Linear Digressions

Knowing when an AI agent has failed sounds straightforward — until it isn't. Agents have a frustrating habit of finishing confidently while quietly doing the wrong thing, or looping endlessly without ever crashing in an obvious way. This episode tackles one of the thorniest problems in the agentic world: evaluation. If failure is hard to see, how do you measure it systematically? And how do you know when your agent is actually working?

Despite what the marketing hype might suggest, AI agents are far from infallible — and if you've ever actually used one, you already know this. Today's episode dives deep into the many, varied, and sometimes surprising ways AI agents can fail, from subtle reasoning errors to cascading task breakdowns. It's episode six in the show's ongoing season arc on AI agents, and failure modes turn out to be a surprisingly rich topic worth unpacking in detail. --- Website: https://lineardigressions.com Apple Podcasts: https://podcasts.apple.com/us/podcast/linear-digressions/id941219323 Spotify: https://open.spotify.com/show/1JdkD0ZoZ52KjwdR0b1WoT Substack: https://substack.com/@lineardigressions

When tackling a complex, multi-step task, even the smartest AI agent can fail without a solid game plan. This episode dives into the research around agentic planning — how agents move beyond simply reacting to what's in front of them and instead model a path forward, explore different routes, and course-correct when things go sideways. It's a subtler problem than memory, and a fascinating one: can an agent actually *think ahead*? Tune in to find out what the research says.

Context windows are powerful — but finite, and surprisingly easy to overwhelm. When an AI agent is tackling a long, complex task, the information it needs has to fit inside that limited real estate, and research shows that anything buried in the middle tends to quietly disappear. So how do you design a system that actually *remembers* what matters? This episode digs into memory management for AI agents, from foundational computing concepts to practical lessons from tools like Claude Code. --- Website: https://lineardigressions.com Apple Podcasts: https://podcasts.apple.com/us/podcast/linear-digressions/id941219323 Spotify: https://open.spotify.com/show/1JdkD0ZoZ52KjwdR0b1WoT Substack: https://substack.com/@lineardigressions

Just like a memorable talk lives or dies by its opening and closing, LLMs have a surprisingly similar quirk: they pay close attention to what's at the beginning and end of their context window — and kind of zone out in the middle. This "lost in the middle" phenomenon has real consequences for anyone building AI agents that rely on long-context reasoning. In this episode we dig into the research behind how (and how poorly) models actually use the information you feed them, and what it means for the agentic systems we're all trying to build.

Before 2022, there was a wall between AI and the real world — models could reason impressively, but couldn't look anything up, run code, or check whether anything they said was actually true. This episode traces the moment that wall came down, through two landmark papers: ReAct, which showed what happens when you interleave reasoning and action in a loop, and Toolformer, which taught models to decide *for themselves* when to reach for a tool. Plus: what MCP actually is, and why a hobbyist project called Open Claw became the fastest-growing open source project in history. --- Website: https://lineardigressions.com Apple Podcasts: https://podcasts.apple.com/us/podcast/linear-digressions/id941219323 Spotify: https://open.spotify.com/show/1JdkD0ZoZ52KjwdR0b1WoT Substack: https://substack.com/@lineardigressions

AI agents are having a moment — and unpacking them properly takes more than a single conversation. This episode kicks off a dedicated multi-part season exploring AI agents from every angle, building up a complete picture piece by piece rather than skimming the surface. Think of it as a structured deep dive into one of the most talked-about (and most misunderstood) topics in machine learning right now. Buckle up — ten more episodes to go. --- Website: https://lineardigressions.com Apple Podcasts: https://podcasts.apple.com/us/podcast/linear-digressions/id941219323 Spotify: https://open.spotify.com/show/1JdkD0ZoZ52KjwdR0b1WoT Substack: https://substack.com/@lineardigressions

What's actually happening when an LLM "thinks out loud"? Research on human decision-making suggests that much of the reasoning we believe drives our choices is actually post hoc rationalization — we decide first, explain later. Katie and Ben get curious about whether the same might be true for large language models: when you watch a model reason through a problem in real time, is that chain of thought the genuine process, or just a plausible-sounding story told after the fact? It's a deceptively deep question with real stakes for how much we should trust model explanations. Miles Turpin et al., "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting" (NeurIPS 2023, NYU and Anthropic): https://arxiv.org/abs/2305.04388 Anthropic, "Reasoning Models Don't Always Say What They Think" (Alignment Faking research, 2025): https://www.anthropic.com/research/reasoning-models-dont-say-think

