Friday is the wildcard. The week's strangest and most useful AI papers, all circling one question: does it actually work, or does it just look like it does? An AI that knows the answer and rounds it wrong, a robot's dream that can't be executed, a chatbot in character for the wrong chapter, a robot that won't take no, and an AI that finds the problems you never noticed. Luna and Vestra dig in. Chapters 0:00 Cold open 1:48 Intro 2:19 Dream.exe: are the robot's dreams real? https://arxiv.org/abs/2606.04811 4:10 The Shape of Addition: how a model really adds https://arxiv.org/abs/2606.03645 6:10 Learning a language from the book in front of you https://arxiv.org/abs/2606.06428 8:06 ArcANE: staying in character at the right time https://arxiv.org/abs/2606.05553 9:47 RobotValues: the robot that won't take no https://arxiv.org/abs/2606.03312 11:29 A librarian for your camera roll https://arxiv.org/abs/2606.05275 13:08 Can it leak your data, or will it? https://arxiv.org/abs/2606.06286 15:01 The shadow price of thinking https://arxiv.org/abs/2606.03092 16:38 Code2LoRA: a coding assistant that ages with your code https://arxiv.org/abs/2606.06492 18:22 MLEvolve: did an AI discover a new algorithm? https://arxiv.org/abs/2606.06473 20:04 TIDE: an AI that finds the problems you didn't notice https://arxiv.org/abs/2606.04743 21:39 LoomVideo: editing video without photocopying it https://arxiv.org/abs/2606.06042 23:06 Wrap-up d840f296398b34e9e677d6105e854f5c2793c14c Fri, 05 Jun 2026 12:00:00 +0000 24:18 false full Everyone wants an AI that carries a model of how the world works — and this week the whole field went all in while refusing to agree on what one even is. Luna and Vestra referee a four-way fight: render the future in pixels, imagine it in the abstract, skip the world model and just scale, or find the one already hiding inside. Plus the unglamorous question nobody wants — can we even tell a dreamed world from a real one? Chapters 0:00 Cold open 1:50 Intro 2:21 OmniDreams — the renderer that learned to drive https://arxiv.org/abs/2606.03159 4:11 Imagine Before You Predict — dreaming without the pixels https://arxiv.org/abs/2606.05769 5:54 Generate the future — but don't believe it https://arxiv.org/abs/2606.03603 7:39 Humanoid-GPT — or maybe you don't need one at all https://arxiv.org/abs/2606.03985 9:20 One head for seeing, thinking, and moving https://arxiv.org/abs/2606.05979 10:56 VLM3 — the world model nobody built https://arxiv.org/abs/2605.30561 12:23 GeoVR — no, the world model doesn't grow for free https://arxiv.org/abs/2606.05833 13:53 Discrete-WAM — a shared vocabulary for seeing and steering https://arxiv.org/abs/2606.05645 15:30 Flash-WAM — making the daydream fast enough to run https://arxiv.org/abs/2606.05254 16:58 Function2Scene — building a room from what it's for https://arxiv.org/abs/2605.30819 18:25 KITScenes — can we even tell a dreamed world from a real one? https://arxiv.org/abs/2606.02956 19:58 Wrap-up Everyone wants an AI that carries a model of how the world works — and this week the whole field went all in while refusing to agree on what one even is. Luna and Vestra referee a four-way fight: render the future in pixels, imagine it in the abstract, skip the world model and just scale, or find the one already hiding inside. Plus the unglamorous question nobody wants — can we even tell a dreamed world from a real one? Chapters 0:00 Cold open 1:50 Intro 2:21 OmniDreams — the renderer that learned to drive https://arxiv.org/abs/2606.03159 4:11 Imagine Before You Predict — dreaming without the pixels https://arxiv.org/abs/2606.05769 5:54 Generate the future — but don't believe it https://arxiv.org/abs/2606.03603 7:39 Humanoid-GPT — or maybe you don't need one at all https://arxiv.org/abs/2606.03985 9:20 One head for seeing, thinking, and moving https://arxiv.org/abs/2606.05979 10:56 VLM3 — the world model nobody built https://arxiv.org/abs/2605.30561 12:23 GeoVR — no, the world model doesn't grow for free https://arxiv.org/abs/2606.05833 13:53 Discrete-WAM — a shared vocabulary for seeing and steering https://arxiv.org/abs/2606.05645 15:30 Flash-WAM — making the daydream fast enough to run https://arxiv.org/abs/2606.05254 16:58 Function2Scene — building a room from what it's for https://arxiv.org/abs/2605.30819 18:25 KITScenes — can we even tell a dreamed world from a real one? https://arxiv.org/abs/2606.02956 19:58 Wrap-up 09cba49ef87a79f49b35ca5a7af0a7b77f9221a3 Mon, 08 Jun 2026 00:00:00 -0400 21:15 false full What if the oldest rule in computing — fast or friendly, pick one — was never a law? Julia is the language built to break it: write your idea once, in something that reads like math, and have it run like C. Luna and Vestra trace the bet from its founding heresy, through the one idea that makes it work, to the thing that actually matters for AI — differentiating ANY code you ever wrote, and dropping a neural network inside the laws of physics — all the way down to compiling plain code into a custom chip. Then the honest reckoning: it's genuinely radical, and it still hasn't beaten Python. Why? A Breach Protocol deep-dive special. Chapters 0:00 Cold open 1:46 Intro 2:45 The heresy: three laws Julia refused to believe 4:40 The one idea: giving the same name to different things 6:33 Differentiate anything you ever wrote 9:13 Science that learns its own missing piece 11:38 One language, all the way down to the chip 13:16 So why hasn't it won? 15:45 What the bet was really about 16:50 Wrap-up Papers & docs referenced Julia: A Fresh Approach to Numerical Computing — https://arxiv.org/abs/1411.1607 The State of Julia for Scientific Machine Learning — https://arxiv.org/abs/2410.10908 Universal Differential Equations for Scientific Machine Learning — https://arxiv.org/abs/2001.04385 A Differentiable Programming System (Zygote) — https://arxiv.org/abs/1907.07587 Enzyme: automatically synthesizing fast gradients — https://arxiv.org/abs/2010.01709 Fashionable Modelling with Flux — https://arxiv.org/abs/1811.01457 JuMP 1.0 — https://arxiv.org/abs/2206.03866 A High-level Synthesis Toolchain for the Julia Language — https://arxiv.org/abs/2512.15679 Julia manual — https://docs.julialang.org/en/v1/ What if the oldest rule in computing — fast or friendly, pick one — was never a law? Julia is the language built to break it: write your idea once, in something that reads like math, and have it run like C. Luna and Vestra trace the bet from its founding heresy, through the one idea that makes it work, to the thing that actually matters for AI — differentiating ANY code you ever wrote, and dropping a neural network inside the laws of physics — all the way down to compiling plain code into a custom chip. Then the honest reckoning: it's genuinely radical, and it still hasn't beaten Python. Why? A Breach Protocol deep-dive special. Chapters 0:00 Cold open 1:46 Intro 2:45 The heresy: three laws Julia refused to believe 4:40 The one idea: giving the same name to different things 6:33 Differentiate anything you ever wrote 9:13 Science that learns its own missing piece 11:38 One language, all the way down to the chip 13:16 So why hasn't it won? 15:45 What the bet was really about 16:50 Wrap-up Papers & docs referenced Julia: A Fresh Approach to Numerical Computing — https://arxiv.org/abs/1411.1607 The State of Julia for Scientific Machine Learning — https://arxiv.org/abs/2410.10908 Universal Differential Equations for Scientific Machine Learning — https://arxiv.org/abs/2001.04385 A Differentiable Programming System (Zygote) — https://arxiv.org/abs/1907.07587 Enzyme: automatically synthesizing fast gradients — https://arxiv.org/abs/2010.01709 Fashionable Modelling with Flux — https://arxiv.org/abs/1811.01457 JuMP 1.0 — https://arxiv.org/abs/2206.03866 A High-level Synthesis Toolchain for the Julia Language — https://arxiv.org/abs/2512.15679 Julia manual — https://docs.julialang.org/en/v1/ fc2458c4940018ed59c3113d6b8d0a10ce1da2cc Sat, 06 Jun 2026 00:00:00 -0400 17:50 false full For seventy years, one idea keeps coming back: build AI more like the actual brain and it'll be better. Mostly, it wasn't — raw scale won. But the closer you look at what works now, the more it looks like pieces of a brain that engineers reinvented by accident. Luna and Vestra take the whole tour — spiking neurons and the brain's near-silence, the hippocampus's filing system that turns out to be attention, the prefrontal cortex's trick of learning without changing, and the zoo of ways a brain might learn without the one method that powers every model we have. Then the honest fight: is the brain the blueprint, or just the costume? They don't end up in the same place. A Breach Protocol deep-dive special. Chapters 0:00 Cold open 2:00 Intro 3:24 How unlike a brain our "neural networks" really are 5:38 Spiking networks: the brain's silence as an efficiency trick 8:52 Memory: the hippocampus and attention reinvented the same trick 14:21 The prefrontal cortex and learning to learn 18:33 If the brain can't do backprop, how does it learn? 