The Curious Engineer

All blogs / AI / ML

Drew Breunig

dbreunig.com

Writing about AI, geo, culture, media, data, and the ways they interact.

AI / MLSoftware / engineering practice

’10’15’20’25
Posts
163 from 5 Mar 2010 to 24 Oct 2026
Posting
weekly (29 in the last year)
Kind
personal blog
Feed
https://www.dbreunig.com/feed.xml

2026 24

  1. The State of Context Engineering
  2. Context Inertia: When Agents Won’t Let Go
  3. What We Can Learn from Claude’s Fable 5.1 System Prompt
  4. Who Taught the Models to Do That?
  5. Fable & The End of the Free Lunch
  6. Harnesses are Situated Agents
  7. Manage Your Agent’s Loadout with Dr. Skill
  8. Understanding the Dynamics of the AI Ecosystem with Pace Layers
  9. The Problem is Prompt Debt
  10. What Do Humans Need From Docs?
  11. The Cost of Overfitting the Harness
  12. 10 Lessons for Agentic Coding
  13. Cybersecurity Looks Like Proof of Work Now
  14. How Claude Code Builds a System Prompt
  15. The 2nd Phase of Agentic Development
  16. The Cathedral, the Bazaar, and the Winchester Mystery House
  17. Learnings from a No-Code Library: Keeping the Spec Driven Development Triangle in Sync
  18. We’re Talking About Terms of Use, But the Issue is Embedded Judgment
  19. Two Beliefs About Coding Agents
  20. Why is Claude an Electron App?
  21. How System Prompts Define Agent Behavior
  22. The Potential of RLMs
  23. The Rise of Spec Driven Development
  24. A Software Library with No Code

2025 54

  1. 2025 in Review: Jagged Intelligence Becomes a Fault Line
  2. Why I Write (And You Should Too!)
  3. How Model Use Has Changed in 2025
  4. Enterprise Agents Have a Reliability Problem
  5. Don’t Fight the Weights
  6. Glimpses of the Future: Speed & Swarms
  7. Enterprise AI Looks Bleak, But Employee AI Looks Bright
  8. AI Companies School Like Fish
  9. Can Chatbots Accomodate Advertising?
  10. Making Sense of AI Job Titles
  11. Building Castles in the Air, but With Surprise Physics
  12. Bottleneck or Bisect: AI-Assisted Coding Will Change Product Management
  13. Initial Thoughts on GPT-OSS
  14. FLUX.1-Krea & the Rise of Opinionated Models
  15. Does the Bitter Lesson Have Limits?
  16. How Kimi K2 RL’ed Qualitative Data to Write Better
  17. How Kimi K2 Became One of the Best Tool-Using Models
  18. How Rewriting Training Data Improved Kimi K2’s Performance
  19. A Comedy Writer on How AI Changes Her Field
  20. Why “Context Engineering” Matters
  21. Delegation is the AI Metric that Matters
  22. AI Creates the Problems it Solves
  23. The Prompt Foreman
  24. Cat Facts Cause Context Confusion
  25. How to Fix Your Context
  26. Prompts vs. Context
  27. How Long Contexts Fail, 2025
  28. An Agentic Case Study: Playing Pokémon with Gemini
  29. The Drawbridges Go Up
  30. Let the Model Write the Prompt
  31. Claude’s System Prompt Changes Reveal Anthropic’s Priorities
  32. Using ‘Slop Forensics’ to Determine Model Ancestry
  33. A Small Model Just for Structured Output
  34. Will the Model Eat Your Stack?
  35. ChatGPT Heard About Eagles Fans
  36. StackOverflow’s LLM-Accelerated Decline
  37. Claude’s System Prompt: Chatbots Are More Than Just Models
  38. You Got Commands in My Prompt!
  39. DuckDB is Probably the Most Important Geospatial Software of the Last Decade
  40. The Wisdom of Artificial Crowds
  41. What We Mean When We Say “Think”
  42. The Dynamic Between Domain Experts & Developers Has Shifted
  43. AI Chatbots Are Like Observational Comics
  44. MCPs are APIs for LLMs
  45. Overcoming Bad Prompts with Help from LLMs
  46. Building an Easier to Use FFmpeg With LLMs
  47. What if AGI is Free?
  48. A Gentle Intro to Running a Local LLM
  49. DeepSeek as a Power Object
  50. Mistral Small & Human-Centric Benchmarks
  51. Is OpenAI Using Trap Tokens?
  52. On Authentic Data
  53. Your Eval is More Important Than the Model
  54. Beware the Cyren’s Song

