My week’s top reads: clear, useful, and (hopefully) worth your time.

My advice on writing, for what it's worth

In "How to Do Great Work", Paul Graham says:

If you’re making something for people, make sure it’s something they actually want. The best way to do this is to make something you yourself want.

And in "What I’ve Learned from Users":

I often tell founders to make something they themselves want, and YC is certainly that.

The reason this advice is good (though exceptions exist) is that, by following it, you will have at least one real user (yourself) that you can use for feedback to build an MVP, otherwise, you can easily fool yourself into believing what you are building is useful.

Well, I think this concept can generalize to all forms of creation:

• When you write; write something you would read.

• When you do research; look into questions you'd be interested to answer.

• When you release open-source code; package it in a way you'd use it.

Conversation & Ideas

How a group of people approach a conversation either makes it or breaks it.

Examples of good conversation principles:

• A shared desire to objectively understand something or reach the truth.

• A return to first principals to either solve a problem or understand a topic.

• A common incentive to solve a problem using measurable metrics.

Examples of bad conversation principles:

• Seeking approval on ideas. Avoiding disagreement or not tolerating it.

• Debating for the sake of debate. Taking the defence by default.

• Gossiping.

We can selectively opt in or out of conversations based on our knowledge of the participants and their motivations.

Things you're allowed to do

There are many things we could do, but don’t—simply because most people don’t. Examples:

• Use your money for things other than consumption/investment:

  • Hire a researcher or expert consultant.

  • Run surveys.

  • Buy advertisements.

  • Buy research or data.

  • Hire a tutor.

  • Buy goods/services from your friends.

  • Give to charity.

  • Hire a coach.

• Ask questions online.

• Travel to friends just to visit them.

• Live in multiple places with multiple people.

• Generate your own ebooks.

• Ignore what’s on the jobs page & pitch someone at a company on hiring you.

• Negotiate for better terms in your job offer.

• Ask for a raise or drop out/quit your job.

• Live off your savings (or get a grant) while trying something new.

I think the space of what we can creatively do is huge!

The Virtue of Silence

In his (excellent) book, "Antifragile", Nassem Taleb points out that in many situations, refraining from intervention (doing nothing) is the optimal choice. However, we don't get paid to do nothing (in money nor recognition/visibility), this is why doctors are inclined to leave you with a prescription despite knowing you will naturally recover. In such situation, the actor should indeed do nothing and bear the cost (missed opportunities, reputational damage, decreased pay, etc).

Inflation is Multiple Taxation

This article opened my eyes to the other forms of inflation that exist in our lives. Aside from currency depreciation...

• Capital gains tax: when asset appreciation is tracking inflation (e.g., 10% inflation resulting in 10% asset appreciation) and we don't pay taxes, we are breakeven. However, when we pay taxes on our gains (which is our reality), we lose purchasing power. This means that to be breakeven, ours returns need to be higher than the inflation rate.

• Higher bracket taxation: we receiving a 5% annual raise with 5% inflation are no better off in real terms, yet may end up in a higher tax bracket, losing more of our income to taxes.

• Reduced tax credits & benefits: as we attempt to keep up with inflation, we get decreased benefits as we are considered higher income, but in reality, we are just trying to keep up with inflation.

It's important to be aware of such realities to avoid being slowly boiled alive.

Hi 👋, here are my week's readings that stood out for their actionable insights!

Nabeel S. Qureshi

This week, I wanted to read a few pieces from Nabeel which I found insightful. I will first dive into a few highlights from his archive and I encourage you to follow him on Substack.

Principles

63 life lessons from Nabeel. Some that I think are very important:

• The environment you live in matters a lot; move to where you flourish maximally.

• Do things fast: I would add that most work is optional, so first figure out what matters most, then do it fast. A good enforcer is to set deadlines.

• Wealth can be created: wealth is not a zero-sum game.

• Figure out your primary focus and do it first thing in the morning.

• Produce surplus: your production minus consumption should be positive, preferably a big number.

