How to Truly Do Great Research?

Today I want to share an article.

Title:

How to be good at research

Author: Vivek Nair, researcher at Anthropic.

On the surface, this article is about “how to do good research.”

But my biggest takeaway after reading it is:

It’s not just for PhD students, AI researchers, or industry researchers.

It’s incredibly insightful for ordinary people’s career development, investment judgments, entrepreneurial direction choices, and even long-term personal growth.

Because truly good research, at its core, is not about “being good at reading papers” or “being good at running experiments.”

It’s about whether you can consistently ask better questions, find better information sources, discover where you’re wrong faster, and turn all that feedback into long-term compounding returns.

Many people think research is an academic skill.

But at a deeper level, research is a way of engaging with the world:

How do you choose your problems. How do you judge what’s important. How do you handle uncertainty. How do you accept being wrong. How do you extract information from failure. How do you turn scattered experiences into reusable judgment over time.

Let me share a few points that deeply resonated with me.

1. Choosing the problem matters more than solving it

Many people’s first reaction to doing research is:

What’s the hottest direction right now? What are the big labs working on? Which paper recently blew up? Which benchmark just got a new high score?

And then they chase after it.

It’s not entirely useless, of course.

But the problem is:

By the time you see a paper, a result, or a product launch, it’s often the outcome of someone else choosing that direction a long time ago.

You see the conclusion.

But you don’t see why they chose that problem in the first place.

You don’t know their original judgment, how many times they iterated and failed internally, or when they were already prepared to pivot.

By the time results become public, a large number of researchers, engineers, and entrepreneurs may have already pounced on it.

When you enter at that stage, you’re usually competing in a very crowded space against people with more head start and more resources.

So the most important reminder in the original text is:

Don’t just receive problems from others; train your ability to choose problems yourself.

Richard Hamming had a famous habit at Bell Labs.

He would ask people at lunch:

What is the most important problem in your field? What problem are you working on right now? If these problems are so important, why aren’t you working on them?

That question cuts deep.

Because it forces you to face a truth:

Often, it’s not that we’re not working hard; it’s that we’re working hard on problems that aren’t important enough.

This applies equally to careers, investing, and entrepreneurship.

In career development, the truly important question is not:

“How do I finish today’s tasks?”

But rather:

“Will the skills I’m building now still be valuable in three years?”

In investing, the truly important question is not:

“Will this stock go up tomorrow?”

But rather:

“What is the core variable driving this industry over the next ten years?”

In entrepreneurship, the truly important question is not:

“How do I build this feature?”

But rather:

“Is this problem really important enough for people to pay for it consistently?”

If you choose the wrong problem, all the hard work in the world just makes the wrong answer look more polished.

2. True originality often comes from problems you genuinely care about

The article also highlights a very important research approach.

It’s not about first looking at the literature for a gap and then filling a small hole.

Instead, it starts with:

What outcome do I truly wish existed in the world? What problem do I genuinely want to solve? What phenomenon has always puzzled me? Where do I repeatedly feel something is “off”?

Then work backwards:

If this problem were to be solved, what experiments would be needed? What data? What tools? What intermediate assumptions need to be validated?

This approach is more likely to generate originality.

Because the problems each person truly cares about are different.

Your life experiences, industry background, failures, aesthetic preferences, and long-standing confusions all allow you to see problems others miss.

Many people think research should be as “objective” as possible, removing yourself from the question.

But to some extent, truly good questions come precisely from your unique experiences.

The traps you’ve fallen into in a certain industry. The long-term torment of a specific workflow in a role. The repeated failures you’ve seen in a certain type of product. The nagging feeling that something is off in a technical direction. The persistent anomaly you’ve observed in a market.

These are not noise.

These may be your entry points into research.

Differentiated inputs lead to differentiated questions.

Differentiated questions lead to differentiated results.

3. Research taste is not a gift; it’s trainable

There’s a key word in the original text:

Taste.

