AI Tools for Data Analysis: from chatbot to coworker

Two days, in under a minute.

Everything in it was made with the tools we use in class.

Every July, the best coders in the world meet in Tokyo.

They compete in the AtCoder World Tour Finals. In July 2025, an AI joined them.

Jul 2025

Today

The whole story, in writing

July 2025

Humanity prevails (for now)

Przemysław Dębiak (“Psyho”) against an OpenAI reasoning model. The model led for most of the ten-hour heuristic final; Psyho caught it at the very end, on three days of almost no sleep. The models weren't good enough for the algorithmic division at all that year.

1Psyho human0

2OpenAIAHC model0

3terry_u16 human0

The heuristic final, top three.

July 2026

Humanity surrenders

The organizers reportedly picked a heuristic problem designed to favor humans. It didn't matter: Psyho reckoned the human field would have needed several more days to match the model. The next day it solved all five algorithmic problems, two that no human in the room cracked. Two “humanity surrenders” awards, one per division.

HeuristicHumanity surrenders

AlgorithmicHumanity surrenders

The ¥600,000 “Humanity Prevails” award went unclaimed in both divisions.

July 2026

And then, new science

Mathematician Levent Alpöge announced that Claude Fable 5 found a counterexample to the Jacobian conjecture, open for 87 years and independently verified within days. Timothy Gowers called it the first example he knew of an LLM solving a famous problem outside his own area.

19392026

Open for 87 years. The thread on X ↗

September 2026

A Millennium Prize problem

Then OpenAI announced that about 10,000 of its agents, working for 88 hours, had proved that the Navier–Stokes equations (the equations of how fluids move) can blow up in finite time when pushed by a smooth force. It is one of the seven Millennium Prize problems, each worth a million dollars. The proof was checked by computer, and the Clay Mathematics Institute said the problem had “apparently been settled.” The unforced version is still open.

P vs NP
Riemann hypothesis
Hodge conjecture
Yang–Mills mass gap
Birch & Swinnerton-Dyer
PoincaréPerelman, 2003
Navier–StokesAI agents, 2026

The seven problems, set in 2000. Two now settled. The Clay statement ↗

Psyho spent twenty years becoming the best coder alive.

The machine passed him in .

twenty years

twelve months

Where will we be five years from now?

What will our jobs look like?
What is left for us to do?
And what should never be handed over?

The truth is, nobody knows.

But the nature of our jobs will radically change, and the responsibility will shift.

For programmers, it already has. In February, Andrej Karpathy, one of the founders of OpenAI, wrote:

Andrej Karpathy on X: it is hard to communicate how much programming has changed due to AI in the last 2 months; coding agents basically didn't work before December and basically work since.

“It is hard to communicate how much programming has changed due to AI in the last 2 months.” Now with one word changed: programming, crossed out, becomes data analysis.

Data analysis has already changed too.

And it is only speeding up.

From chatbot to coworker in less than four years. A short history.

scroll to climb ↓

Milestones are spaced evenly; the bands show the years.

The course poster: an analyst in a rowing boat with her laptop; beside her only the robot’s head shows above the water, and below the surface the whole robot works at a laptop, with files, charts, a magnifying glass and a gear floating around it. — Most of us use this much. This is what’s underneath.

Three levels of letting go.

Each level hands the machine a bigger piece of the work and changes where we stand.

Chatbot

we are the loop

day oneLevel one, live

Coworker

we are at lunch

day twoLevel two, live

Manager

we design the loop

day twoLevel three, live

Delegate a task.

It writes the code, runs it, and hands back the receipts.

Now cooking OLS + confidence bands

Every number below is real. Pick another dish

you▤ happiness_2023.csv · 2,363 rows

claude 0.0s

Day one, à la carte

Prompting, the lean way

Information, not incantations. ~~“You are an expert data scientist…”~~ Start plain, look at what comes back, add what was missing: context, goal, output, done. Once you know what you like, write it down as a skill, then one line is enough.

The cold start: a default matplotlib scatter of GDP per capita against happiness.

After two rounds: a magma-colored scatter across 153 countries with a dashed OLS line, a gray 95% band and named outliers.

With the watercolor-graphs skill: the same question as a watercolor chart: 129 countries in 2023, a painted fit line with a 95% band, plus 0.51 points per doubling of GDP, and the three countries farthest above and below the line named.

The same data, as a world you can paint.

Move the brush across the map: each country floods with how its people rate their lives, and plays its note.Swipe to paint. Tap a country for its numbers.

How happy is the world?

Sound is on.

less happyhappier· life satisfaction, 0–10, 2021–23

0 of 165 countries painted

Delegate a project.

A folder, a goal, permission. Then we go to lunch.

the-project/ · goal: “a full report on what moves happiness”idle

the-project/
data/happiness_2023.csv
papers/

The house rules. Read first, every session.

Noon. We press Run and go to lunch.

Delegate a process.

We write the check and the stop rule. The manager does the rest.

The managergoal: “a full report on what moves happiness” agents 0loops 0checks 0/0

How high is the bar? lenientstrict

stop when: every check passesafter 3 rounds

One manager. It decides how many agents the job needs.

“A loop is an agent repeating cycles of work until a stop condition is met.”

Stop when the check passes, not when an answer merely exists.

The worst errors are silent.

The code runs, the chart looks fine, and the answer is still wrong. That is how OpenAI came to withdraw three of its 722 AI-written math papers a day after posting them.

“The world has got sadder: average life satisfaction fell 0.83 points since 2005.”

or open the files yourself:

Both means are right. The comparison isn't: about two-thirds of the “decline” is the survey growing.

Four checks a machine can run. The fifth is ours.

Beautiful is the default. Judgment isn't.

The machine drew this chart. Which headline would you sign?

Pick a headline to sign

Four headlines, one chart, the same dots.

What people took away.

…interactive dashboards … were invaluable.

Christine Davis, The National Center for Construction Education and Research

read it all

“The diversity in applications for using ChatGPT for data analysis, especially the creation of interactive dashboards and HTML to host on a website, were invaluable.”

…I learned so much new information about what AI can do…

Jane C. Daquin, Sam Houston State University

read it all

“I have been using ChatGPT every day since 2023 for research, teaching, and even creative writing, though primarily to code advanced statistical models. Despite this experience, I learned so much new information about what AI can do and its potential.”

…explained all the topics very well… really interactive.

Laura Fabregat, Universitat Rovira i Virgili

read it all

“I think Boris explained all the topics very well. He knows a lot and was really interactive. Data analysis and AI are very, very important topics that really help our academic career.”