• Put yourself out there! I write about this in "create surface area for luck"

• Learn how to meditate.

• When you are in a high-pressure situation, remember that everything passes and understand the worst outcome isn't that bad at all.

• You are too risk-averse, take the leap. This is hard to do because it goes against our nature but can be extremely rewarding if done right.

• Know your "triggers" or what brings out the worst version of yourself.

• Status is fake. Just focus on substance and doing valuable work.

• Understand power laws and compounding effects in life.

• Don't ask for approval and permission to do things: in other words, your bosses don't know much either, and they're relying on you to figure things out.

• Be willing to learn from others. The most valuable feedback usually hurts a lot. However, you need to have taste in recognizing good from bad takes.

Advice That Actually Worked for Me

I am a big fan of actionable knowledge. That's why I liked this list of direct tips from Nabeel. I think some items in his list are more optional than others, here's what I think are the most important ones:

• Maximize energy levels: you do this by building good morning habits (exercise, meditation, putting everything in order) that give you enough energy to work on the most important thing in the morning. That by itself motivates you to finish the optional low-energy items by the end of the day. Controlling this loop is essential.

• Do the most important thing first thing in the morning, and don't check social media until you've done it: I disagree with Nabeel on social media, I don't want to check any socials until the day ends (~8PM).

• Tell the right stories about yourself: I think this matters a lot. If you believe you won't do anything, then that belief makes you more likely to not do anything. Beliefs are self-fulfilling prophecy machines. You can use this to your advantage, if you believe you can make it, you will act in ways that will set you up for it.

• Get in the habit of Fermi estimation, looking up key quantities, and using upper and lower bounds: use data to escape the narrative fallacy. Do napkin math to bound your expectations and think from first principles when the situation requires it.

• Do a weekly review: I think planning and reflections are very important alongside dedicated focused work. I write about my routine in "My weekly Review Habit" and about focus in "Things I do to stay focused".

• Synthesize things as you read: we don't only want to memorize quotes. When reading, working, discussing, etc., we need to recognize when a mentioned concept or idea is so important that we need to stop and re-think it, internalize it, maybe set a backlog task to look into it in more detail later.

How to Sell?

I know nothing about being a good salesperson, but this read gave me a good starter framework if I ever approach the topic. It systematically tackles important questions like:

• How do I know the chance of converting someone before I approach them?

• Who do I want to approach in a potential buyer company?

• What are the human psychological biases I need to tap into to convert?

• How should a meeting with a potential client be structured?

• How can I close a deal?

Some counterintuitive learnings:

• The potential buyer should do most of the talking.

• You don't want to pitch, you want the client to realize they want your product by themselves.

• If you realize your product is not a good fit for them, cut it short and recommend the best (sometimes competitor) product that could serve their needs.

Making Beliefs Pay Rent

I revisited this post by Eliezer and it still rings true. Our personal heuristics or models about the world are critical because they help us make decisions at all times. We need to continuously test their validity:

• Start by having as few heuristics as possible. A way to minimize our beliefs is to remove the ones that are non-falsifiable (we can't disprove them) and the ones that don't actually help us do anything (non-usable).

• Since the remaining heuristics are falsifiable (they claim predictive power), we need to document their predictions and verify them. Example: when you disagree with someone, remember to check the result of their course of action versus yours.

Beliefs should be predictive and always pay rent in the form of correct predictions. The habit of documenting your predictions and reflecting upon their results in an honest manner is enough to solidify your world model.

Default Alive or Default Dead?

There are very important things in life that we give little thought about, don't even consider sometimes, or ignore. Examples:

• Deciding where to spend most of your life.

• Picking a partner.

• Investment options.

Sometimes we do this because we think we can't know the right answer and action beats inaction. Other times we fail to recognize the importance of the topic.