Research taste means:

Can you judge whether a problem is important? Whether a direction has potential? Whether an experimental result is truly meaningful? Whether a paper is a real breakthrough or just well-packaged? Whether a product addresses a real need or just looks sexy? Whether an opportunity is a real trend or short-term hype?

Many people think taste is a gift.

But the author’s view is:

Taste is more like a muscle; it can be trained.

How?

Simple, but counterintuitive:

Write down your predictions before every experiment.

What do you think the result will be? Which variable do you think matters most? Where do you think this method will fail? What do you think will be the biggest bottleneck in the end?

Then, after the experiment, compare against your predictions.

Where were you right? Where were you wrong? Why were you wrong? Lack of information, or wrong model? Bad judgment, or emotionally driven?

You can practice this with papers too.

Don’t rush to the conclusion.

First, look at the problem and method. Then pause and ask yourself:

If I were the author, what would I roughly expect from this experiment? How strong would this method be? Which ablations might hurt performance? What limitations might the authors put in the limitations section?

Do this consistently, and your brain will develop a feedback system.

Every prediction and correction trains your judgment.

This is useful for research, and also extremely useful for investing.

Many people lose money in investing not because they lack information, but because they never record their own judgments.

Why did you buy before you bought? What did you expect to happen? What would indicate you were wrong? If wrong, was it the logic or the timing? Was it incomplete information, or being carried away by market sentiment?

If you don’t record, your memory will automatically beautify every loss.

In the end, you only remember “I was almost right.”

But real growth comes from being willing to see, repeatedly, how specifically wrong you were.

4. Upgrade your information sources, or you’ll only get average ideas

If every day you consume the same hot topics, same reposts, same summaries, same opinions, then the ideas you generate will also be highly similar.

Shared information sources produce shared conclusions.

And shared conclusions are usually not very valuable.

That’s why the author emphasizes:

You need to upgrade your inputs.

The first upgrade: Go back to the classics.

Many things that seem new today were thought of long ago.

For example, architectures like MoE had related research as early as 1991.

Backpropagation wasn’t important just today; the 1986 Nature paper on backpropagation was already a key milestone in deep learning history.

Richard Sutton’s “The Bitter Lesson” isn’t long, but its judgment on the direction of AI development remains penetrating even today.

Classics are classics not because they’re old.

But because they’ve survived the filter of time, and they can still explain things happening today.

Trends expire.

Classics keep generating new explanatory power.

The second upgrade: Cross-disciplinary learning.

Many ideas for AI interpretability come from neuroscience. Many evaluation methods are essentially related to statistics, psychology, and mechanism design. Whether a model architecture can run efficiently depends on hardware, memory bandwidth, and data flow. Many product problems are ultimately behavioral science, organizational structure, and business model problems.

If you only stay in a narrow information pool, you risk becoming more and more refined, but more and more closed.

Truly valuable ideas often come from the gaps between different fields.

The third upgrade: Read the original sources.

Don’t just read Twitter threads. Don’t just watch short video summaries. Don’t just read secondhand opinions curated by others.

Read the paper. Read the main text. Read the appendix. Read the limitations. Read the details the authors didn’t put in the abstract.

Because a lot of truly valuable information is hidden in places that aren’t easy to share.

The abstract is for making you interested.

The main text tells you the logic.

The appendix and limitations often tell you things closer to the truth.

5. Writing is not a communication tool; it’s a thinking tool

I really liked the part of the article about writing.

Many people think writing is:

I’ve already figured it out, and now I’m just writing it down.

But the reality is often the opposite:

You think you’ve figured it out, until you start writing and realize you haven’t figured it out at all.

Paul Graham has made a similar point:

Writing makes thoughts more precise and more complete.

An idea might seem complete in your head.

But once you write it down, you discover all the gaps.

Where’s the missing step in the deduction? Which assumption hasn’t been validated? Which two points actually contradict each other? Which example is just emotion, not evidence? Which conclusion just sounds nice, but doesn’t hold up?

