Before my son arrived, I had an “AI addiction” problem that consumed nearly all my free time outside of work.
A $200/month Claude Max subscription burning through $1,500+/month in API-equivalent value, with ChatGPT, Gemini, and Grok subscriptions on the side. 400+ Claude Code/Codex sessions in the past 30 days. 20+ AI agents built and running in orchestrated teams for all aspects of my personal and professional life. 100+ scientific journals and news sources summarized and ready for my review every day. 200+ podcast episodes on AI saved on my playlist. Within 1 hour of a new model like GPT-5.4 dropping, I ran a series of evals on Codex and had my “personal benchmark” against Opus 4.6.
I thought I was living in the future, but I was more anxious than ever. AI is moving so fast, and I was gripped by a deep FOMO, the fear that if I slowed down for even a week, I’d fall behind permanently.
And I had a plan for my unborn son. Before he took his first breath, I was going to build an AI knowledge system to document his entire life. I extracted and synthesized six parenting books, plus everything relevant on PubMed, into a 37-file, 7,600-line “parenting knowledge base” for a dedicated AI agent. I even started scoping an AI companion (“openclaw”) for my son: a longitudinal experiment in human-AI co-development, beginning at birth.
Then my son arrived in early March. And he had absolutely no interest in my infrastructure.
In the fog of 4am feedings, failed sleep schedules, and a crying baby who hadn’t followed the script of any of my carefully curated knowledge bases, my Claude usage dropped 87% overnight. Between diaper changes, I found myself back in real life, reflecting on what actually matters from my AI frenzy.
Here’s what I learned.
Learning 1: AI Captures a Lot, But the Residuals Are Often the Signal
Before my son was born, I collected six parenting books, from AAP to Montessori to best-sellers. I extracted the key evidence from each and organized findings by topic: sleep training, feeding schedules, soothing techniques, developmental milestones.. I cross-referenced contradictions across all six and built a structured knowledge base with 37 files and comparison tables.
I was running PCA on parenting, taking a massive, high-dimensional problem and reducing it to its principal components. The consensus recommendations. The standard advice.
For the non-data people: Principal Component Analysis compresses complex data into a handful of dimensions that capture the most information. The leftover information, the residuals, gets discarded. Usually that’s fine. The residuals are noise.
But sometimes the residuals contain exactly the signal you need.
My parenting knowledge base was comprehensive. But it was also useless at 3am when my son was screaming and none of the principal components were working. What worked was a specific bouncing rhythm I discovered through three sleepless nights. A particular angle of holding him that no book mentioned. One specific white noise, not the others.
These were residuals. Tiny, specific signals that no compression algorithm would preserve. And they were exactly the things that mattered in my specific context: comforting my baby in a specific situation.
This is the dimension reduction problem at the core of AI. AI is extraordinary at capturing explicit knowledge, the stuff that can be written down, searched, and summarized. What are the AAP safe sleep guidelines? What’s the evidence on sleep training? Ask it anything that’s been extensively documented, and it performs at a level that would have seemed impossible 2 years ago.
But there’s a whole category of knowledge that lives outside the training data entirely. Tacit knowledge. The feel for when a baby is sick versus just fussy. The judgment call about whether to wake a sleeping baby to feed. The instinct that something is “off” before you can articulate why. This knowledge doesn’t transfer through text. It doesn’t compress into a context window. It lives in the body, the hands, the nervous system of the person who has done the work.
Michael Polanyi, the philosopher who coined the term, put it simply: “We can know more than we can tell.” If you can’t put it in language, it can’t enter the training data. If it’s not in the training data, the model can’t learn it.
Through my experience, AI without specific context captures maybe 50–70% of what you need in domains without clearly defined evaluation criteria. The generic direction, the standard protocol, the consensus view. That’s genuinely valuable. But the remaining 30–50% is where the real signal hides: the edge cases, the situational judgment, the muscle memory, the pattern recognition built from years of getting it wrong.
The plumber who charges $1,000 to tap one pipe isn’t charging for the tap. He’s charging for the 20 years of tapping wrong pipes that taught him which one to tap. The best biotech investors and operators don’t just have information to make decisions. They also have process knowledge from years of experience. They know which KOL’s opinion to weight and which to discount in a given scenario. They know that a 40% response rate in one clinical trial with a defined patient population and baseline may or may not translate to another trial.
AI raises the floor. Your tacit knowledge is the ceiling. The distance between them is where the real value lives.
Learning 2: The Fake Productivity Trap. When AI Creates Work That Doesn’t Matter
In addition to my baby knowledge base, I also built a “growth tracker terminal” for my baby. It tracked all possible analytics every day: feeding, diaper changes, sleep, developmental milestones. It integrated with my knowledge base through short-term and long-term “agentic memory.” I could pull data and generate nice visualizations whenever needed, and the AI agent would give me customized advice daily.
Nobody uses it.