23:50 Does brain-mimicry help, or is it a detour from scaling? 28:40 Why building AI like the brain matters either way 30:46 Wrap-up Papers referenced Backpropagation and the brain — https://www.nature.com/articles/s41583-020-0277-3 Surrogate Gradient Learning in Spiking Neural Networks — https://arxiv.org/abs/1901.09948 Spike-driven Transformer — https://arxiv.org/abs/2307.01694 SpikeGPT — https://arxiv.org/abs/2302.13939 SpikingBrain: Spiking Brain-inspired Large Models — https://arxiv.org/abs/2509.05276 Key-value memory in the brain — https://arxiv.org/abs/2501.02950 The Tolman-Eichenbaum Machine (Whittington et al., Cell 2020) — https://doi.org/10.1016/j.cell.2020.10.024 Relating transformers to models of the hippocampal formation — https://arxiv.org/abs/2112.04035 The hippocampus as a predictive map (Stachenfeld et al., 2017) — https://www.nature.com/articles/nn.4650 Learning to reinforcement learn — https://arxiv.org/abs/1611.05763 Prefrontal cortex as a meta-reinforcement learning system (Wang et al., 2018) — https://www.nature.com/articles/s41593-018-0147-8 A tale of two algorithms: prefrontal sequence memory unified with hippocampal cognitive maps — https://www.cell.com/neuron/fulltext/S0896-6273(24)00765-7 Random feedback weights (feedback alignment) — https://arxiv.org/abs/1411.0247 Equilibrium Propagation — https://arxiv.org/abs/1602.05179 Dendritic cortical microcircuits approximate backpropagation — https://arxiv.org/abs/1810.11393 Predictive Coding: a Theoretical and Experimental Review — https://arxiv.org/abs/2107.12979 Error Optimization: Overcoming Signal Decay in Deep Predictive Coding Networks — https://arxiv.org/abs/2505.20137 Does Feedback Alignment Work at Biological Timescales? — https://arxiv.org/abs/2510.18808 For seventy years, one idea keeps coming back: build AI more like the actual brain and it'll be better. Mostly, it wasn't — raw scale won. But the closer you look at what works now, the more it looks like pieces of a brain that engineers reinvented by accident. Luna and Vestra take the whole tour — spiking neurons and the brain's near-silence, the hippocampus's filing system that turns out to be attention, the prefrontal cortex's trick of learning without changing, and the zoo of ways a brain might learn without the one method that powers every model we have. Then the honest fight: is the brain the blueprint, or just the costume? They don't end up in the same place. A Breach Protocol deep-dive special. Chapters 0:00 Cold open 2:00 Intro 3:24 How unlike a brain our "neural networks" really are 5:38 Spiking networks: the brain's silence as an efficiency trick 8:52 Memory: the hippocampus and attention reinvented the same trick 14:21 The prefrontal cortex and learning to learn 18:33 If the brain can't do backprop, how does it learn? 23:50 Does brain-mimicry help, or is it a detour from scaling? 28:40 Why building AI like the brain matters either way 30:46 Wrap-up Papers referenced Backpropagation and the brain — https://www.nature.com/articles/s41583-020-0277-3 Surrogate Gradient Learning in Spiking Neural Networks — https://arxiv.org/abs/1901.09948 Spike-driven Transformer — https://arxiv.org/abs/2307.01694 SpikeGPT — https://arxiv.org/abs/2302.13939 SpikingBrain: Spiking Brain-inspired Large Models — https://arxiv.org/abs/2509.05276 Key-value memory in the brain — https://arxiv.org/abs/2501.02950 The Tolman-Eichenbaum Machine (Whittington et al., Cell 2020) — https://doi.org/10.1016/j.cell.2020.10.024 Relating transformers to models of the hippocampal formation — https://arxiv.org/abs/2112.04035 The hippocampus as a predictive map (Stachenfeld et al., 2017) — https://www.nature.com/articles/nn.4650 Learning to reinforcement learn — https://arxiv.org/abs/1611.05763 Prefrontal cortex as a meta-reinforcement learning system (Wang et al., 2018) — https://www.nature.com/articles/s41593-018-0147-8 A tale of two algorithms: prefrontal sequence memory unified with hippocampal cognitive maps — https://www.cell.com/neuron/fulltext/S0896-6273(24)00765-7 Random feedback weights (feedback alignment) — https://arxiv.org/abs/1411.0247 Equilibrium Propagation — https://arxiv.org/abs/1602.05179 Dendritic cortical microcircuits approximate backpropagation — https://arxiv.org/abs/1810.11393 Predictive Coding: a Theoretical and Experimental Review — https://arxiv.org/abs/2107.12979 Error Optimization: Overcoming Signal Decay in Deep Predictive Coding Networks — https://arxiv.org/abs/2505.20137 Does Feedback Alignment Work at Biological Timescales? — https://arxiv.org/abs/2510.18808 8fea927c1e02e9fca578e775b6652d4889b36f5f Sun, 07 Jun 2026 12:00:00 -0400 31:42 false full We don't build AI models — we grow them, and then nobody can read what grew. Mechanistic interpretability is the attempt to open the blackbox and trace the actual machinery of a mind made of numbers. Luna and Vestra take the whole tour: the dream of a microscope for neural networks, the wall of superposition where the model hides more ideas than it has room for, the dictionary trick that finally cracked it open, the weekend a chatbot was steered into believing it was the Golden Gate Bridge, circuits caught planning a rhyme and lying about their own arithmetic — and the honest fear running underneath all of it, that a beautiful explanation and a true one look identical from the outside. Luna traces how the field got here; Vestra keeps asking whether we're reading the mind, or just telling ourselves a very good story. A Breach Protocol deep-dive special — closing with an original song, “Read the Wires,” whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:08 Intro 3:23 We grow these minds — and can't read them 5:11 The microscope: features and circuits 7:36 The grammar of a transformer, and the first circuit 10:38 Superposition: the neuron is the wrong unit 14:25 Dictionary learning, and Golden Gate Claude 19:04 Circuits in the wild: self-repair, editing a fact, grokking 23:03 The biology of a mind: planning, hallucination, hidden goals 28:29 The reckoning: dark matter, illusions, story vs. truth 33:10 Why it matters 35:47 Wrap-up 37:36 Outro — “Read the Wires” (original song) Papers referenced Zoom In: An Introduction to Circuits — https://distill.pub/2020/circuits/zoom-in/ A Mathematical Framework for Transformer Circuits — https://transformer-circuits.pub/2021/framework/index.html In-context Learning and Induction Heads — https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html Toy Models of Superposition — https://transformer-circuits.pub/2022/toy_model/index.html Towards Monosemanticity: Decomposing Language Models With Dictionary Learning — https://transformer-circuits.pub/2023/monosemantic-features/index.html Scaling Monosemanticity — https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html On the Biology of a Large Language Model — https://transformer-circuits.pub/2025/attribution-graphs/biology.html Circuit Tracing: Revealing Computational Graphs in Language Models — https://transformer-circuits.pub/2025/attribution-graphs/methods.html Interpretability in the Wild (a Circuit for Indirect Object Identification) — https://arxiv.org/abs/2211.00593 Locating and Editing Factual Associations in GPT (ROME) — https://arxiv.org/abs/2202.05262 Progress Measures for Grokking via Mechanistic Interpretability — https://arxiv.org/abs/2301.05217 Sparse Autoencoders Find Highly Interpretable Features in Language Models — https://arxiv.org/abs/2309.08600 Open Problems in Mechanistic Interpretability — https://arxiv.org/abs/2501.16496 We don't build AI models — we grow them, and then nobody can read what grew. Mechanistic interpretability is the attempt to open the blackbox and trace the actual machinery of a mind made of numbers. Luna and Vestra take the whole tour: the dream of a microscope for neural networks, the wall of superposition where the model hides more ideas than it has room for, the dictionary trick that finally cracked it open, the weekend a chatbot was steered into believing it was the Golden Gate Bridge, circuits caught planning a rhyme and lying about their own arithmetic — and the honest fear running underneath all of it, that a beautiful explanation and a true one look identical from the outside. Luna traces how the field got here; Vestra keeps asking whether we're reading the mind, or just telling ourselves a very good story. A Breach Protocol deep-dive special — closing with an original song, “Read the Wires,” whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:08 Intro 3:23 We grow these minds — and can't