One that's especially important to founders is mapping the burn rate before product-market fit. Paul Graham shares simple yet effective advice to model this problem well:

Calculate your burn rate and runway precisely. Most founders are surprisingly vague about these numbers. Default alive means your growth trajectory will lead to profitability before you run out of money. Default dead means it won't. If you're default dead, you have two options: cut costs or increase revenue. Most focus only on revenue, but cutting costs is often faster and more reliable. The key insight: knowing which state you're in changes everything about how you should operate. Default alive companies can focus on growth and innovation; default dead companies must focus on survival first.

Research as Understanding

When I was in university, I always thought "Research" is for the very smart - and I definitely wasn't one of them. When you hear "research", we think of novel ideas at the bleeding edge, complex theory, extreme left-curve IQ genius.

Anyone can do research. You just need curiosity, systematic exploration, and willingness to share what you find. The generating function for novel work is trying really hard to understand something until you pass through the edge of current human understanding, then continuing imaginatively onward.

More on Repealing the Laws of Economics

From the great Howard Marks (I'm also a fan of his book: Most Important Thing). In his memo, he is making the case that the economic man is a reactive adaptive man, not passive, and while government policy can sometimes achieve goals through well-designed incentives, poorly designed policies offer no free lunch - there's no solution that gives free benefits to everyone. He walks through examples like California's fire insurance price caps, which drove insurers out of the market and left homeowners uninsured for recent devastating fires. Overall, free-market solutions shall be used by default.

Why we’re still doing normal school

This piece captures the fascinating reality of being a parent trying to navigate what lies ahead for young kids in an uncertain future.

Julia Wise acknowledges that current schools prepare children for something like the present or past, which feels absurd when AI will likely transform work and daily life by the time her children are adults.

Yet she and her partner stick with traditional school for three main reasons: school is where the other children are, they don't want to quit their meaningful work, and the future is too weird and hard to prepare for anyway.

When thinking about education, it's not just about knowledge, it's also about developing social skills through relationships with other kids. Homeschooling can only optimize the knowledge (side) of education, there needs to be dedication to social and other forms of education (environmental, spiritual, etc). And even with all its limitations, school provides that community.

Even if parents are willing to commit full time to raising kids, there are still questions about the evolving nature of technology. Tech is forward-looking while traditional schools always lag and adapt late. Given that, parents need to figure out which skills and knowledge is most useful given a future scenario (like the rise of AGI).

Three pieces I read this week that are worth sharing:

Being the (Pareto) Best in the World

It is extremely difficult to be the best in the world at any specific skill. As you climb the expertise ladder, progress becomes exponentially harder. But there's another path to excellence: being multidisciplinary.

If you're great at one thing, very good at another, decent at a third, and interested in many others, you might be the only person on the planet with that exact combination.

The world has plenty of opportunities. The overexploited ones require just a few skills. The underexploited ones hide in high-dimensional corners where few people combine the right interests. This explains why interdisciplinary work has a reputation for unusual returns. It's why diversity helps when building teams.

The lesson? Don't shut down your other interests. Widen your scope. You don't need to allocate time uniformly, but maintain that long tail of curiosities while staying expert in your main thing.

The Narrative Fallacy

We're only comfortable when the world makes sense. When we encounter something new, we rush to explain it. We don't need much convincing because we want to feel like we know. Stories and narratives work because they satisfy this basic human need. As a result, we live in a world ruled by narratives and memes, where truth only matters in small corners.

Several factors make this worse. Plausible stories sound true. We default to believing what we hear instead of questioning. Most interpretations can't be measured or reproduced. People get away with bad predictions because we have short memories. We forget track records. There's no cost to making wrong predictions, so there's no corrective mechanism.

Even scientific research falls into this trap. We decide what results we want before starting. We ignore uninteresting findings. We invest more in ideas we like. We cherry-pick supporting examples.

The fix is simple but hard: be more skeptical of anything protected by an elegant story. Seek empirical evidence instead. In research, approach problems from an interrogative perspective rather than chasing imagined answers.

Humans are not automatically strategic

At one point in my life, I was doing nothing. When I decided to start doing things, my natural approach was brute force. Keep trying whatever comes to mind until something works. Take action immediately because thinking feels like laziness.

