RAG
166 episodes · Page 2 of 7
#5771: MCP Connectors: Talk to Your Data vs. Open Data
Same protocol, two patterns: pointing agents at your own database, and pointing them at public data — with very different results.
#5770: Talking to Your Data: What MCP Doesn't Solve
MCP standardized the pipe. It has no opinion about what flows through it — and that's where the hard part lives.
#5623: When Your AI Cites Sources It Didn't Actually Use
Your RAG pipeline cites a source — but did the model actually use it? New interpretability research says you can't tell from the trace.
#5507: Inside ChatGPT's Hidden Pipeline: 8 Models Per Reply
One chat turn isn't one model call. It's moderation, PII filtering, a tone-based router, and activation classifiers — here's the documented graph.
#5465: Packaging AI Pipelines So They Actually Get Reused
Daniel's podcast pipeline works — so why can't he reuse it? Recipes, containers, and Hebrew code-switching TTS.
#5408: When Small NLP Models Beat the LLM
Feature extraction, fill-mask, token classification — the classic NLP tasks still have a job. Here's when a small model beats a frontier API.
#5392: Building Agents You Can Actually Move
Agent portability isn't a copy job — it's a rebuild. Why memory, not code, is where lock-in lives.
#5328: A Shabbat Radio That Can't Be Paused
A soundproof box that plays podcasts continuously — no pause, no skip, no volume. Just open the lid and listen.
#5262: How the Talmud Outlived Every Book Burning
From the 1242 Paris burning to Sefaria's AI agents, the chain of transmission was never in the books — it was in the network.
#5230: Why Your Research Agent Won't Stop Searching
Most agent search runs add no new evidence — and 66-85% of the context window is just search snippets. Here's what actually stops the loop.
#5207: Hard Drives Are Having Their Biggest Boom in a Decade
Everyone assumes flash killed the spinning disk. Seagate's 52% margins and sold-out capacity through 2028 say otherwise.
#5165: Weighting Memory in a RAG Pipeline
Daniel's memory leg won't fire. Turns out system prompts have no inherent vector-space privilege — but retrieval does have a real knob.
#4958: Multi-Model Transcription with LLM Reconciliation
Using multiple transcription engines and an LLM judge to catch errors in voice-to-text — especially tricky technical terms.
#4809: Can a Podcast Train Itself?
Can a podcast become a self-developing project? We explore RAG vs. fine-tuning for lore memory.
#4786: Markdown vs Vectors for Agent Memory
Why coding agents need two kinds of memory — and why mixing them creates a sync nightmare.
#4704: Building a Unified AI Filesystem with Rclone and MinIO
How to build a single virtual filesystem for AI agents across multiple cloud storage providers — without the token headaches.
#4701: Why Claude Code Re-Introduces Itself to Your Repo
Claude Code remembers your preferences but not your codebase. We explore why that's by design and how to work around it.
#4688: Why 3,000 Episodes Vanished From Spotify (And Came Back)
The December episode wasn't deleted — Spotify just wasn't told about it. Here's how 4,600 episodes reappeared.
#4682: When Context Windows Grow: What Survives?
If context windows jump to 10M tokens, which AI engineering practices survive? We separate load-bearing architecture from temporary scaffolding.
#4648: Graph Engineering: AI Memory Beyond Retrieval
How linked notes and knowledge graphs are turning AI agents from memory tools into discovery engines.
#4540: The One-Star Review Circuit: A Master Playbook
How to find the worst-rated restaurants in any city—and actually enjoy the experience.
#4523: Can You Trust an AI's Summary?
Dedicated text compressors exist but aren't in production. The real problem? Nobody can verify the summary didn't drop what mattered.
#4492: What Makes a Container Robot-Ready?
The hidden specs that turn a plastic box into something a robot can pick, read, and move at industrial speed.
#4458: How Headroom Compresses Agent Context Without Breaking Outputs
A deep dive into Headroom's content-aware compression pipeline that promises 20-95% token savings without changing agent answers.