2024 35

  1. Generating a Glossary from a Jekyll Blog Using DSPy & Claude
  2. On Test-Time Compute: The New Game in Town
  3. On Synthetic Data: How It’s Improving & Shaping LLMs
  4. Pipelines & Prompt Optimization with DSPy
  5. Why LLM Advancements Have Slowed: The Low-Hanging Fruit Has Been Eaten
  6. Turning Your Root URL Into a DuckDB Remote Database
  7. Will Surging Crypto Prices Damper Interest in AI?
  8. Transitioning from the Attention Era to the Automation Era
  9. AI Search Apps Succeed Because of Their UI, Not their LLMs
  10. Generating Descriptive Weather Reports with LLMs
  11. The 3 AI Use Cases: Gods, Interns, and Cogs, 2024
  12. Wikidata is a Giant Crosswalk File
  13. We Need Help With Content Discovery More Than Content Generation
  14. Conflating Overture Places Using DuckDB, Ollama, Embeddings, and More
  15. What Remains After an AI Collapse?
  16. When is a Minicar as Dangerous as a 3-Ton Truck?
  17. We Need an Ari Emanuel for Small OSS
  18. The Right to Confront Your AI Accuser
  19. Stewart Brand on Finding Where the Future is Being Made
  20. Domain Expertise is the Difference Maker for Data Scientists
  21. How to Build a Bigger Bubble
  22. Towards Standardizing Place, 2024
  23. Prompt Engineering at the End of the Universe
  24. Enriching CSVs with H3
  25. Smaller, Cheaper, Faster, Sober
  26. Be Better, Not Smaller
  27. Using DuckDB to Map Overture GERS IDs to US Census FIPs Codes
  28. The Eras of Domain Names
  29. Sober AI is the Norm
  30. A Plea for Sober AI
  31. Introducing StepList
  32. A POI Database in One Line
  33. The Rise of Audio AR
  34. Paying the AI Premium
  35. In Pursuit of Quiet AI Features

2023 14

  1. My CCPA Dialog With OpenAI
  2. Déjà Connu: AI Copilots Reveal Well-Traveled Paths
  3. A Big Year For Small Models
  4. A Foundation Model For Every Culture
  5. The Tyranny of the Blank Textbox
  6. Extreme Compression with AI: Fitting a 45 Minute Podcast into 40kbs
  7. The AI Market is Firming Up Fast
  8. Finding Bathroom Faucets with Embeddings
  9. Manipulating AI: Prompt Injection, Trap Tokens, & Post-Training Suggestion
  10. Considering AI Privacy Scenarios
  11. The Platypus In The Room
  12. AI Lies, Privacy, & OpenAI
  13. Practical AI For App Makers
  14. We Need AI Safewords

2022 1

  1. Why Media Metrics Matter

2020 3

  1. The Gauntlet Growing Social Networks Face Just Got Harder
  2. How to Build a Buzzword
  3. What are We Waiting For?

2016 10

  1. How the iPhone Led to Today’s Internet Outage
  2. Keeping Up with the Joneses, Today
  3. I Dream of the Post Office Buying Twitter
  4. Spectacles are the Anti-Glass
  5. Do Algorithms Find Depression or Cause Depression?
  6. Apple Shouldn’t Kill the Headphone Jack
  7. A Community Camera Manifesto
  8. The Business Implications of Machine Learning, 2016
  9. Motivations Behind Calls for Podcast Data
  10. Discussions We Will Have in 2016

2015 3

  1. Fake Can Be Just as Good
  2. Ways of Thinking About Products: Starter Kits
  3. What I’ve Learned About the Gold Rush So Far

2014 3

  1. On Convenience-Tech and Trusting Technology
  2. On Nuclear Weapons and Complex Systems
  3. The State of Technology in 2013: What We Learn from Cameras

2013 6

  1. Media Misbehavior Through the Ages
  2. The 5 Blows to the Human Ego
  3. Things Being Hacked
  4. Google Glass is Just Like the Segway – and Similarly Doomed
  5. Sports by Ideal Medium
  6. The “Remote President”

2012 3

  1. Habits are Cues for Innovation
  2. Things AR Should ‘See’
  3. Topics that will be Discussed on Antiques Roadshow, circa 2050

2011 6

  1. Frontiers through the Ages
  2. Multisignal Tools & Mutable Art
  3. Siri, Violin Vibratto, and Bing Crosby’s Croon
  4. The Idea of a Computer is Diffusing
  5. On Facebook Timeline: Teaching Data to Speak Humanely
  6. Loyal Customer Myopia

2010 1

  1. On Chat Roulette. Or, Never Judge a New Medium by its Early Adopters