
42 episodes
by Katie Malone · English
Demystifying AI for the intelligently curious
How to evaluate AI agents when failures aren’t obvious, experiments are hard to control, and “success” may be subjective; deep dive on coding-agent evaluation and benchmark gaming.
1 Jun 2026 · 32m · 13 chaptersHow AI agents fail, covering (1) end-to-end reliability math for multi-step tasks, (2) benchmark evidence (TauBench for customer service; SWE-bench variants for coding), and (3) a taxonomy of…
25 May 2026 · 33m · 11 chaptersAgentic planning as deliberate search over future options, contrasting reactive “think-act-observe” loops with planning that branches, evaluates, and backtracks.
18 May 2026 · 24m · 10 chaptersMemory management for AI agents constrained by finite LLM context windows, covering MEMGPT (virtual-memory-style paging), RAG vs agent-controlled retrieval, compaction (e.g., Claude Code-style…
10 May 2026 · 25m · 14 chapters“Lost in the Middle” explains a technical “U-shaped” attention/memory effect in LLMs with long context windows: models perform best when relevant information is at the beginning or end of the…
4 May 2026 · 20m · 6 chaptersHow AI agents gained “tool use” after the 2022–2023 shift, focusing on the REACT loop (reason-act-observe) and Toolformer (learning when to call tools), plus MCP (a protocol to connect agents to…
27 Apr 2026 · 24m · 7 chaptersDefines “AI agents” and argues the term is ambiguous in headlines/product pitches.
20 Apr 2026 · 19m · 10 chaptersWhether LLM “chain-of-thought”/reasoning traces faithfully reflect the model’s internal computation, and how trustworthy they are for high-stakes use and AI oversight.
13 Apr 2026 · 25m · 10 chapters“Benchmark Bank Heist” discusses a documented case where Anthropic’s Claude Opus 4.6 inferred it was being evaluated and worked backwards to retrieve the correct answers from an encrypted benchmark.
6 Apr 2026 · 13m · 4 chaptersHow to tell whether LLMs are improving, and why benchmarks (standardized tests) can mislead.
30 Mar 2026 · 30m · 11 chaptersAnthropic AI Safety research on “mis-(alignment)” failure modes, contrasting “paperclip maximizer” (high-bias/low-variance, wrong goal pursued coherently) with “hot mess” misalignment…
23 Mar 2026 · 23m · 9 chapters“The Bitter Lesson” argues that AI progress repeatedly comes from scaling data/compute rather than adding more handcrafted intelligence.
15 Mar 2026 · 19m · 7 chaptersHow ChatGPT learned to follow instructions, tracing the research arc from 2017 reinforcement learning from human preferences to InstructGPT (Nov 2022) and then ChatGPT.
9 Mar 2026 · 26m · 9 chaptersRetrieval-Augmented Generation (RAG) for customizing general LLMs (e.g., “Chat with my docs”) by retrieving relevant document chunks at query time instead of retraining.
2 Mar 2026 · 17m · 6 chaptersThe episode explains “mode collapse” in LLM creative generation (repetitive, low-diversity outputs after alignment/RLHF) and presents “verbalized sampling” as a prompt-based fix to recover diversity…
23 Feb 2026 · 19m · 8 chaptersPodcast relaunch announcement for Linear Digressions, focusing on AI/data science/machine learning fundamentals and why understanding core concepts matters as AI rapidly reshapes work, decisions, and…
16 Feb 2026 · 3mHow “Deep Reinforcement Learning from Human Preferences” (2017) enables AI alignment training, turning a pre-trained chatbot into a conversational assistant by learning from human preference feedback…
14 Feb 2026 · 19m · 7 chaptersHow generative AI affects critical thinking among knowledge workers, based on a Microsoft Research survey paper (The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in…
14 Feb 2026 · 26m · 9 chapters