Video summary

Nicholas de Leon’s AI Newsroom

Main summary

Key takeaways

News and Commentary

Nicholas De Leon (Consumer Reports): AI-driven “AI newsroom” experiments

Nicholas De Leon discusses his AI-driven “AI newsroom” experiments, centered on a project called Tucson Daily Brief—an automated local-news site and content pipeline designed to keep up with city/county coverage that has largely disappeared from traditional local journalism.

Main points and analyses

How Tucson Daily Brief started (automation + experimentation)

  • De Leon began after noticing interest in “OpenAI/Claw”-type automation tools earlier in the year.
  • His goal wasn’t “world domination,” but a practical experiment in AI plus journalism.
  • The initial version largely aggregated ~30 sources (RSS and light scraping), and overnight selected and summarized the top 10 stories across categories.
  • It published each morning (around 6:00 a.m. Arizona time) with links back to the original outlets.

Scaling from aggregation to “real journalism”

  • He wanted to go beyond summarizing other media and add reporting value using tools that can create new content from public records and streams.
  • He describes a “tool-assisted speedrun” approach: instead of manual work, the system uses LLMs, transcriptions, public-data scrapers, scheduling, and Python glue to generate outputs while he sleeps.

“Live AI reporter” for town halls (most distinctive feature)

  • He covers multiple municipalities including the city of Tucson, Pima County, and nearby towns, with plans to add school districts.
  • Town halls are streamed publicly. He captures audio in real time using FFmpeg, transcribes via Deepgram, then sends the transcript to Anthropic (Sonnet 4.6 mentioned) to generate an AP-style ~1000-word news article.
  • He edits manually to catch transcription errors (notably mangled Spanish names).
  • Turnaround is fast—stories can go up within about an hour, sometimes before other local outlets.
  • He emphasizes this is not stealing: it uses public streams and public records and links back to primary sources.

Blending AI drafting with human verification (“semi-automated”)

  • For some stories (including town-hall-derived pieces), AI drafts while De Leon performs a quick quality pass.
  • He frames his role as ensuring outputs aren’t “crazy,” and says he wouldn’t be comfortable fully publishing AI-only work without oversight.

Case example: integrating AI with traditional reporting (data-center election story)

  • He cites a New York Times profile of Marana and an election related to data center opposition.
  • De Leon describes co-writing a follow-up story “with AI,” using:
    • county recorder vote totals,
    • context from his own prior reporting about data centers and town halls,
    • AI for synthesis.
  • He presents the result as balanced and factual, not a simplistic “anti-data center lost” narrative.

Distribution: multi-platform content from the same pipeline

Daily briefs become:

  • a newsletter
  • a podcast (voice cloned using 11Labs)
  • YouTube content including shorts
  • social posts (Instagram/BlueSky noted; X described as more manual)

He spends significant time on social/media workflows (visual creation and captions) because some platforms can’t be fully automated.

Voice cloning with 11Labs (localized delivery)

  • He cloned his own voice to read the briefs.
  • He notes needing custom phonetic “interrupt” files for Spanish names and symbol rendering issues (e.g., dollar sign handling).

ChatTDB: a RAG tool for local Q&A

  • He introduces chat.tdb.com (ChatTDB) as a retrieval-augmented generation (RAG) system that refreshes from his databases (including items like liquor license info).
  • The goal is to answer questions such as: what did a particular town focus on regarding data centers, based on his corpus.

Cost and model selection

  • He estimates costs around $30–$40/month, with 11Labs as the largest cost contributor.
  • He argues for using cheaper models where possible, reserving more capable ones only when needed.

Other projects (broad AI-for-media experimentation)

  • Crossword-style news site (“crossword the situation.com” adapted locally): a small crossword puzzle tied to recent headlines, including Tucson-specific clues.
  • DeepDugout.com: an AI baseball simulation where models act like managers; stronger reasoning models show diminishing returns beyond a point.
  • Matchdayinference.com: personalized World Cup “matchday programs” per user preferences/lenses using sports data plus context.
  • MyNextAssignment.com: job-finder tool for media/editorial roles by scanning job boards, with options to tailor resumes/cover letters for editing and submission.

Overall takeaway

De Leon presents his work as a demonstration of what AI can do for local journalism infrastructure: rapidly turning public records and livestreams into publishable stories, while maintaining human editorial checks. The broader theme is that AI can help restore coverage that local news organizations struggle to sustain—especially for routine governance events—by automating the “commodity” parts of reporting.

Presenters / contributors

  • Leo Laporte
  • Paris Martineau
  • Nicholas De Leon (Consumer Reports)

Original video