How I AIHow to digest 36 weekly podcasts without spending 36 hours listening | Tomasz Tunguz
CHAPTERS
- 0:00 – 1:00
The “36 podcasts a week” problem and a daily podcast-ripping solution
Tomasz explains why he wants to read podcasts instead of listening and introduces his system that automatically downloads and transcribes dozens of shows. The goal is to turn long-form audio into quickly skimmable text and actionable insights without spending 36 hours a week.
- •Motivation: too many valuable podcasts, too little time
- •Preference for reading to skim, search, and skip ahead
- •Daily automation to download episodes and transcribe them
- •End output is not just transcripts, but structured takeaways
- 1:00 – 4:05
Show context + sponsor segment (Notion) and episode framing
Claire introduces the show, Tomasz’s background, and tees up the live walkthrough of his workflow. This section also includes the Notion sponsor message and a channel giveaway announcement.
- •Who Tomasz is (Theory Ventures, widely followed blogger)
- •What viewers will learn: podcast digestion + AI drafting workflow
- •Notion ad: AI meeting notes, enterprise search, research mode
- •Giveaway mechanics and how to enter
- 4:05 – 5:14
System architecture: Parakeet Podcast Processor, FFmpeg, Whisper/Parakeet transcription
Tomasz breaks down the core components of his pipeline: ingesting feeds, downloading audio, converting formats, and generating transcripts. He contrasts earlier Whisper-based transcription with Nvidia’s Parakeet running locally on a Mac.
- •Processor reads podcast file/feed and downloads episodes
- •FFmpeg converts audio into the right format for transcription
- •Transcription options: Whisper vs Nvidia Parakeet (local Mac performance)
- •Outputs are fed into prompts to generate requested summaries/structures
- 5:14 – 6:14
Cleaning transcripts with an LLM (Gemma) and when it matters
A hands-on look at the transcript cleanup step: removing filler words while preserving technical content and length. Tomasz explains that cleanup used to be critical for downstream extraction, but became less necessary after switching to more capable LLM-based extraction.
- •Prompting Gemma to act as a transcript editor (remove ums/ahs, keep content)
- •Batch processing multiple transcripts per day
- •Cleaning helped earlier ML/NLP extraction (proper nouns, entity recognition)
- •With stronger LLMs, transcript cleanliness matters less than before
- 6:14 – 7:15
Orchestration + storage: DuckDB tracking and daily summary document generation
Tomasz shows how he orchestrates processing, tracks what was processed each day, and compiles outputs into a single daily digest. DuckDB serves as a lightweight local system of record for episodes and transcripts.
- •Podcast Orchestrator runs the end-to-end daily pipeline
- •DuckDB stores processing metadata (episode/date processed)
- •System pulls all daily transcripts and runs a summarizer prompt
- •Generates a dated daily summary file for quick consumption
- 7:15 – 10:19
Extracting what’s actionable: quotes, theses, observations, tweet drafts, and company mentions
The digest is structured into parts Tomasz can directly use: notable quotes, venture investment theses, and company/startup mentions for enrichment. He also experiments with turning observations into tweet drafts and automating relationship-building via links/mentions.
- •Comprehensive summary plus key topics/themes per episode
- •High-signal quotes as the most valuable output
- •Actionable investment theses to kick off market maps and internal work
- •Company/startup mentions sent to CRM for enrichment
- •Noteworthy observations drafted into potential tweets (still iterating)
- 10:19 – 12:29
Why the terminal: latency, scripting power, and Claude Code as a force multiplier
Claire asks why Tomasz stays in the terminal instead of building a UI. Tomasz argues terminal workflows are faster (low latency) and more scriptable, and he highlights Claude Code as a tool for rapidly modifying large sets of files and building utilities on top of his blog corpus.
- •Terminal has low interaction latency, reducing frustration
- •Comfort and speed gained since learning it during COVID
- •Scripting enables batch actions and automations (email, CRM actions)
- •Claude Code helps refactor/modify many files and build new tools quickly
- 12:29 – 14:21
Personalized software vs off-the-shelf: building a “glove-like fit” workflow
Claire reframes the system as a prime example of hyper-personalized software—something that wouldn’t be worth productizing for a mass market but is extremely valuable for an individual workflow. Tomasz emphasizes how quickly he can adjust the pipeline as needs change.
