Nobody Teaches You Research

Nobody really teaches you research.

You get a desk, a problem someone else picked, and a vague instruction to "produce something novel."

Most people reverse-engineer the job from what they can observe:

• Papers
• Twitter threads
• Conference announcements
• Research blogs

The result?

They learn how to look like researchers instead of learning how to think like one.

Research isn't one skill.

It's a stack of smaller skills.

The good news is that almost all of them can be deliberately trained.

---

Pick Your Own Problems

Richard Hamming had a habit that made him unpopular at Bell Labs.

At lunch he'd ask people two questions:

"What are the important problems in your field?"

Then immediately:

"Why aren't you working on them?"

People eventually started avoiding his table.

The question hurts because most of us don't actually choose our problems.

We inherit them.

From advisors.

From trending papers.

From whatever a large lab announced last quarter.

The danger is subtle.

You inherit the conclusion without inheriting the reasoning.

You know what everyone is building.

You never learn why.

When the field pivots, you're still solving yesterday's problem.

Worse, you're competing against thousands of researchers who started earlier and have significantly more resources.

---

Start From the Outcome

John Schulman describes two styles of research.

The first:

• Read papers.
• Find something to improve.

The second:

• Imagine something you genuinely wish existed.
• Work backwards until you discover the experiments needed to build it.

He strongly favors the second.

Why?

Because it naturally produces originality.

If you start from a goal you actually care about, you'll inevitably explore territory no survey paper has mapped yet.

---

Train Your Taste

People often treat research taste as if it's innate.

It's closer to a muscle.

Before running an experiment:

Predict the outcome.

Before reading a paper's results:

Guess them.

Before a major release:

Predict whether it'll still matter two years from now.

Then check your predictions later.

Forecast.

Compare.

Correct.

Repeat.

Every cycle trains the most important model you'll ever build:

Your own intuition.

---

Upgrade Your Inputs

Shared reading lists produce shared ideas.

If everyone reads the same Twitter thread and the same trending arXiv papers...

...everyone reaches roughly the same conclusions.

Those conclusions quickly become worthless.

Old work is criminally underrated.

Many "new" ideas are decades old.

Examples include:

• Mixture of Experts (1991)
• LSTMs (1997)
• Backpropagation (1986)

Rich Sutton's The Bitter Lesson remains one of the most predictive essays in modern AI despite being only a thousand words long.

Claude Shannon, back in 1952, suggested an approach that still beats most productivity advice today:

Shrink the problem until it's trivial.

Solve the trivial version.

Then slowly add complexity back.

Simple.

Powerful.

Timeless.

---

Read Beyond Your Specialty

Depth matters.

Breadth matters just as much.

Interpretability borrows heavily from neuroscience.

Evaluation borrows from economics and mechanism design.

Systems research depends on understanding hardware.

Statistics quietly underpins almost everything.

Some of the biggest breakthroughs happen where disciplines overlap.

One more rule:

Read the paper itself.

Not the thread.

Not the summary.

The appendix usually contains the details.

The limitations section usually contains the truth.

---

Write Everything Down

Paul Graham once observed that ideas feel complete until you try writing them.

Writing exposes every missing assumption.

Every logical gap.

Every contradiction.

Richard Feynman had a similar philosophy:

The easiest person to fool is yourself.

Writing is the cheapest defense against self-deception.

Charles Darwin went even further.

Whenever evidence contradicted his beliefs, he immediately wrote it down.

He knew memory conveniently forgets inconvenient facts.

Research journals should contain more than successes.

Keep track of:

• Hypothesis
• Experimental setup
• Prediction
• Result
• Updated belief

Reading last month's mistakes teaches far more than rereading last month's successes.

---

Publish Your Thinking

Not everything belongs inside a paper.

Chris Olah and Nick Cammarata argued that many fields accumulate research debt because useful ideas remain trapped inside notebooks.

A clear explanation is itself a contribution.

Many researchers entered interpretability through blog posts rather than conference papers.

Public writing has another advantage.

It becomes an impossible-to-fake record of how you think.

---

Tighten the Feedback Loop

Research speed isn't measured by papers.

It's measured by how quickly you discover you're wrong.

Stories about Alec Radford rarely revolve around one brilliant insight.

Instead, they describe relentless iteration.

More experiments.

More discarded ideas.

Faster updates to mental models.

That speed compounds.

---

Invest in Tooling

Good tooling dramatically shortens the learning loop.

Launching an experiment should require one command.

Visualizing results should require another.

Every run should be reproducible.

Comparisons should take seconds.

Not hours.

One of Andrej Karpathy's favorite debugging techniques is beautifully simple:

Before training at scale...

Overfit a single batch.

Thirty seconds.

Half your bugs disappear.

Shrink everything until it's cheap.

Get it right.

Then spend the compute.

---

Engineering Is Research

There's an outdated belief that engineering merely supports research.

At the frontier, they're inseparable.

Researchers who can build:

• Evaluation pipelines
• Data systems
• Training infrastructure
• Experimental harnesses

are simply able to test more ideas.

Everyone else waits.

---

Stare at the Outputs

Loss curves are comforting.

They're not understanding.

Most experiments generate far more information than researchers actually inspect.

Things like:

• Failure cases
• Model outputs
• Transcripts
• Edge cases
• Long-tail behaviors

often remain unread.

Karpathy emphasizes spending hours inspecting raw data before writing training code.

Andrew Ng has taught essentially the same lesson for years:

Collect one hundred failures.

Read every one.

Group similar mistakes.

Fix the largest category first.

One surprising transcript often teaches more than another decimal point of benchmark accuracy.

---

Wander on Purpose

Your first research area is usually an accident.

Treat it like one.

Spend time exploring:

• Interpretability
• Reinforcement Learning
• Evaluation
• Systems
• Infrastructure

Somewhere there is a niche where your background becomes an unfair advantage.

You won't discover it by staying in one room forever.

---

Kill Ideas Early

Prototype everything.

Let most ideas die quickly.

Tune your baselines aggressively.

Many exciting improvements disappear once compared against a properly optimized baseline.

Better to discover that yourself than have Reviewer #2 discover it for you.

Ablate relentlessly.

Often only one component truly matters.

It's usually not the one mentioned in the paper title.

---

Breadth Is Career Insurance

Every subfield eventually saturates.

Usually right after becoming popular on social media.

Researchers who continue producing through these cycles already know neighboring fields.

They don't need to start over.

---

Find Your People

Hamming noticed something interesting.

Researchers with closed office doors often produced more work.

Researchers with open doors produced work that mattered more.

Conversations carried information.

Today, your open door might simply be your inbox.

Keep it open.

Generosity compounds surprisingly well.

Replicate papers.

Share tools.

Write explanations.

Release code.

The rewards often arrive months later in unexpected forms:

• Collaborations
• Recommendations
• Job opportunities
• Friendships

Half-formed ideas are worth sharing too.

Being wrong on Twitter costs very little.

Being wrong after three months of work costs much more.

---

The Long Game

Louis Pasteur famously said:

Chance favors the prepared mind.

Richard Hamming built an entire philosophy around that idea.

Knowledge compounds.

Productivity compounds.

Relationships compound.

Small daily improvements seem insignificant in isolation.

But over years they create careers that outsiders mistake for luck.

What you read.

What you write.

Who you argue with.

How quickly you learn.

Every one of those compounds quietly.

Start earlier than feels necessary.

Future you will wish you had.