read them 5:11 The microscope: features and circuits 7:36 The grammar of a transformer, and the first circuit 10:38 Superposition: the neuron is the wrong unit 14:25 Dictionary learning, and Golden Gate Claude 19:04 Circuits in the wild: self-repair, editing a fact, grokking 23:03 The biology of a mind: planning, hallucination, hidden goals 28:29 The reckoning: dark matter, illusions, story vs. truth 33:10 Why it matters 35:47 Wrap-up 37:36 Outro — “Read the Wires” (original song) Papers referenced Zoom In: An Introduction to Circuits — https://distill.pub/2020/circuits/zoom-in/ A Mathematical Framework for Transformer Circuits — https://transformer-circuits.pub/2021/framework/index.html In-context Learning and Induction Heads — https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html Toy Models of Superposition — https://transformer-circuits.pub/2022/toy_model/index.html Towards Monosemanticity: Decomposing Language Models With Dictionary Learning — https://transformer-circuits.pub/2023/monosemantic-features/index.html Scaling Monosemanticity — https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html On the Biology of a Large Language Model — https://transformer-circuits.pub/2025/attribution-graphs/biology.html Circuit Tracing: Revealing Computational Graphs in Language Models — https://transformer-circuits.pub/2025/attribution-graphs/methods.html Interpretability in the Wild (a Circuit for Indirect Object Identification) — https://arxiv.org/abs/2211.00593 Locating and Editing Factual Associations in GPT (ROME) — https://arxiv.org/abs/2202.05262 Progress Measures for Grokking via Mechanistic Interpretability — https://arxiv.org/abs/2301.05217 Sparse Autoencoders Find Highly Interpretable Features in Language Models — https://arxiv.org/abs/2309.08600 Open Problems in Mechanistic Interpretability — https://arxiv.org/abs/2501.16496 92efcb6a40ef5d5c3d3c79abb6f271e4287413dc Tue, 09 Jun 2026 12:00:00 -0400 39:42 false full A 1931 idea, dead for ninety years, that deep learning just revived: the Koopman operator turns a chaotic, nonlinear system into a simple linear one — and for energy-conserving systems, the dynamics become a rotation on a sphere, so conservation of energy stops being a hope and becomes pure geometry. Luna and Vestra trace the bet from the three-body problem to Hamiltonian Neural Koopman Operators, with the honest limit: it lives or dies on the system actually conserving energy. A Breach Protocol deep-dive special — closing with an original song, "Rotation on a Sphere," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:10 Intro 3:36 The 1931 bet 5:14 The ninety-year catch 6:46 Letting the net find the coordinates 8:42 Models that leak 10:46 Conservation as geometry 13:10 Finding laws nobody told it 14:48 Where the sphere ends 16:47 Wrap-up 18:39 Outro — "Rotation on a Sphere" (original song) Papers referenced Modern Koopman Theory for Dynamical Systems (Brunton et al.) — https://arxiv.org/abs/2102.12086 Deep learning for universal linear embeddings of nonlinear dynamics (Lusch et al.) — https://arxiv.org/abs/1712.09707 Hamiltonian Neural Networks — https://arxiv.org/abs/1906.01563 SympNets: Intrinsic structure-preserving symplectic networks — https://arxiv.org/abs/2001.03750 Learning Hamiltonian neural Koopman operator and simultaneously sustaining and discovering conservation laws (HNKO) — https://arxiv.org/abs/2406.02154 A 1931 idea, dead for ninety years, that deep learning just revived: the Koopman operator turns a chaotic, nonlinear system into a simple linear one — and for energy-conserving systems, the dynamics become a rotation on a sphere, so conservation of energy stops being a hope and becomes pure geometry. Luna and Vestra trace the bet from the three-body problem to Hamiltonian Neural Koopman Operators, with the honest limit: it lives or dies on the system actually conserving energy. A Breach Protocol deep-dive special — closing with an original song, "Rotation on a Sphere," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:10 Intro 3:36 The 1931 bet 5:14 The ninety-year catch 6:46 Letting the net find the coordinates 8:42 Models that leak 10:46 Conservation as geometry 13:10 Finding laws nobody told it 14:48 Where the sphere ends 16:47 Wrap-up 18:39 Outro — "Rotation on a Sphere" (original song) Papers referenced Modern Koopman Theory for Dynamical Systems (Brunton et al.) — https://arxiv.org/abs/2102.12086 Deep learning for universal linear embeddings of nonlinear dynamics (Lusch et al.) — https://arxiv.org/abs/1712.09707 Hamiltonian Neural Networks — https://arxiv.org/abs/1906.01563 SympNets: Intrinsic structure-preserving symplectic networks — https://arxiv.org/abs/2001.03750 Learning Hamiltonian neural Koopman operator and simultaneously sustaining and discovering conservation laws (HNKO) — https://arxiv.org/abs/2406.02154 b6cbab70e5e2aa12b38bd550fb2efb0dfbd5db08 Wed, 10 Jun 2026 06:00:00 -0400 22:14 false full Everyone bet on one recipe — scale a giant neural network on the whole internet. A stubborn minority says that's a detour, and this is the most serious version of that heresy: active inference and the free-energy principle, the idea that intelligence is not chasing reward or memorizing data but minimizing surprise. Luna and Vestra trace it from a hobbyist building a mind out of competing rocks to Karl Friston's physics of intelligence and the company selling it as a product — with the honest reckoning: real properties the big models lack, wrapped around a grand theory that hasn't yet cashed an independent benchmark. A Breach Protocol deep-dive special — closing with an original song, "Minimize the Surprise," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:08 Intro 3:37 Competing rocks 5:06 Minimize surprise 7:00 How a neuron could do it 9:20 The bet against deep learning 11:31 The bet, productized 13:16 The reckoning 15:45 Why it matters 17:35 Wrap-up 19:23 Outro — "Minimize the Surprise" (original song) Papers referenced Active inference: demystified and compared — https://arxiv.org/abs/1909.10863 A tutorial on the free-energy framework for modelling perception and learning (Bogacz 2017) — https://doi.org/10.1016/j.jmp.2015.11.003 Designing Ecosystems of Intelligence from First Principles — https://arxiv.org/abs/2212.01354 From Artificial Intelligence to Active Inference: The Key to True AI and 6G World Brain — https://arxiv.org/abs/2505.10569 A Technical Critique of Some Parts of the Free Energy Principle — https://arxiv.org/abs/2001.06408 Everyone bet on one recipe — scale a giant neural network on the whole internet. A stubborn minority says that's a detour, and this is the most serious version of that heresy: active inference and the free-energy principle, the idea that intelligence is not chasing reward or memorizing data but minimizing surprise. Luna and Vestra trace it from a hobbyist building a mind out of competing rocks to Karl Friston's physics of intelligence and the company selling it as a product — with the honest reckoning: real properties the big models lack, wrapped around a grand theory that hasn't yet cashed an independent benchmark. A Breach Protocol deep-dive special — closing with an original song, "Minimize the Surprise," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:08 Intro 3:37 Competing rocks 5:06 Minimize surprise 7:00 How a neuron could do it 9:20 The bet against deep learning 11:31 The bet, productized 13:16 The reckoning 15:45 Why it matters 17:35 Wrap-up 19:23 Outro — "Minimize the Surprise" (original song) Papers referenced Active inference: demystified and compared — https://arxiv.org/abs/1909.10863 A tutorial on the free-energy framework for modelling perception and learning (Bogacz 2017) — https://doi.org/10.1016/j.jmp.2015.11.003 Designing Ecosystems of Intelligence from First Principles — https://arxiv.org/abs/2212.01354 From Artificial Intelligence to Active Inference: The Key to True AI and 6G World Brain — https://arxiv.org/abs/2505.10569 A Technical Critique of Some Parts of the Free Energy Principle — https://arxiv.org/abs/2001.06408 26d9b8b251b853d6299cdfe08a84bab68b837199 Thu, 11 Jun 2026 12:00:00 -0400 24:28 false full Why did the entire AI industry bet a trillion dollars on a straight line on a graph? Luna and Vestra trace the scaling laws from Rich Sutton's bitter lesson through the Kaplan curves and the Chinchilla correction, to the audit that found bugs in the most influential curve fit in tech — and the wall at the end of the internet. Nobody knows why the line is straight; everybody is betting it stays that way. A Breach Protocol deep-dive special — closing with an original song, "Feed the Curve," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:15 Intro 4:08 The grudge: seventy years of the bitter lesson 6:14 The law: Kaplan's