Writing exposes all of this.

So writing isn’t the final polish.

Writing is part of the research process itself.

Feynman said something similar:

The person you must be most wary of is yourself.

Because it’s easiest to deceive yourself.

Darwin also had a habit:

Whenever he observed a fact inconsistent with his theory, he would write it down immediately.

Because he knew that if he didn’t write it down, memory would quickly delete these “inconvenient pieces of evidence.”

This point is so important.

Our brain is not a reliable database.

It’s more like an editor that automatically maintains self-esteem.

It keeps evidence that supports you and downplays evidence that contradicts you.

It remembers moments when you almost succeeded and forgets times when your judgment was completely off.

So you need to write it down.

Write down the hypothesis. Write down the evidence. Write down counterexamples. Write down failures. Write down why you believed it at the time. Write down why you changed your mind later.

These records are proof that you actually thought.

6. Research speed depends on how fast you discover you’re wrong

The article also has a very hard-hitting point:

Research speed is essentially how fast you discover you’re wrong.

Many people think experts are great because they get it right from the start.

But truly productive researchers are often not always right.

Instead, they have extremely fast feedback loops.

They generate more hypotheses per day. Run more small experiments. See more failure samples. Update their models faster. Abandon unpromising directions sooner.

There’s a very important method behind this:

Break down complex problems.

Don’t try to solve the full version from the beginning.

Start with a small version you can fully understand.

Get it working at a small scale. Understand the relationships between variables. Know where it fails. Then gradually add complexity.

Many of Karpathy’s teaching projects exemplify this approach.

Instead of jumping straight into a huge black-box system, he starts with the smallest model, smallest training loop, smallest runnable code, allowing you to gradually understand the structure behind large models.

The same applies to entrepreneurship.

Don’t try to build a full platform from day one.

Start with a small tool. Serve a small audience. Validate a clear pain point. Get the shortest possible feedback loop. Find out why users don’t use it, don’t buy it, don’t stay.

Complex systems aren’t built by imagination.

They grow from a small, working system over time.

7. Don’t just look at metrics; stare at the failure cases

I think this point is especially relevant in the AI era.

Many people working on models, products, or growth love looking at top-line metrics.

Accuracy went up. Retention went up. Conversion went up. Loss curve went down. DAU hit a new high.

These are important, of course.

But metrics are not analysis.

Metrics are just the entry point.

Real information is often hidden in the failure cases.

Andrew Ng talked about a method long ago:

If you want to improve a system, don’t just stare at the average metrics.

Pull out a batch of failure cases and look at them one by one.

Then categorize them.

Which are data issues? Which are labeling issues? Which are model misunderstandings? Which are edge cases? Which are mismatches between user expectations and product design? Which are not technical problems at all, but a non-existent need?

Once you’ve sorted your failure cases into piles, the biggest pile is often the most worthwhile direction to attack.

This applies to entrepreneurship too.

Don’t just look at “user growth is good.”

You need to look at:

Why didn’t these users stay? Why didn’t they use the product after signing up? Why didn’t they buy after viewing the page? Why did they say they liked it but never paid? Why did they share it but not really use it? Why did they bookmark it but never open it again?

Failure cases are the diagnostic report your system gives you.

The problem is, most people don’t want to look at them.

Because looking at failure cases is tedious and painful.

But real progress often hides in that tedium and pain.

8. Tools are part of research capability

The article also has a very practical suggestion:

You need to polish your tools.

If launching an experiment requires ten steps, you won’t do many per day.

If compiling results takes half an hour, you’ll subconsciously avoid reviewing.

If finding an article you read three months ago takes twenty minutes, your train of thought gets broken.

So good researchers constantly reduce the friction between idea and validation.

