2026·05·18
The Cheapest Idea I Ever Had Cost Me a Year— on premature optimization of beliefs
12 min
$
cat ~/now.md # what's loaded into memory
// week of may 25 — jun 1, 2026
›Re-reading "Thinking in Bets" alongside building a small bankroll-sizing notebook.
›Training a sparse autoencoder on intraday order-book imbalance; first results are interesting.
›Slow project: a personal "research OS" — capture, link, distill. Mostly markdown + scripts.
›Walking the shiba twice a day. The cat is unimpressed.
$
ls -lah ~/projects/
permissions
name
brief
updated
status
drwxr-xr-x
orderbook-sae/
Sparse autoencoders on L2 book features. Toy → live.
2d
wip
drwxr-xr-x
research-os/
A reading + note + spaced-recall pipeline. Markdown all the way down.
1w
live
-rw-r--r--
bet-sizing.ipynb
Kelly with uncertainty: small notebook, big consequences.
3w
live
drwxr-xr-x
shiba-tracker/
A weekend project that started as a joke. Pose-estimation + walk routes.
1mo
wip
drwxr-xr-x
factor-zoo/
A small, honest survey of factor decay in mid-cap equities.
2mo
live
-rw-r--r--
retrieval-lab/
BM25 vs dense vs hybrid on my own reading corpus. Hybrid wins, barely.
4mo
live
drwxr-xr-x
slowread/
A CLI that paces ebooks to a steady WPM. For people who skim too much.
6mo
live
-r--r--r--
microcap-mean-rev/
First serious strategy. Worked. Then didn't. Lessons preserved.
1y
archive
$
cat ~/writing/index.md
2026·04·02
Reading as Compression— notes on what survives a year of forgetting
8 min
2026·02·14
A Boring Defense of Boring Models— why I keep reaching for logistic regression
6 min
2026·01·09
Trading the Edges of My Own Attention— a small, honest journal
14 min
2025·11·30
Letters to a Junior Researcher— five rules I keep breaking
9 min
2025·10·06
On Keeping a Shiba and a Habit— small loops, large effects
4 min
$
tail -f ~/notes/scratch.log
#0412
A model that refuses to predict is more useful than one that always does. Abstention is information.
2026·05·22 · trading
#0411
Reading speed is a vanity metric. Reading residue — what you can still cite a year later — is the real one.
2026·05·19 · reading
#0410
The shiba's selective hearing is a great analogy for attention heads. He attends only to treat and walk.
2026·05·15 · ml
#0409
"Idea not cheap" — because the cheap part is generation. The expensive part is selection, then commitment.
2026·05·12 · meta
#0408
Most factor decay isn't from competition. It's from regime change you didn't notice because your backtest was too clean.
2026·05·08 · trading
#0407
If you can't explain your model to the cat, the cat is right.
2026·05·05 · life
$
cat ~/reading/2026.md # in-progress, finished, queued
Thinking in Bets
62%
The Alignment Problem
28%
Trading and Exchanges
done · ★★★★★
How to Take Smart Notes
done · ★★★★
Probabilistic Machine Learning
done · ★★★★★
Advances in Financial Machine Learning
queued
The Embedding Space
queued
$
whoami --verbose
"Ideas are cheap; the expensive thing is to live one out loud, for years, in front of people."
I'm Tao, a data scientist who treats reading, trading, and machine learning as three views of the same problem: how to take a vague hunch and shape it into something the world will pay you to be right about.
I work mostly at the intersection of quantitative research and ML systems — sparse models, retrieval, decision-making under uncertainty — and write to keep myself honest about what I actually believe.
Outside of work: long walks with a stubborn shiba, a cat who supervises, and a slow stack of books I refuse to skim.
based insomewhere with good coffee, UTC+8
tools
pythonjaxduckdbpolarsobsidiantmuxneovimmake
studying
interpretabilitymarket microstructurebayesian dm
currentlybuilding a small, opinionated research OS for one user (me)
open tocollaborations · paid research · long emails
nevercrypto pump groups, "10x" anything
$
curl -X POST tao@ideanotcheap.com
粤公网安备44010402003714号