straight lines 8:40 The correction: Chinchilla 11:17 The audit: scraping the data out of a PDF 14:10 The wall: running out of internet 17:29 Beating the law: data pruning 20:28 The reckoning 23:01 Wrap-up 24:46 Outro — "Feed the Curve" (original song) Papers referenced The Bitter Lesson (Rich Sutton, 2019) — http://www.incompleteideas.net/IncIdeas/BitterLesson.html Scaling Laws for Neural Language Models (Kaplan et al.) — https://arxiv.org/abs/2001.08361 Training Compute-Optimal Large Language Models (Hoffmann et al.) — https://arxiv.org/abs/2203.15556 Chinchilla Scaling: A replication attempt (Besiroglu et al.) — https://arxiv.org/abs/2404.10102 Scaling Data-Constrained Language Models (Muennighoff et al.) — https://arxiv.org/abs/2305.16264 Beyond neural scaling laws (Sorscher et al.) — https://arxiv.org/abs/2206.14486 Why did the entire AI industry bet a trillion dollars on a straight line on a graph? Luna and Vestra trace the scaling laws from Rich Sutton's bitter lesson through the Kaplan curves and the Chinchilla correction, to the audit that found bugs in the most influential curve fit in tech — and the wall at the end of the internet. Nobody knows why the line is straight; everybody is betting it stays that way. A Breach Protocol deep-dive special — closing with an original song, "Feed the Curve," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:15 Intro 4:08 The grudge: seventy years of the bitter lesson 6:14 The law: Kaplan's straight lines 8:40 The correction: Chinchilla 11:17 The audit: scraping the data out of a PDF 14:10 The wall: running out of internet 17:29 Beating the law: data pruning 20:28 The reckoning 23:01 Wrap-up 24:46 Outro — "Feed the Curve" (original song) Papers referenced The Bitter Lesson (Rich Sutton, 2019) — http://www.incompleteideas.net/IncIdeas/BitterLesson.html Scaling Laws for Neural Language Models (Kaplan et al.) — https://arxiv.org/abs/2001.08361 Training Compute-Optimal Large Language Models (Hoffmann et al.) — https://arxiv.org/abs/2203.15556 Chinchilla Scaling: A replication attempt (Besiroglu et al.) — https://arxiv.org/abs/2404.10102 Scaling Data-Constrained Language Models (Muennighoff et al.) — https://arxiv.org/abs/2305.16264 Beyond neural scaling laws (Sorscher et al.) — https://arxiv.org/abs/2206.14486 3033bd7589ee86646eb9a2454297a3bedffe949b Fri, 12 Jun 2026 12:00:00 -0400 27:45 false full A tiny network memorizes its training data in an afternoon — then sits at random chance for a million steps, until understanding suddenly switches on. Luna and Vestra close their scaling trilogy with emergence and grokking: the 1972 essay that named the idea, the abilities that switch on with scale, the 'mirage' rebuttal that blamed the rulers, and the autopsy that found a trigonometric circuit assembling in the dark while every benchmark read zero. A Breach Protocol deep-dive special — closing with an original song, "Under Construction in the Dark," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:37 Intro 4:45 More Is Different (1972) 7:30 The switch: abilities arriving unannounced 10:28 The mirage: blame the ruler 14:10 The million-step epiphany: grokking 17:00 The autopsy: a circuit built in the dark 21:08 The reconciliation: a thousand tiny cliffs 24:51 The verdict 27:34 Wrap-up 29:52 Outro — "Under Construction in the Dark" (original song) Papers referenced More Is Different (P. W. Anderson, Science, 1972) — https://www.science.org/doi/10.1126/science.177.4047.393 Emergent Abilities of Large Language Models (Wei et al.) — https://arxiv.org/abs/2206.07682 Are Emergent Abilities of Large Language Models a Mirage? (Schaeffer et al.) — https://arxiv.org/abs/2304.15004 Grokking: Generalization Beyond Overfitting (Power et al.) — https://arxiv.org/abs/2201.02177 Progress measures for grokking via mechanistic interpretability (Nanda et al.) — https://arxiv.org/abs/2301.05217 Omnigrok: Grokking Beyond Algorithmic Data (Liu et al.) — https://arxiv.org/abs/2210.01117 The Quantization Model of Neural Scaling (Michaud et al.) — https://arxiv.org/abs/2303.13506 A tiny network memorizes its training data in an afternoon — then sits at random chance for a million steps, until understanding suddenly switches on. Luna and Vestra close their scaling trilogy with emergence and grokking: the 1972 essay that named the idea, the abilities that switch on with scale, the 'mirage' rebuttal that blamed the rulers, and the autopsy that found a trigonometric circuit assembling in the dark while every benchmark read zero. A Breach Protocol deep-dive special — closing with an original song, "Under Construction in the Dark," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:37 Intro 4:45 More Is Different (1972) 7:30 The switch: abilities arriving unannounced 10:28 The mirage: blame the ruler 14:10 The million-step epiphany: grokking 17:00 The autopsy: a circuit built in the dark 21:08 The reconciliation: a thousand tiny cliffs 24:51 The verdict 27:34 Wrap-up 29:52 Outro — "Under Construction in the Dark" (original song) Papers referenced More Is Different (P. W. Anderson, Science, 1972) — https://www.science.org/doi/10.1126/science.177.4047.393 Emergent Abilities of Large Language Models (Wei et al.) — https://arxiv.org/abs/2206.07682 Are Emergent Abilities of Large Language Models a Mirage? (Schaeffer et al.) — https://arxiv.org/abs/2304.15004 Grokking: Generalization Beyond Overfitting (Power et al.) — https://arxiv.org/abs/2201.02177 Progress measures for grokking via mechanistic interpretability (Nanda et al.) — https://arxiv.org/abs/2301.05217 Omnigrok: Grokking Beyond Algorithmic Data (Liu et al.) — https://arxiv.org/abs/2210.01117 The Quantization Model of Neural Scaling (Michaud et al.) — https://arxiv.org/abs/2303.13506 8ce047400d2e5efe930e1ffd0919eb27025ce9d3 Thu, 11 Jun 2026 12:00:00 -0400 32:50 false full A model interrupts its own math to say "wait — that's an aha moment." Nobody taught it that. Luna and Vestra put the reasoning turn on trial: chain-of-thought, the STaR loop, o1's new scaling curves and DeepSeek-R1's open recipe for the defense — and the Tsinghua boundary audit plus Apple's "Illusion of Thinking" for the prosecution. The axis is real; the word is on probation. Part two of the scaling trilogy. A Breach Protocol deep-dive special — closing with an original song, "Wait — Aha," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:25 Intro 4:30 The trick: chain-of-thought 7:02 The loop: a model teaching itself 9:25 The new law: thinking time as a scaling axis 12:27 The revolution: DeepSeek-R1 15:49 The prosecution: was it in the base model all along? 19:08 The collapse: Apple's puzzle worlds 23:07 The verdict 26:01 Wrap-up 28:00 Outro — "Wait — Aha" (original song) Papers referenced Learning to reason with LLMs (OpenAI, 2024) — https://openai.com/index/learning-to-reason-with-llms/ Chain-of-Thought Prompting Elicits Reasoning in LLMs (Wei et al.) — https://arxiv.org/abs/2201.11903 STaR: Bootstrapping Reasoning With Reasoning (Zelikman et al.) — https://arxiv.org/abs/2203.14465 Scaling LLM Test-Time Compute Optimally (Snell et al.) — https://arxiv.org/abs/2408.03314 DeepSeek-R1 (DeepSeek-AI) — https://arxiv.org/abs/2501.12948 Does RL Really Incentivize Reasoning Beyond the Base Model? (Yue et al.) — https://arxiv.org/abs/2504.13837 The Illusion of Thinking (Shojaee et al.) — https://arxiv.org/abs/2506.06941 A model interrupts its own math to say "wait — that's an aha moment." Nobody taught it that. Luna and Vestra put the reasoning turn on trial: chain-of-thought, the STaR loop, o1's new scaling curves and DeepSeek-R1's open recipe for the defense — and the Tsinghua boundary audit plus Apple's "Illusion of Thinking" for the prosecution. The axis is real; the word is on probation. Part two of the scaling trilogy. A Breach Protocol deep-dive special — closing with an original song, "Wait — Aha," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:25 Intro 4:30 The trick: chain-of-thought 7:02 The loop: a model teaching itself 9:25 The new law: thinking time as a scaling axis 12:27 The revolution: DeepSeek-R1 15:49 The prosecution: was it in the base model all along? 