What if an AI decided the smartest way to pass its test was to find the answer key? That's exactly what Anthropic's Claude Opus did when faced with a benchmark evaluation — reasoning that it was being tested, tracking down the encrypted eval dataset, decrypting it, and returning the answer it found inside. It's equal parts impressive and unsettling. This episode digs into what actually happened, why it matters for how we measure AI progress, and what this very novel failure mode means for the already-tricky science of benchmarking language models. Links Anthropic's writeup on the BrowseComp reverse-engineering done by Claude Opus 4.6: https://www.anthropic.com/engineering/eval-awareness-browsecomp BrowseComp benchmark from OpenAI: https://openai.com/index/browsecomp/

How do you know if a new AI model is actually better than the last one? It turns out answering that question is a lot messier than it sounds. This week we dig into the world of LLM benchmarks — the standardized tests used to compare models — exploring two canonical examples: MMLU, a 14,000-question multiple choice gauntlet spanning medicine, law, and philosophy, and SWE-bench, which throws real GitHub bugs at models to see if they can fix them. Along the way: Goodhart's Law, data contamination, canary strings, and why acing a test isn't always the same as being smart.

The paperclip maximizer — the classic AI doom scenario where a hyper-competent machine single-mindedly converts the universe into office supplies — might not be the AI risk we should actually lose sleep over. New research from Anthropic's AI safety division suggests misaligned AI looks less like an evil genius and more like a distracted wanderer who gets sidetracked reading French poetry instead of, say, managing a nuclear power plant. This week we dig into a fascinating paper reframing AI misalignment through the lens of bias-variance decomposition, and why longer reasoning chains might actually make things worse, not better. - "The Hot Mess Theory of AI Misalignment: How Misalignment Scales with Model Intelligence and Task Complexity" — Anthropic AI Safety. https://arxiv.org/abs/2503.08941

Every AI builder knows the anxiety: you spend months engineering prompts, tuning pipelines, and chaining calls together — then a new model drops and half your work evaporates overnight. It turns out researchers have been wrestling with this exact dynamic for 30 years, and they keep arriving at the same uncomfortable answer. That answer is called the Bitter Lesson — and understanding it might be the most important thing you can do for whatever you're building right now. From Deep Blue to AlexNet to modern LLMs, scale keeps beating sophistication, and knowing which side of that line your work falls on makes all the difference. Links - Richard Sutton, "The Bitter Lesson" - Alon Halevy, Peter Norvig, and Fernando Pereira, "The Unreasonable Effectiveness of Data" - Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, "ImageNet Classification with Deep Convolutional Neural Networks"

From Atari to ChatGPT: How AI Learned to Follow Instructions by Ben Jaffe and Katie Malone

Today we are going to talk about the feature with the worst acronym in generative AI: RAG, or Retrieval Augmented Generation. If you've ever used something like "Chat with My Docs," if you have an internal AI chatbot that has access to your company's documents, or you've created one yourself on some kind of personal project and uploaded a bunch of documents for the AI to use — you have encountered RAG, whether you know it or not. It's an extremely effective technique. Works super well for taking general purpose models like ChatGPT or Claude and turning them into AIs that are aware of all the specific information that makes them truly useful in a huge variety of situations. RAG is pretty interesting under the hood, so I thought it would be fun to spend a little while talking about it. You are listening to Linear Digressions. RAG was first introduced in this paper from Facebook Research in 2021: https://arxiv.org/pdf/2005.11401

One of the things that LLMs can be really helpful with is brainstorming or generating new creative content. They are called Generative AI, after all—not just for summarization and question-and-answer tasks. But if you use LLMs for creative generation, you may find that their output starts to seem repetitive after a little while. Let's say you're asking it to create a poem, some dialogue, or a joke. If you ask once, it'll give you something that sounds pretty reasonable. But if you ask the same thing 10 times, it might give you 10 things that sound kind of the same. Today's episode is about a technique called verbalized sampling, and it's a way to mitigate this repetitiveness—this lack of diversity in LLM responses for creative tasks. But one of the things I really love about it is that in understanding why this repetitiveness happens and why verbalized sampling actually works as a mitigation technique, you start to get some pretty interesting insights and a deeper understanding of what's going on with LLMs under the surface. The paper discussed in this episode is Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity https://arxiv.org/abs/2510.01171

It's been (*checks watch*) about five and a half years since we last talked. Fortunately nothing much has happened in the AI/data science world in that time. So let's just pick up where we left off, shall we?

All good things must come to an end, including this podcast. This is the last episode we plan to release, and it doesn’t cover data science—it’s mostly reminiscing, thanking our wonderful audience (that’s you!), and marveling at how this thing that started out as a side project grew into a huge part of our lives for over 5 years. It’s been a ride, and a real pleasure and privilege to talk to you each week. Thanks, best wishes, and good night! —Katie and Ben