If you can launch an experiment in one line, don’t make it two. If you can log results automatically, don’t copy them manually. If you can auto-generate comparison plots, don’t redo them each time. If you can put resources into a searchable system, don’t scatter them across five apps. If you can turn something into a template, don’t start from zero every time.

In the AI era, this is even more important.

Because AI can handle many repetitive tasks for you.

But the prerequisite is that you design the workflow.

You need your own research log. Your own knowledge base. Your own experiment templates. Your own review format. Your own list of questions. Your own repository of failure cases. Your own record of judgments.

This is not formalism.

This is a compounding system.

What truly separates people is not a single moment of inspiration.

It’s whether you have a mechanism that allows every reading, experiment, failure, writing, and review to leave a trace, and to be recalled for the next use.

9. Public output is the strongest filter

There’s another point that’s easy to overlook:

Write some of your ideas publicly.

Many people think that only publishing a paper counts as output.

But high-quality public explanations are contributions in themselves.

Clearly explaining a complex problem. Reproducing a result. Pointing out the boundaries of a method. Mapping out a field’s landscape. Articulating a failure experience clearly. Packaging a tool chain. Laying out a judgment process for others to see.

All of these attract like-minded people.

Public output has another benefit:

It helps you find true collaborators.

Resumes can be polished. Titles can be polished. A line like “I’m very interested in AI” can be polished.

But long-term public writing is hard to fake.

Because it exposes your way of thinking.

How you ask questions. How you reason. How you handle counterexamples. How you acknowledge uncertainty. How you update your views. How you confront your past mistakes.

These things are far more convincing than any self-introduction.

For ordinary people, it’s the same.

Continuously publishing your judgments, reviews, and methodologies publicly is essentially accumulating a verifiable record of your thinking on the internet.

That is far more powerful than saying “I have great insight.”

10. Find people who can tell you you’re wrong, early

The last point is about people.

Research is not done in isolation.

Of course, you need deep work, long periods of solitude, and the ability to think through problems yourself.

But you also need people who can point out your issues early.

A good collaborator might tell you before you invest three months:

This direction isn’t worth it. This assumption doesn’t hold. This experimental design is flawed. This metric doesn’t support your claim. This problem has already been solved. This conclusion sounds nice but lacks evidence. You’re being fooled by novelty. This isn’t what you should be validating.

This kind of feedback is incredibly valuable.

Because it doesn’t just save you one discussion.

It saves you months of your life.

That’s why truly good research relationships are not about mutual praise.

They’re about improving each other’s judgment quality.

This is also why public output, reproducing others’ results, sharing your tools, and seriously commenting on others’ work all matter.

You’re not doing it “for networking.”

You’re letting others see how you think.

When you consistently contribute useful information, relationships form naturally around the work.

Summary

The biggest inspiration I got from this article is:

Doing great research is not a mysterious skill.

It’s a set of trainable abilities:

First, choose truly important problems, not just hot topics.

Second, train your research taste by predicting and correcting to improve judgment.

Third, upgrade your information sources: read classics, read originals, read across disciplines.

Fourth, use writing to expose gaps in your thinking, not just as a communication tool.

Fifth, shorten your feedback loop: generate hypotheses faster, discover you’re wrong faster.

Sixth, stare at failure cases, because that’s where the system’s real problems often lie.

Seventh, polish your tools and workflows so every thought can compound.

Eighth, publish some of your ideas publicly so like-minded people can find you through your work.

Ninth, find people who can tell you you’re wrong early; don’t just stay in comfortable information environments.

This set of methods seems to be about research.

But I think it’s more about how a person can become more clear-headed.

Making career choices is essentially researching the match between yourself and the market.

Making investments is essentially researching how the world changes.

Building a startup is essentially researching what problems are truly worth solving.

Creating content is essentially researching what ideas are worth expressing.

So maybe the truly important question in the end is:

Are the things you’re busy with every day truly important problems?

Or are they problems that others have already chosen, already crowded, already overheated?

If you can start seriously answering this question, you might already be closer to “good research” than most people.