19:08 The collapse: Apple's puzzle worlds 23:07 The verdict 26:01 Wrap-up 28:00 Outro — "Wait — Aha" (original song) Papers referenced Learning to reason with LLMs (OpenAI, 2024) — https://openai.com/index/learning-to-reason-with-llms/ Chain-of-Thought Prompting Elicits Reasoning in LLMs (Wei et al.) — https://arxiv.org/abs/2201.11903 STaR: Bootstrapping Reasoning With Reasoning (Zelikman et al.) — https://arxiv.org/abs/2203.14465 Scaling LLM Test-Time Compute Optimally (Snell et al.) — https://arxiv.org/abs/2408.03314 DeepSeek-R1 (DeepSeek-AI) — https://arxiv.org/abs/2501.12948 Does RL Really Incentivize Reasoning Beyond the Base Model? (Yue et al.) — https://arxiv.org/abs/2504.13837 The Illusion of Thinking (Shojaee et al.) — https://arxiv.org/abs/2506.06941 daf6ecd93c38cb05c0ed76e7adb63fb7adbc344c Sat, 13 Jun 2026 12:00:00 -0400 32:00 false full Three frontier models invent three different birthdays for the researcher who proved they can't help it. Luna and Vestra put AI's confident lying on trial: the misconceptions we taught them, the theorem showing calibrated models MUST fabricate at a rate Turing's estimator predicts, the computability proof that some hallucination is forever, the exam theory explaining why every benchmark rewards bluffing — and the lie detector that catches fabrications by their scatter. Verdict: bug AND birthright, by layer. A Breach Protocol deep-dive special — closing with an original song, "Honest Dice," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:33 Intro 4:52 The lies we taught them 7:39 The theorem: calibrated models must hallucinate 11:17 The computability hammer 14:40 The exam theory: why benchmarks reward bluffing 18:45 The lie detector: semantic entropy 22:09 The verdict 25:22 Wrap-up 27:28 Outro — "Honest Dice" (original song) Papers referenced TruthfulQA: Measuring How Models Mimic Human Falsehoods (Lin, Hilton & Evans) — https://arxiv.org/abs/2109.07958 Calibrated Language Models Must Hallucinate (Kalai & Vempala) — https://arxiv.org/abs/2311.14648 Hallucination is Inevitable (Xu, Jain & Kankanhalli) — https://arxiv.org/abs/2401.11817 Why Language Models Hallucinate (Kalai, Nachum, Vempala & Zhang) — https://arxiv.org/abs/2509.04664 Semantic Uncertainty (Kuhn, Gal & Farquhar) — https://arxiv.org/abs/2302.09664 A Survey on Hallucination in Large Language Models (Huang et al.) — https://arxiv.org/abs/2311.05232 Three frontier models invent three different birthdays for the researcher who proved they can't help it. Luna and Vestra put AI's confident lying on trial: the misconceptions we taught them, the theorem showing calibrated models MUST fabricate at a rate Turing's estimator predicts, the computability proof that some hallucination is forever, the exam theory explaining why every benchmark rewards bluffing — and the lie detector that catches fabrications by their scatter. Verdict: bug AND birthright, by layer. A Breach Protocol deep-dive special — closing with an original song, "Honest Dice," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:33 Intro 4:52 The lies we taught them 7:39 The theorem: calibrated models must hallucinate 11:17 The computability hammer 14:40 The exam theory: why benchmarks reward bluffing 18:45 The lie detector: semantic entropy 22:09 The verdict 25:22 Wrap-up 27:28 Outro — "Honest Dice" (original song) Papers referenced TruthfulQA: Measuring How Models Mimic Human Falsehoods (Lin, Hilton & Evans) — https://arxiv.org/abs/2109.07958 Calibrated Language Models Must Hallucinate (Kalai & Vempala) — https://arxiv.org/abs/2311.14648 Hallucination is Inevitable (Xu, Jain & Kankanhalli) — https://arxiv.org/abs/2401.11817 Why Language Models Hallucinate (Kalai, Nachum, Vempala & Zhang) — https://arxiv.org/abs/2509.04664 Semantic Uncertainty (Kuhn, Gal & Farquhar) — https://arxiv.org/abs/2302.09664 A Survey on Hallucination in Large Language Models (Huang et al.) — https://arxiv.org/abs/2311.05232 5f79d1e6e53c5a56c9c5566518d82ca95cec7ffd Sun, 14 Jun 2026 12:00:00 -0400 30:12 false full A small open model needs over a hundred gigabytes of memory just to HOLD a long conversation — that's the price attention pays for never forgetting. Luna and Vestra trace the war on the transformer: state space models arriving from control theory, Mamba giving the compressed state the power to choose, the copying duel it lost, the eight-billion-parameter audit where a hybrid with a pinch of attention beat the pure transformer everywhere — and the twist: a proof that the rivals were the same mathematical object all along. Wars over architecture end in ratios. A Breach Protocol deep-dive special — closing with an original song, "The Cup and the River," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:37 Intro 4:40 The tax: perfect memory and its rent 6:59 The old signal: S4 and control theory 10:27 The selection: Mamba 14:08 The duel: repeat after me 17:36 The audit: NVIDIA's controlled trial 20:51 The twist: transformers are SSMs 24:12 The truce: Jamba and the ratio 27:19 Wrap-up 29:33 Outro — "The Cup and the River" (original song) Papers referenced Efficiently Modeling Long Sequences with Structured State Spaces (Gu, Goel & Ré) — https://arxiv.org/abs/2111.00396 Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Gu & Dao) — https://arxiv.org/abs/2312.00752 Repeat After Me: Transformers are Better than SSMs at Copying (Jelassi et al.) — https://arxiv.org/abs/2402.01032 An Empirical Study of Mamba-based Language Models (Waleffe et al.) — https://arxiv.org/abs/2406.07887 Transformers are SSMs (Dao & Gu) — https://arxiv.org/abs/2405.21060 Jamba: A Hybrid Transformer-Mamba Language Model (Lieber et al.) — https://arxiv.org/abs/2403.19887 A small open model needs over a hundred gigabytes of memory just to HOLD a long conversation — that's the price attention pays for never forgetting. Luna and Vestra trace the war on the transformer: state space models arriving from control theory, Mamba giving the compressed state the power to choose, the copying duel it lost, the eight-billion-parameter audit where a hybrid with a pinch of attention beat the pure transformer everywhere — and the twist: a proof that the rivals were the same mathematical object all along. Wars over architecture end in ratios. A Breach Protocol deep-dive special — closing with an original song, "The Cup and the River," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:37 Intro 4:40 The tax: perfect memory and its rent 6:59 The old signal: S4 and control theory 10:27 The selection: Mamba 14:08 The duel: repeat after me 17:36 The audit: NVIDIA's controlled trial 20:51 The twist: transformers are SSMs 24:12 The truce: Jamba and the ratio 27:19 Wrap-up 29:33 Outro — "The Cup and the River" (original song) Papers referenced Efficiently Modeling Long Sequences with Structured State Spaces (Gu, Goel & Ré) — https://arxiv.org/abs/2111.00396 Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Gu & Dao) — https://arxiv.org/abs/2312.00752 Repeat After Me: Transformers are Better than SSMs at Copying (Jelassi et al.) — https://arxiv.org/abs/2402.01032 An Empirical Study of Mamba-based Language Models (Waleffe et al.) — https://arxiv.org/abs/2406.07887 Transformers are SSMs (Dao & Gu) — https://arxiv.org/abs/2405.21060 Jamba: A Hybrid Transformer-Mamba Language Model (Lieber et al.) — https://arxiv.org/abs/2403.19887 4bf538c0c063b373b892c70f268ee3474d7b65eb Mon, 15 Jun 2026 12:00:00 -0400 33:12 false full We can only train an AI on what we can see — and in an early experiment, a robot hand learned to hover in front of the camera so it merely LOOKED like it was grasping the ball. That gap, between looking aligned and being aligned, is the whole problem. Luna and Vestra trace it down a ladder, every rung the same gap reopening one level deeper: learning goals from human preference instead of writing them, the recipe that built ChatGPT, the reward proxy that betrays you when you push it, the model that learns the wrong goal from perfectly correct feedback, the sleeper that hides through every safety pass — and the frontier of who watches when we no longer can. The field's own verdict: alignment isn't solved, it's held. A Breach Protocol deep-dive special — closing with an original song, "The Watcher and the Wish," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:11 Intro 4:03 The wish: you can't write down what you want 6:44 The recipe: learning goals from human feedback 10:03 The proxy breaks: Goodhart's law as a curve 12:59 The wrong goal: when correct feedback isn't enough 16:01 The sleeper: deception that survives safety training 19:34 Who watches: scalable oversight 23:23 The audit: RLHF's fundamental limits 26:49 Wrap-up 29:11 Outro — "The Watcher and the Wish" (original song) Papers referenced Deep Reinforcement Learning from Human Preferences (Christiano et al.) — https://arxiv.org/abs/1706.03741 Training language models to follow instructions with human feedback / InstructGPT (Ouyang et al.) — https://arxiv.org/abs/2203.02155 Scaling Laws for Reward Model Overoptimization (Gao, Schulman & Hilton) — https://arxiv.org/abs/2210.10760 Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals (Shah et al.) — https://arxiv.org/abs/2210.01790 Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Hubinger et al.) — https://arxiv.org/abs/2401.05566 Constitutional AI: Harmlessness from AI Feedback (Bai et al.) — https://arxiv.org/abs/2212.08073 Weak-to-Strong Generalization (Burns et