Learn – VerbaGPT

Four-panel chart of MSFT institutional ownership across the last three quarters, by trend, manager location, buyer/seller counts, and investment discretion

Use Case

More Managers Bought Microsoft Than Sold It. The Sellers Still Won.

3,222 managers increased their MSFT stake in Q1 2026, only 2,365 cut it — yet institutional ownership fell 6.7%. The average seller sold about twice what the average buyer bought.

5 min read · Aug 9, 2026

invalid identifier 'RXXPX'
→ checked, not guessed again

Use Case

It Got the Schema Wrong Twice. Both Times, It Checked Instead of Guessing Again.

A typo'd SQL column, then a case-mismatched DataFrame column. Both times the fix was a one-line probe for ground truth, not another guess.

4 min read · Jul 27, 2026

Four-panel chart: diabetic vs non-diabetic GLP-1 users by drug in 2023, non-diabetic share of GLP-1 use trending 2020-2023, cost per fill, and payer mix

Use Case

The Newest GLP-1's Off-Label Share Is Already Double Ozempic's

Tirzepatide had barely two years on the market, yet 33.6% of its MEPS users were already non-diabetic — Trulicity is still at 4.3%.

5 min read · Jul 24, 2026

Fair/Poor → Good+
47.3% within a year

Use Case

Self-Rated Health Isn't Sticky. 47% Went From Poor to Good in a Year.

Nearly half of MEPS respondents who called their health fair or poor called it good or better just one panel year later.

4 min read · Jul 24, 2026

$Bar chart of weighted uninsured fraction among high-risk MEPS respondents by region and year, y-axis capped at 0.02$

Use Case

The Executive Summary Says 11.8%. The Table Two Lines Below It Says 1.18%.

Every table and the chart's own axis agree on the real range. Only the opening summary sentence is off — by exactly 10x.

4 min read · Jul 14, 2026

Four-panel chart of weighted GLP-1 use by education, income, age, and gender, 2019-2023

Use Case

The Income Group With the Lowest GLP-1 Use Isn't the Poorest One

In 2021, the poorest Americans had higher GLP-1 use than the wealthiest bracket. The group in between lagged behind almost everyone.

5 min read · Jul 13, 2026

Four-panel chart of GLP-1 users, total spend, share of pharmacy spend, and share of healthcare spend, 2019-2023, all trending upward

Use Case

GLP-1 Spending Nearly Quadrupled. Cost Per Patient Peaked in 2021.

Four charts show users, spend, and cost share all climbing. A fifth number from the same table — spend per patient — peaked in 2021 and fell since.

5 min read · Jul 13, 2026

Bar chart of average out-of-pocket health spending by employment status, nearly flat across three groups

Use Case

This Group Has the Highest Average Health Spending and the Lowest Typical Spending

Not-in-labor-force spends the most on average and the least at the median. Employed spends the least on average and the most at the median.

4 min read · Jul 13, 2026

Line chart of China's population from 1952 to 2007, showing a visible flattening between 1957 and 1962

Use Case

Inside China's "Rapid Expansion" Phase, One Five-Year Stretch Barely Grew at All

A two-phase summary calls 1952-1982 the fast era. One five-year stretch inside it grew slower than almost everything that came after.

4 min read · Jul 13, 2026

Bar chart of correlation coefficients between diabetes dataset features and disease progression

Use Case

Two of the Five Strongest Correlations Here Both Come From HDL Cholesterol

s4 is defined as total cholesterol ÷ HDL. s3 is HDL. Two of the chart's strongest bars are mostly the same lab value, counted twice.

4 min read · Jul 13, 2026

Line chart of decade-over-decade average annual precipitation in Dubai, 1940-2020s

Use Case

Every Point on This Rainfall Chart Represents a Decade. Except One.

Eight points on Dubai's 85-year rainfall chart average ten years each. The ninth averages six — and nothing on the chart says so.

4 min read · Jul 13, 2026

Line chart of diagnosis rates for the top 10 most-diagnosed conditions in MEPS data, 2019-2023

Use Case

Back Pain Diagnoses Dropped in 2020. Anxiety Diagnoses Never Paused.

Of the ten most-diagnosed conditions in five years of MEPS data, only two break trend in 2020 — both musculoskeletal. Anxiety and depression climb every single year.

4 min read · Jul 13, 2026

1.030 vs. 0.983
Same diabetes, opposite miss

Use Case

Mild Diabetes Is Overpriced. Severe Diabetes Is Underpriced. Same Risk Score.

Six HCC predictive ratios landed within 3% of 1.0 — except the mild and severe tiers of the same diabetes diagnosis, which missed in opposite directions.

4 min read · Jul 7, 2026

People likely with asthma but missing an HCC code in 2023, by evidence type

Use Case

One of the Top 10 "Asthma Drugs" in This Query Wasn't a Drug At All

A frequency-based drug list used as evidence of undercoded asthma included a missing-data code, not a medication.

