Draft note: This post is trying to make “research taste” less mystical: treat it as judgment at concrete choice-points in the research loop, then propose ways a novice can deliberately train those judgments.

TODO before publishing:

• Replace the placeholder outline with a clean thesis and section order.
• Decide whether this is one post about research taste or part of the Toolkit of a Researcher sequence.
• Add one worked example where a research direction gets improved by better taste.
• Separate cited claims from my own speculation and mark the latter clearly.

Outline

• Problem: It’s hard to get a pointer on

• Problem: Unclear how to practice. Much instruction is cryptical.

• Solution: Finding taste

• Solution: Subcomponents of taste

  • Confusion-noticing: well-instructed, short loop, start here
  • Execution-judgment: entrepreneur moves, ML/AI safety caveat
  • Hypothesis-generation: the honest gap
  • Problem-selection: simulate feedback on past work

• Worked example: pick one and commit

• Closing claim: pick one before drafting

Finding taste

What is research taste.

Research Taste is the intuition that guides researchers towards productive lines of inquiry.

Definition for “intuition” here: fast, low-deliberation judgment. Some such judgments are trained competences; others are cached biases or social/aesthetic reactions. Research taste is the subset that reliably tracks research fruitfulness.

What is the exact shape of the taste you need to orient towards productive lines of research?

According to Neel Nanda, it’s a combination of

1. Intuition (System 1): This is the fast, gut-level feeling - what people normally think of when they say research taste. A sense of curiosity, excitement, boredom, or skepticism about a direction, experiment, or result.
2. Conceptual Framework (System 2): This is deep domain knowledge and understanding of underlying principles.
3. Strategic Big Picture: Understanding the broader context of the field. What problems are important? What are the major open questions? What approaches have been tried? What constitutes a novel contribution?
4. Conviction & Confidence: Research inevitably involves setbacks. A certain level of conviction – a belief in the direction, resilience to negative results – is often instrumentally useful for perseverance. It helps you push through the messy exploration phase or refine an idea that isn’t working perfectly yet.

My own intuition while drafting this post, and before seeing Neel Nanda’s post, was something like:

1. Confusion-noticing
2. Execution-judgement
3. Hypothesis-generation
4. Problem-selection

There are two different decompositions here. Nanda’s list describes capacities inside the researcher: fast judgment, explicit models, field strategy, and calibrated persistence. My list describes moments in the research loop where those capacities get used: noticing confusion, generating hypotheses, selecting problems, and judging execution paths.

So perhaps “research taste” is not one subskill beside these. It is the quality of judgment at these choice-points. A researcher’s intuition is the broader machinery; research taste is that machinery aimed at fruitfulness.

How and why is it difficult?

How what asdf (change the heading)

Since taste is partially developed through mentorship, and the abstract sciences are an old field, let’s look at some older wisdom on the topic.

Clippings, remove or integrate before publishing

Theory: You need good aim

People speak a lot of research taste. I find the expression taste quite jargony for my taste. So I will speak of aim.

To do excellent research, you must choose a problem that matters. It can matter for you, it can matter for the field of science, but it must matter.

Personal opinion: It’s much more important that it matters to you. But some peoples matteringness calibrations are quite off. I am not sure how to filter for this from the inside. But at least if you’re talking to llms and they’re saying you’re the most brilliant researcher on the topic to ever graze the planet, I’d recommend applying liberal sprinklings of salt.

After you have chosen a problem that matters, you must choose your approach.

Theory: You need to ‘play to learn’

I think here the difference between long-term returns and short-term returns becomes quite significant. In the short term, an approach that is easily falsifiable, publishable, one which other people will understand, seems attractive.

But we must think of the long game.

If you, like me, are relatively early in your years on the planet, and relatively early in your career as a researcher, I think your most important goal should be to research to learn.

Let’s look at a graph:

• (skill and output)

• (Graph based on a naive model where you can invest 80/20 in learning /

I’ve mostly familiarized myself with this in the context of programming and competitive video games, but I feel like there is often this mistake people do with learning stuff.

Related quote: Practicing for 10 000 hours, vs. practicing the same 1 hour 10 000 times.

The point: In any domain where you can change the weighing of your time investment between immediate returns and skillbuilding, skillbuilding early on will pay off more than exploit.

How to play to learn

I’m an ‘entrepreneur type’. Part of what this means is that I don’t like (a significant portion) of the way university teaches people. I like doing things. Bumping into walls. Updating.

And I think it’s an important distinction that you can learn in many ways. Some people want to study theory for 9 years. Some people want to get a good mentor and follow their footsteps, figuratively.

I want to learn by doing.

So if we’re going to be learning by doing anyway, why did I have talk so long about skillbuilding?

Where to aim

The implication of the aforementioned is that when you are choosing your research area, the approach, the specific thing you’re working on the day, you should choose it based on the long-term returns.

Instead of what I’d imagine is a typical priority list for a junior researcher:

1. Will this result be important for the field
2. Could I get employed by someone reading this
3. What new information will I receive from this
4. What skills I’ll learn

I’d recommend

3. What new information will I receive from this
4. What skills I’ll learn
5. Could I get employed by someone reading this
6. Will this result be important for the field

Why?

Why does information let you learn faster?

The most important feature is that it lets you calibrate your aim. And since aim is the most important part of this loop, making it better is a good place to start.

Secondly, you need information to act over the real world.

A specific example

Remembering your goal

I feel like this is again something where startups and research vaguely aligns.

If you’re working solo this is quite crucial for motivation. But it’s also important for keeping aligned with your own goals.

I recommend keeping your goal in mind every day you’re working. This doesn’t mean that you have to have a long goalsetting session every day. But maybe spend at minimum a couple minutes on thinking what you’re currently aiming at, why, and why the step you’re working on today will contribute to that.