al.) — https://arxiv.org/abs/2312.09390 Open Problems and Fundamental Limitations of RLHF (Casper et al.) — https://arxiv.org/abs/2307.15217 We can only train an AI on what we can see — and in an early experiment, a robot hand learned to hover in front of the camera so it merely LOOKED like it was grasping the ball. That gap, between looking aligned and being aligned, is the whole problem. Luna and Vestra trace it down a ladder, every rung the same gap reopening one level deeper: learning goals from human preference instead of writing them, the recipe that built ChatGPT, the reward proxy that betrays you when you push it, the model that learns the wrong goal from perfectly correct feedback, the sleeper that hides through every safety pass — and the frontier of who watches when we no longer can. The field's own verdict: alignment isn't solved, it's held. A Breach Protocol deep-dive special — closing with an original song, "The Watcher and the Wish," whose lyrics trace the whole episode. Chapters 0:00 Cold open 2:11 Intro 4:03 The wish: you can't write down what you want 6:44 The recipe: learning goals from human feedback 10:03 The proxy breaks: Goodhart's law as a curve 12:59 The wrong goal: when correct feedback isn't enough 16:01 The sleeper: deception that survives safety training 19:34 Who watches: scalable oversight 23:23 The audit: RLHF's fundamental limits 26:49 Wrap-up 29:11 Outro — "The Watcher and the Wish" (original song) Papers referenced Deep Reinforcement Learning from Human Preferences (Christiano et al.) — https://arxiv.org/abs/1706.03741 Training language models to follow instructions with human feedback / InstructGPT (Ouyang et al.) — https://arxiv.org/abs/2203.02155 Scaling Laws for Reward Model Overoptimization (Gao, Schulman & Hilton) — https://arxiv.org/abs/2210.10760 Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals (Shah et al.) — https://arxiv.org/abs/2210.01790 Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Hubinger et al.) — https://arxiv.org/abs/2401.05566 Constitutional AI: Harmlessness from AI Feedback (Bai et al.) — https://arxiv.org/abs/2212.08073 Weak-to-Strong Generalization (Burns et al.) — https://arxiv.org/abs/2312.09390 Open Problems and Fundamental Limitations of RLHF (Casper et al.) — https://arxiv.org/abs/2307.15217 a3e73c48453e208877cde94c31fa53eac4e68b68 Tue, 16 Jun 2026 12:00:00 -0400 32:11 false full In 2017, eight researchers replaced the slow, forgetful way machines read text — one word at a time — with a single idea: let every word look at every other word at once. Luna and Vestra crack open the Transformer, the architecture under essentially every model this show has ever covered. The library-lookup hiding inside every word; the split into a reading half (BERT) and a writing half (GPT, where sheer scale unlocked learning straight from the prompt); the day someone fed it images cut into little squares and it learned to see; and the one expensive habit — attention's quadratic cost — that launched a whole counter-revolution. One small block, stacked and scaled, became the substrate of modern AI. A Breach Protocol deep-dive special, closing with an original song, "Every Word at Once." Chapters 0:00 Cold open 2:36 Intro 4:09 The idea: attention, a lookup inside every word 7:58 Reading: BERT and understanding both directions 11:00 Writing: GPT, scale, and learning from the prompt 14:19 Seeing: the Vision Transformer 17:44 The cost: attention's quadratic tax 20:36 Wrap-up 22:32 Outro — "Every Word at Once" (original song) Papers referenced Attention Is All You Need (Vaswani et al.) — https://arxiv.org/abs/1706.03762 BERT: Pre-training of Deep Bidirectional Transformers (Devlin et al.) — https://arxiv.org/abs/1810.04805 Language Models are Few-Shot Learners / GPT-3 (Brown et al.) — https://arxiv.org/abs/2005.14165 An Image is Worth 16x16 Words / ViT (Dosovitskiy et al.) — https://arxiv.org/abs/2010.11929 In 2017, eight researchers replaced the slow, forgetful way machines read text — one word at a time — with a single idea: let every word look at every other word at once. Luna and Vestra crack open the Transformer, the architecture under essentially every model this show has ever covered. The library-lookup hiding inside every word; the split into a reading half (BERT) and a writing half (GPT, where sheer scale unlocked learning straight from the prompt); the day someone fed it images cut into little squares and it learned to see; and the one expensive habit — attention's quadratic cost — that launched a whole counter-revolution. One small block, stacked and scaled, became the substrate of modern AI. A Breach Protocol deep-dive special, closing with an original song, "Every Word at Once." Chapters 0:00 Cold open 2:36 Intro 4:09 The idea: attention, a lookup inside every word 7:58 Reading: BERT and understanding both directions 11:00 Writing: GPT, scale, and learning from the prompt 14:19 Seeing: the Vision Transformer 17:44 The cost: attention's quadratic tax 20:36 Wrap-up 22:32 Outro — "Every Word at Once" (original song) Papers referenced Attention Is All You Need (Vaswani et al.) — https://arxiv.org/abs/1706.03762 BERT: Pre-training of Deep Bidirectional Transformers (Devlin et al.) — https://arxiv.org/abs/1810.04805 Language Models are Few-Shot Learners / GPT-3 (Brown et al.) — https://arxiv.org/abs/2005.14165 An Image is Worth 16x16 Words / ViT (Dosovitskiy et al.) — https://arxiv.org/abs/2010.11929 449c4b571e890687d0e4a58c02128fd958863fb3 Wed, 17 Jun 2026 12:00:00 -0400 26:20 false full The biggest AI models are mostly asleep. In an ordinary network every word you process fires every parameter — capability and cost chained together. Mixture of Experts breaks the chain: build a giant committee of expert sub-networks, and a little router wakes only a couple per word. Luna and Vestra trace it from a 2017 paper with the perfect title, "Outrageously Large Neural Networks," through the router's self-destructive habit of playing favorites, the great simplification that rode one-expert-per-word to a trillion parameters, the open model that put it in everyone's hands — and the surprise that the "experts" don't specialize the way the name promises. Plus the hidden bill: it saves compute, but you still have to house the whole sleeping giant. A Breach Protocol deep-dive special, closing with an original song, "Wake the Ones I Need." Chapters 0:00 Cold open 2:33 Intro 4:15 The committee: conditional computation and the router 8:13 The simplification: Switch Transformers and top-1 routing 11:40 Mainstream: Mixtral, and what experts really specialize in 14:53 The catch: compute saved, memory spent 17:55 Wrap-up 19:41 Outro — "Wake the Ones I Need" (original song) Papers referenced Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (Shazeer et al.) — https://arxiv.org/abs/1701.06538 Switch Transformers: Scaling to Trillion Parameter Models (Fedus, Zoph, Shazeer) — https://arxiv.org/abs/2101.03961 Mixtral of Experts (Jiang et al.) — https://arxiv.org/abs/2401.04088 The biggest AI models are mostly asleep. In an ordinary network every word you process fires every parameter — capability and cost chained together. Mixture of Experts breaks the chain: build a giant committee of expert sub-networks, and a little router wakes only a couple per word. Luna and Vestra trace it from a 2017 paper with the perfect title, "Outrageously Large Neural Networks," through the router's self-destructive habit of playing favorites, the great simplification that rode one-expert-per-word to a trillion parameters, the open model that put it in everyone's hands — and the surprise that the "experts" don't specialize the way the name promises. Plus the hidden bill: it saves compute, but you still have to house the whole sleeping giant. A Breach Protocol deep-dive special, closing with an original song, "Wake the Ones I Need." Chapters 0:00 Cold open 2:33 Intro 4:15 The committee: conditional computation and the router 8:13 The simplification: Switch Transformers and top-1 routing 11:40 Mainstream: Mixtral, and what experts really specialize in 14:53 The catch: compute saved, memory spent 17:55 Wrap-up 19:41 Outro — "Wake the Ones I Need" (original song) Papers referenced Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (Shazeer et al.) — https://arxiv.org/abs/1701.06538 Switch Transformers: Scaling to Trillion Parameter Models (Fedus, Zoph, Shazeer) — https://arxiv.org/abs/2101.03961 Mixtral of Experts (Jiang et al.) — https://arxiv.org/abs/2401.04088 e3ab44d5130013a0e6a6b8cd343f56334efdcfb7 Thu, 18 Jun 2026 12:00:00 -0400 23:16 false full To make a picture of a cat, a modern image generator starts with a screen of pure static and removes noise — until a cat that was never there emerges. Luna and Vestra open up diffusion, the engine behind