- •Many want “podcast digest apps,” but not with this exact UX/workflow
- •Owning the full stack enables tailoring to personal preferences
- •Fast iteration: fix ordering issues and add features in seconds/minutes
- •AI lowers the cost/friction of building custom utilities
- 14:21 – 16:08
Sponsor break (Miro) and pivot to podcast-to-blog workflows
After a Miro sponsor segment, Claire recaps what the system already extracts and transitions into Tomasz’s second workflow: using podcast insights to generate blog post drafts. The emphasis shifts from digestion to content creation.
- •Miro ad: AI copilot in canvas for briefs, diagrams, prototypes
- •Recap: summaries, themes, quotes, companies, tweet drafts
- •Transition: turning extracted insights into blog post drafts
- •Set up for demoing the blog generator pipeline
- 16:08 – 17:39
Generating a blog post from a specific podcast: context, retrieval, and a demo hiccup
Tomasz walks through generating a post based on a podcast (e.g., GitHub CEO discussion on AI/coding). He describes passing the transcript as context, categorizing content, and retrieving related materials—while noting a live bug in his local vector search (LanceDB).
- •Use case: write a post anchored to a quote/idea from a podcast
- •Transcript supplied as context + an output file target
- •Search for related items via categorization (e.g., AI)
- •Local vector DB (LanceDB) used for retrieval; demo fails due to a bug
- 17:39 – 18:25
The “AP English teacher” grading loop: letter grades, iteration, and why A- is the bar
Tomasz introduces his most effective writing-improvement trick: have the model grade the draft like an AP English teacher, then iterate until it reaches an A-minus. He explains a three-iteration process and why hook and conclusion matter most for reader experience.
- •Ask the model to grade drafts like an AP English teacher
- •Iterate until reaching ~A- (around 90–91)
- •Three-pass loop: grade → revise → grade, focusing on hook and conclusion
- •Uses critique models (e.g., Gemini critiquing Claude) to refine prompts
- 18:25 – 26:20
Capturing Tomasz’s writing style: using 2,000 prior posts, dynamic style patterns, and constraints
The generator retrieves Tomasz’s prior blog posts to mimic topic-specific style, but he notes voice matching is still very hard. He shares concrete constraints (word count, no headers, short paragraphs) derived from analyzing reader behavior, plus the challenge of internal linking and AI’s “too perfect” grammar.
- •Only his blog corpus (not podcasts) is embedded for style/context
- •Dynamic extraction of stylistic patterns from similar posts
- •Constraints: ~500 words, no section headers (headers hurt dwell time)
- •Short paragraphs (often two long sentences), smooth transitions desired
- •Challenges: AI voice feels “computer-like,” struggles to link to other posts
- 26:20 – 28:16
The grading rubric, variability, and the explore/exploit pattern across iterations
Tomasz shares the grading prompt criteria and discusses how scores can vary—especially when grading his dictated drafts versus AI-generated ones. The three-loop process can dip then recover as the model explores changes and then tightens for brevity, with transitions being the most persistent weakness.
- •Grading criteria: hook, clarity, evidence, structure, conclusion, engagement
- •Models can be overly positive; still sees Bs/Cs depending on input source
- •Dictated human drafts can score much lower (e.g., C-)
- •Iteration variability: first score high, second dips, third improves
- •Common failure mode: overly verbose transitional text
- 28:16 – 35:14
AI in writing education + closing: tiny teams, model competitions, and where to find Tomasz
They discuss AI as a first-pass writing evaluator in education, freeing teachers to focus on creativity once grammar/logic is handled. The episode closes with Tomasz’s view of a future 30-person $100M company, his tactic of having models “duke it out,” and final plugs.
- •AI as a fast feedback loop for students: grade-first, not write-for-me
- •Teachers can focus on creativity after mechanical issues are addressed
- •Vision: small, engineer-heavy $100M companies powered by PLG + automation
- •Prompting tactic: pit models against each other to improve outputs
- •Where to find Tomasz (tomtunguz.com) and invitation for AI founders