4 min read · Jul 7, 2026

"R² score of 0.5248"
vs. R² Score: 0.5482

Use Case

A Model Guessed Its Own Result Wrong. The User Never Saw It.

The commentary guessed before the code ran. Only the number the code actually computed reached the chat.

4 min read · Jul 7, 2026

Correlation heatmap of Iris sepal and petal measurements

Use Case

Why the Decision Tree Ignored Sepal Width: A Correlation Heatmap Explains

Petal length and petal width correlate at 0.96. Sepal width is the outlier — negatively correlated with everything else.

3 min read · Jul 6, 2026

Titanic survival rates by passenger class and sex

Use Case

"Women First" Had an Asterisk: What the Titanic Data Actually Shows

The female survival advantage wasn't constant — it was strongest in 2nd class and cratered in 3rd.

3 min read · Jul 6, 2026

Decision tree structure for Iris species prediction

Use Case

A Decision Tree That Ignores Half Its Features — And Still Hits 98%

One sentence in, a trained classifier out. Petal measurements decide everything; sepals contribute exactly 0%.

3 min read · Jul 6, 2026

Healthcare expenditure and risk scores by decile, MEPS 2019-2023

Use Case

Healthcare Spending Doesn't Scale Linearly: A Decile Analysis

The top 10% of spenders cost 4x the next decile down. Risk scores track it almost linearly — until they don't.

4 min read · Jul 6, 2026

Video

New Sample Dataset: 13F SEC

Loading 16M+ rows of SEC holdings data and auto-generating data notes to understand the structure.

4 min watch · Jan 2026

"Start now with messy data.
Document as you go."

Tutorial

Stop Waiting for Perfect Data Documentation

The old model was: clean → document → analyze. AI makes it iterative.

5 min read · Mar 14, 2026

Use Case

On Accuracy When LLMs Analyze Real-World Data

6,000+ exoplanets analyzed. The code was right. The commentary hallucinated.

4 min read · Dec 21, 2025

Use Case

Agent Skills: A Weather Analysis Case Study

Domain expertise loaded at runtime. Expert-quality climate analysis without being an expert.

3 min read · Dec 26, 2025

Use Case

The Unreasonable Effectiveness of Code

Raw FITS astrophotography in. Enhanced Orion Nebula out. 15 minutes instead of hours.

2 min read · Jan 15, 2026

Vision

What AI Future Do We Want?

Our AI future isn't written. The choices we make now about what to build shape the next decade.

3 min read · Jan 7, 2026

"50 apps for 50 problems
isn't scale — it's bloat."

Vision

We're Drowning in Tools

AI makes narrow apps trivial to build. One harness, many solutions on demand, is the better path.

2 min read · Jan 14, 2026

"Train on one model's tools,
and their pricing becomes yours."

Vision

Model Lock-in is the New Vendor Lock-in

AI labs are building moats with proprietary tooling. Here's why that threatens your margins.

3 min read · Jan 17, 2026

"True intelligence knows
when not to think too hard."

Vision

Brute Force AI

Agents burning 50,000 tokens to answer questions that need 50. Efficiency is a form of intelligence.

5 min read · Sep 17, 2025

"One question.
Three separate meters running."

Vision

The Lakehouse Tax

Metering storage, compute, and AI as three separate products doesn't make cost proportional to value — it multiplies it.

3 min read · Aug 5, 2026

Tutorials, use cases,
and ideas

End-to-End Autonomous Data Analysis

More Managers Bought Microsoft Than Sold It. The Sellers Still Won.

It Got the Schema Wrong Twice. Both Times, It Checked Instead of Guessing Again.

The Newest GLP-1's Off-Label Share Is Already Double Ozempic's

Self-Rated Health Isn't Sticky. 47% Went From Poor to Good in a Year.

The Executive Summary Says 11.8%. The Table Two Lines Below It Says 1.18%.

The Income Group With the Lowest GLP-1 Use Isn't the Poorest One

GLP-1 Spending Nearly Quadrupled. Cost Per Patient Peaked in 2021.

This Group Has the Highest Average Health Spending and the Lowest Typical Spending

Inside China's "Rapid Expansion" Phase, One Five-Year Stretch Barely Grew at All

Two of the Five Strongest Correlations Here Both Come From HDL Cholesterol

Every Point on This Rainfall Chart Represents a Decade. Except One.

Back Pain Diagnoses Dropped in 2020. Anxiety Diagnoses Never Paused.

Mild Diabetes Is Overpriced. Severe Diabetes Is Underpriced. Same Risk Score.

One of the Top 10 "Asthma Drugs" in This Query Wasn't a Drug At All

A Model Guessed Its Own Result Wrong. The User Never Saw It.

Why the Decision Tree Ignored Sepal Width: A Correlation Heatmap Explains

"Women First" Had an Asterisk: What the Titanic Data Actually Shows

A Decision Tree That Ignores Half Its Features — And Still Hits 98%