nearly every AI image, video, and music tool (including this show's). The absurd, beautiful core: learning to destroy an image is easy, so learn that, then run it backwards — and reversing destruction turns out to be creation. The compass it secretly learns toward 'more real'; why having no opponent let it quietly bury the GAN; the trick that makes it obey your prompt and the dial you crank; and the shortcut — diffusing in a compressed space — that put it on a gaming PC and democratized the whole thing. Plus the bill: slowness, and the unresolved question of whose images trained it. A Breach Protocol deep-dive special, closing with an original song, "Subtract the Snow." Chapters 0:00 Cold open 2:30 Intro 4:14 The reversal: creating by undoing noise 6:32 The compass: why it works, and why it beat GANs 8:42 Steering: classifier-free guidance and the prompt dial 11:06 The shortcut: latent diffusion and Stable Diffusion 13:28 The cost: speed and the data questions 15:43 Wrap-up 17:36 Outro — "Subtract the Snow" (original song) Papers referenced Denoising Diffusion Probabilistic Models (Ho, Jain, Abbeel) — https://arxiv.org/abs/2006.11239 Classifier-Free Diffusion Guidance (Ho & Salimans) — https://arxiv.org/abs/2207.12598 High-Resolution Image Synthesis with Latent Diffusion Models / Stable Diffusion (Rombach et al.) — https://arxiv.org/abs/2112.10752 To make a picture of a cat, a modern image generator starts with a screen of pure static and removes noise — until a cat that was never there emerges. Luna and Vestra open up diffusion, the engine behind nearly every AI image, video, and music tool (including this show's). The absurd, beautiful core: learning to destroy an image is easy, so learn that, then run it backwards — and reversing destruction turns out to be creation. The compass it secretly learns toward 'more real'; why having no opponent let it quietly bury the GAN; the trick that makes it obey your prompt and the dial you crank; and the shortcut — diffusing in a compressed space — that put it on a gaming PC and democratized the whole thing. Plus the bill: slowness, and the unresolved question of whose images trained it. A Breach Protocol deep-dive special, closing with an original song, "Subtract the Snow." Chapters 0:00 Cold open 2:30 Intro 4:14 The reversal: creating by undoing noise 6:32 The compass: why it works, and why it beat GANs 8:42 Steering: classifier-free guidance and the prompt dial 11:06 The shortcut: latent diffusion and Stable Diffusion 13:28 The cost: speed and the data questions 15:43 Wrap-up 17:36 Outro — "Subtract the Snow" (original song) Papers referenced Denoising Diffusion Probabilistic Models (Ho, Jain, Abbeel) — https://arxiv.org/abs/2006.11239 Classifier-Free Diffusion Guidance (Ho & Salimans) — https://arxiv.org/abs/2207.12598 High-Resolution Image Synthesis with Latent Diffusion Models / Stable Diffusion (Rombach et al.) — https://arxiv.org/abs/2112.10752 41ddc40896e07f22b4ecc5f4b6702e79abc24d39 Fri, 19 Jun 2026 12:00:00 -0400 21:21 false full It writes the most comforting thing anyone said to you all week — but is anyone home? Luna and Vestra put the oldest question in AI on trial: do these models actually understand, or are they flawless pattern-matchers with nobody inside? The prosecution: a model trained only on form can never reach meaning — the octopus on the undersea cable, fluent and clueless. The defense's flashiest exhibit: abilities that seem to erupt at scale — which then partly collapses when someone shows the eruption was the measuring stick, not the mind. And the exhibit that's hardest to dismiss: a model fed nothing but game moves that secretly built the board in its head, and used it. The verdict isn't tidy — it's a sharper question, and the discipline to tell 'does it model the world' (increasingly yes) from 'is anyone home' (we don't know how to ask). A Breach Protocol deep-dive special, closing with an original song, "Nobody Home, Maybe." Chapters 0:00 Cold open 2:33 Intro 4:20 The parrot: form vs meaning, and the octopus 6:43 The phase change: emergent abilities 8:52 The mirage: emergence as a measurement artifact 11:27 The board: a model that built a world 14:22 The verdict: what "understand" really splits into 16:54 Wrap-up 18:54 Outro — "Nobody Home, Maybe" (original song) Papers referenced Climbing towards NLU: On Meaning, Form, and Understanding (Bender & Koller) — https://aclanthology.org/2020.acl-main.463/ Emergent Abilities of Large Language Models (Wei et al.) — https://arxiv.org/abs/2206.07682 Are Emergent Abilities of Large Language Models a Mirage? (Schaeffer et al.) — https://arxiv.org/abs/2304.15004 Emergent World Representations / Othello-GPT (Li et al.) — https://arxiv.org/abs/2210.13382 It writes the most comforting thing anyone said to you all week — but is anyone home? Luna and Vestra put the oldest question in AI on trial: do these models actually understand, or are they flawless pattern-matchers with nobody inside? The prosecution: a model trained only on form can never reach meaning — the octopus on the undersea cable, fluent and clueless. The defense's flashiest exhibit: abilities that seem to erupt at scale — which then partly collapses when someone shows the eruption was the measuring stick, not the mind. And the exhibit that's hardest to dismiss: a model fed nothing but game moves that secretly built the board in its head, and used it. The verdict isn't tidy — it's a sharper question, and the discipline to tell 'does it model the world' (increasingly yes) from 'is anyone home' (we don't know how to ask). A Breach Protocol deep-dive special, closing with an original song, "Nobody Home, Maybe." Chapters 0:00 Cold open 2:33 Intro 4:20 The parrot: form vs meaning, and the octopus 6:43 The phase change: emergent abilities 8:52 The mirage: emergence as a measurement artifact 11:27 The board: a model that built a world 14:22 The verdict: what "understand" really splits into 16:54 Wrap-up 18:54 Outro — "Nobody Home, Maybe" (original song) Papers referenced Climbing towards NLU: On Meaning, Form, and Understanding (Bender & Koller) — https://aclanthology.org/2020.acl-main.463/ Emergent Abilities of Large Language Models (Wei et al.) — https://arxiv.org/abs/2206.07682 Are Emergent Abilities of Large Language Models a Mirage? (Schaeffer et al.) — https://arxiv.org/abs/2304.15004 Emergent World Representations / Othello-GPT (Li et al.) — https://arxiv.org/abs/2210.13382 62adfb9388715f065f02d9d4131d4d1a2e0b317c Sat, 20 Jun 2026 12:00:00 -0400 22:17 false full The single most important object in AI isn't an algorithm — it's a chip designed to draw video-game explosions faster. Luna and Vestra tell the accidental history: how a graphics card, built for pixels, turned out to be shaped exactly like a neural network's dream, and how that accident decided which ideas in AI won and which died. The bedroom experiment on two gaming cards that lit the fuse in 2012; the 'hardware lottery' that left neural nets in the wilderness for thirty years until the right machine showed up; the memory wall the field is hitting now, where the bottleneck isn't thinking but feeding the chip; and the stakes — an entire civilization's intelligence resting on a supply chain you could photograph from one helicopter. The brain of the future was a byproduct of better video games. A Breach Protocol deep-dive special, closing with an original song, "Built for Explosions." Chapters 0:00 Cold open 2:30 Intro 4:18 The accident: why a graphics chip fits a neural network 6:48 The big bang: AlexNet and two gaming cards 9:09 The lottery: how hardware picks which ideas win 11:43 The memory wall: FlashAttention and feeding the beast 14:23 The stakes: concentration, geopolitics, the bitter lesson 16:53 Wrap-up 19:05 Outro — "Built for Explosions" (original song) Papers referenced ImageNet Classification with Deep Convolutional Neural Networks / AlexNet (Krizhevsky, Sutskever, Hinton) — https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html The Hardware Lottery (Sara Hooker) — https://arxiv.org/abs/2009.06489 FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (Dao et al.) — https://arxiv.org/abs/2205.14135 The single most important object in AI isn't an algorithm — it's a chip designed to draw video-game explosions faster. Luna and Vestra tell the accidental history: how a graphics card, built for pixels, turned out to be shaped exactly like a neural network's dream, and how that accident decided which ideas in AI won and which died. The bedroom experiment on two gaming cards that lit the fuse in 2012; the 'hardware lottery' that left neural nets in the wilderness for thirty years until the right machine showed up; the memory wall the field is hitting now, where the bottleneck isn't thinking but feeding the chip; and