My wife likes the Huckleberry app. Tap, tap, done. My mother-in-law isn’t tech-savvy and prefers a whiteboard on the fridge. When the pediatrician asks how things are going, they don’t want to see my fancy visualization. They just want the numbers to do a gut check.
I built a Bloomberg Terminal for a baby. The market participants prefer Post-it notes.
Here’s what actually works in our house: the whiteboard. Tally marks for feedings. A note about the last diaper change. That’s it. Everyone in the family can read it, update it, and act on it. At 3am, in the dark, without logging into anything. I spent 10 hours on my fancy knowledge base. The whiteboard cost $15.
In the podcast AI Daily Brief, host Nathaniel Whittemore calls this “workslop”: “AI is revealing just how much of work is people doing things pretty much just to be seen to be doing things.”
My knowledge base wasn’t productivity. It was fake productivity: automation for the sake of automation, complexity for the sake of complexity.
This pattern extends far beyond baby dashboards. Think about the AI-generated morning briefings that land in your inbox, beautifully formatted summaries of news you never read. The AI-polished emails where both sender and recipient spend extra time crafting and parsing elevated language that adds no information. The automated Slack summaries, the Zoom AI meeting notes, the generated reports. Each one individually reasonable, collectively creating a rising tide of polished outputs without necessarily increasing productivity.
Every one of these passes the “Can AI do this?” test. But the real question is: does AI actually increase productivity, or does it create the illusion of productivity while consuming more time from everyone?
Learning 3: The First Time Is Sacred. Why AI Can’t Automate Experience
AI can automate your 100th competitive analysis. It can templatize your 50th earnings summary. Once you know the pattern and give exact instructions, AI is extraordinary at replicating it at scale.
But it cannot automate the first time.
My first diaper change in a hospital room with a crying baby and no healthcare professional on the side was nerve-wracking. The first time I troubleshot a crying baby through process of elimination (hunger? Diaper? Gas? Temperature? Overstimulation?) I was building a decision tree in real time from zero experience. After the first time of “figuring it out,” the steps registered in my memory and gradually became natural.
Struggle is not a bug in the learning process. It IS the learning process. The friction of figuring something out without a clear answer is what converts information into skill, knowledge into judgment.
AI removes the struggle. You get an answer from a chatbot, a beautiful presentation from an AI agent, without the need to struggle. For productivity, that’s a feature. For skill development, it’s a catastrophic bug.
I recently read a book named The Skill Code, which highlights exactly this issue. One quote I can’t stop thinking about: “Imagine how much worse it can get when we insert technology that separates the expert from the novice and gives that novice a steady stream of hyper-convincing literal snapshots.”
The most important AI skill isn’t prompt engineering. It’s knowing when NOT to use AI. The first time is sacred because it’s when the learning happens. Automate the hundredth time. Protect the first. Without that first-hand experience, it’s impossible to even evaluate whether AI did a 60/100 or a 90/100.
Learning 4: Your Attention Is the Scarce Resource
After the baby, I missed an entire week of AI news. Podcast episodes piled up across all 65 shows. My curated daily newsfeed summary emails went unopened.
And nothing bad happened.
I did manage to catch up with the news afterward. Anthropic launched several great features, but nothing was a game changer to my workflow. The new Nature Communications article on BCMA in vivo CAR-T offered some food for thought but nothing urgent. The discourse I was terrified of missing turned out to be a repackaging of ideas I already understood.
But I did enjoy a magical first week with my wife and the newborn. First eye contact, first diaper change, first bath, first smile. All time much better spent.
Chris Hayes at a16z asks: “Does AI so ruthlessly optimize for what you WILL pay attention to that it alienates you from what you WANT to pay attention to?”
For months, I was paying attention to changelogs. What I wanted to pay attention to was right in front of me. A tiny human discovering the ceiling fan, making a face before a sneeze that is objectively the funniest thing I’ve ever seen.
The moments that truly compound don’t have a changelog: genuine expertise, deep relationships, your kid’s first real look at you. They don’t get superseded next quarter. They are irreversible, unrepeatable, and they will not wait for you to finish scrolling.
AI will keep updating. My son’s first month won’t.
Final thoughts
Here’s the pattern across all four:
AI is extraordinary at the explicit, generic, compressible parts of life. The principal components. The stuff that’s been written down, that applies broadly, that scales. For these tasks, it’s genuinely transformative.
The parts that matter most remain stubbornly human. Tacit knowledge that lives in your hands and brain. First-time experiences that build judgment. The residuals that no compression algorithm preserves. The decisions about what’s worth doing in the first place. The attention you choose to give to a person instead of a screen.
If I had to bet: AI will keep getting better at the principal components. The models will improve. The context windows will expand. The agents will get more capable.
But the residuals will remain yours. The tacit, the specific, the first-time, the sacred.
Words can't express how I enjoyed reading this article, Liang!
As a first time parent myself and AI enthusiast, I can deeply resonate. I love that you made your own Bloomberg terminal for tracking!
I look forward to what you'll create for when your little one starts eating solids! :)