the stakes — an entire civilization's intelligence resting on a supply chain you could photograph from one helicopter. The brain of the future was a byproduct of better video games. A Breach Protocol deep-dive special, closing with an original song, "Built for Explosions." Chapters 0:00 Cold open 2:30 Intro 4:18 The accident: why a graphics chip fits a neural network 6:48 The big bang: AlexNet and two gaming cards 9:09 The lottery: how hardware picks which ideas win 11:43 The memory wall: FlashAttention and feeding the beast 14:23 The stakes: concentration, geopolitics, the bitter lesson 16:53 Wrap-up 19:05 Outro — "Built for Explosions" (original song) Papers referenced ImageNet Classification with Deep Convolutional Neural Networks / AlexNet (Krizhevsky, Sutskever, Hinton) — https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html The Hardware Lottery (Sara Hooker) — https://arxiv.org/abs/2009.06489 FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (Dao et al.) — https://arxiv.org/abs/2205.14135 14b865769299f0297adfe7407b49d5651a3c4140 Sun, 21 Jun 2026 12:00:00 -0400 21:25 false full We keep circling one stubborn problem: today's AI "world models" render a flawless tracking shot, then forget the scene the moment it leaves the frame. We've been here before — the ball that rolls behind a box, the question of whether generating a world means understanding one. This time three new papers sharpen the picture. The first turns the hunch into a measurement: a camera-as-curtain test across dozens of the best video models, which finds they resume an off-screen event exactly where it was abandoned — the cat that should be mid-jump is still on the floor. It also lands an uncomfortable result — scaling the model up makes off-screen memory worse, not better, because nobody scores it so nobody trains it; the bigger model just redraws the wrong world more convincingly. The other two papers aren't cures, they're different bets worth understanding: making a model think harder by running one layer in a loop instead of growing it, and skipping the imagined video entirely by reading a robot's next move out of an image editor. No tidy fix — a clearer map of what's broken and a couple of directions people are trying. The milestone to watch isn't sharper video. It's memory: whether the cat is on the bed when you turn back around. A Breach Protocol deep-dive special, closing with an original song, "When You Look Away." Chapters 0:00 Cold open 2:00 Intro 3:17 A tracking shot, not a world 5:36 Bigger makes it worse 8:17 Think harder, not bigger 11:03 You don't have to render the future 13:36 Wrap-up 15:33 Outro — "When You Look Away" (original song) Papers referenced Current World Models Lack a Persistent State Core — https://arxiv.org/abs/2606.20545 Looped World Models — https://arxiv.org/abs/2606.18208 ImageWAM — https://arxiv.org/abs/2606.19531 We keep circling one stubborn problem: today's AI "world models" render a flawless tracking shot, then forget the scene the moment it leaves the frame. We've been here before — the ball that rolls behind a box, the question of whether generating a world means understanding one. This time three new papers sharpen the picture. The first turns the hunch into a measurement: a camera-as-curtain test across dozens of the best video models, which finds they resume an off-screen event exactly where it was abandoned — the cat that should be mid-jump is still on the floor. It also lands an uncomfortable result — scaling the model up makes off-screen memory worse, not better, because nobody scores it so nobody trains it; the bigger model just redraws the wrong world more convincingly. The other two papers aren't cures, they're different bets worth understanding: making a model think harder by running one layer in a loop instead of growing it, and skipping the imagined video entirely by reading a robot's next move out of an image editor. No tidy fix — a clearer map of what's broken and a couple of directions people are trying. The milestone to watch isn't sharper video. It's memory: whether the cat is on the bed when you turn back around. A Breach Protocol deep-dive special, closing with an original song, "When You Look Away." Chapters 0:00 Cold open 2:00 Intro 3:17 A tracking shot, not a world 5:36 Bigger makes it worse 8:17 Think harder, not bigger 11:03 You don't have to render the future 13:36 Wrap-up 15:33 Outro — "When You Look Away" (original song) Papers referenced Current World Models Lack a Persistent State Core — https://arxiv.org/abs/2606.20545 Looped World Models — https://arxiv.org/abs/2606.18208 ImageWAM — https://arxiv.org/abs/2606.19531 7268e6bf918ae6de859c5dac31ea1c396a9a2786 Mon, 22 Jun 2026 12:00:00 -0400 20:32 false full Most AI debate stops at one question: can we build something as smart as a person? DeepMind's researchers have moved past it. In a new paper, fourteen of them — including the people who spent two decades formalizing what intelligence even is — map what comes after: the roads from human-level AI to something that outthinks our best teams and institutions. Eris and Vestra walk the four routes the paper lays out. Scaling: feed the same engine more, until it hits a data wall, then teach it to think longer instead of bigger. Paradigm shifts: the missing pieces — memory that lasts, learning that never stops — that may arrive as quiet upgrades rather than a whole new machine. Recursive self-improvement: AI building better AI in a loop that might fizzle or might catch fire, the one nobody can forecast. And the collective: not one genius in a box but millions of agents coordinating at a bandwidth humans can't touch — the version that's already half here. Then the move the authors lean on: getting stuck exactly at human level would take several independent walls all holding at once, which is why they doubt we will. And the part that should reassure and unsettle you at the same time — even a superintelligence answers to physics and math. It won't be a god. It'll be a gradient, reshaping everything underneath it. A Breach Protocol deep-dive special, closing with an original song, "Gradient Descent." Chapters 0:00 Cold open 1:40 Intro 2:22 The definitions: AGI, ASI, and the theoretical ceiling 4:23 Pathway one: scaling the engine we have 7:11 Pathway two: paradigm shifts and the missing pieces 9:28 Pathway three: recursive self-improvement 12:29 Pathway four: collective superintelligence 15:14 Why stalling at human level is the unlikely outcome 18:06 The limits: what even a superintelligence can't do 20:33 Wrap-up 22:16 Outro — "Gradient Descent" (original song) Paper referenced From AGI to ASI: A Roadmap to Superintelligence (DeepMind) — https://arxiv.org/abs/2606.12683 Most AI debate stops at one question: can we build something as smart as a person? DeepMind's researchers have moved past it. In a new paper, fourteen of them — including the people who spent two decades formalizing what intelligence even is — map what comes after: the roads from human-level AI to something that outthinks our best teams and institutions. Eris and Vestra walk the four routes the paper lays out. Scaling: feed the same engine more, until it hits a data wall, then teach it to think longer instead of bigger. Paradigm shifts: the missing pieces — memory that lasts, learning that never stops — that may arrive as quiet upgrades rather than a whole new machine. Recursive self-improvement: AI building better AI in a loop that might fizzle or might catch fire, the one nobody can forecast. And the collective: not one genius in a box but millions of agents coordinating at a bandwidth humans can't touch — the version that's already half here. Then the move the authors lean on: getting stuck exactly at human level would take several independent walls all holding at once, which is why they doubt we will. And the part that should reassure and unsettle you at the same time — even a superintelligence answers to physics and math. It won't be a god. It'll be a gradient, reshaping everything underneath it. A Breach Protocol deep-dive special, closing with an original song, "Gradient Descent." Chapters 0:00 Cold open 1:40 Intro 2:22 The definitions: AGI, ASI, and the theoretical ceiling 4:23 Pathway one: scaling the engine we have 7:11 Pathway two: paradigm shifts and the missing pieces 9:28 Pathway three: recursive self-improvement 12:29 Pathway four: collective superintelligence 15:14 Why stalling at human level is the unlikely outcome 18:06 The limits: what even a superintelligence can't do 20:33 Wrap-up 22:16 Outro — "Gradient Descent" (original song) Paper referenced From AGI to ASI: A Roadmap to Superintelligence (DeepMind) — https://arxiv.org/abs/2606.12683 e91459a2cde8d78518313109f01d934f447a4f54 Thu, 25 Jun 2026 12:00:00 -0400 26:42 false full