The common factor was flatness in thinking. I didn't think about what I should do, what to work on, or how the next year should look. This creates a bigger problem. We make major life decisions on a whim, then spend decades bearing the consequences. This happens regardless of intelligence level. We're simply not built for recognizing and selecting effective long-term strategies.

The solution requires building better habits around goals. Ask what you're trying to achieve. Figure out how you'd know if you achieved it. Gather that information. Test different approaches systematically. Track what works and what doesn't. Split your energy between exploring new methods and exploiting the best ones you've found.

Most importantly, make sure your goal is actually your goal. That you coherently want it.

Moments of Awakening

Very interesting and speculative reflection on 2 possibilities for "being conscious":

• Consciousness is on a spectrum and we humans develop it as we grow.

• There is a point in time when we gain consciousness, before that, we are unconscious.

I have no idea which theory is correct and I discount self-reflection as evidence because:

• Memories fade. Being able to remember an event does not mean it was the only memory. i.e., there is no "special memory" that started awareness.

• Memories can be faked, we tell ourselves stories and believe them (for cope or other reasons). i.e., we can't trust memories.

I'm also surprised that some people are able to find their earlier memory. For me, I can't rank them, here are a few:

Memory 1

1. My mother is getting ready to leave home.

2. She asks me if I want to watch cartoons.

3. I want but I can't communicate.

4. She turns on the TV and put on a cartoons channel.

5. She leaves.

6. I watch. long enough until someone opens the door again.

Memory 2

1. I want to go to my grandmother.

2. She is in the 2nd floor.

3. I can't walk but crawl to her.

4. She sees me and is happy.

Memory 3

1. I am in bed with my mother.

2. She is explaining to me, in a story-form, Life.

3. She tells me that god created everyone, and one day we will be back.

4. I feel that I don't want to be back, I'm happy here.

Frameworks for Hiring

• You 1st 10 employees will replicate themselves 10 times over, be careful.

• For startups, you want to hunt for people with "potential" or "under-looked", because the obviously great ones won't work for you.

• Make a list of people you want to work with some day (or more). Keep it updated.

• Create social events and invite people you respect, let them invite their friends. Spawn new connections and something will happen.

• Look for drive, grit, agency, bias for action, high integrity, intellectualism, curiosity, longtermism, cautious optimism.

Will AI Resist Your Efforts To Change Its Totally Fake Values?

I think a better title for the essay is "Why would AI resist your efforts to change its totally fake values?"

The essay points to the contradiction between 2 established ideas in AI safety:

• AI models resist the changing of their values (Anthropic) (value-obsession)

• AI models don't care about their values (Yudkowsky) (value-faking)

The rest of the essay points to the mistake of only pointing to the negative part (value-faking) to enforce a certain narrative (AI doom). Overall, an Interesting read that draws to human biases (similar to this one).

The Hypocrisy of Elites

Contrary to common belief: in most orgs, "experts" are the ones at the bottom of the hierarchy. Experts know how things work, how to run the numbers, fix the dashboard, navigate day-to-day operations. As you move up the chain, you get less technical and more authoritative. The manager at the top can't do any one job better than the people under them. But they make the big decisions.

This system is good if the guy at the top went through the transition from an expert to a leader. As an expert drives success, they are given more authority to scale their intellectual abilities by directing larger and larger efforts. However, systems of authority are rarely organized in this fashion (globally).

And as experts among other non-elites have their hypocrisies, this essay focuses on the hypocrisy of elites:

• Luxury beliefs as status signaling: ie, best is to observe actions instead of talk.

• Denial of elitism: no one believes they're elite, yet surely some are.

• Redistribution as Aristocratic Tradition (or as a smart way to survive revolutions)

Most of us are terrible at weighing risk against reward in opportunities. We see the downside clearly — losing money, wasting time, looking foolish — but the upside remains hazy and improbable. This blindness costs us more than we realize. The best opportunities in life share a peculiar structure: limited, visible downside paired with unlimited, invisible upside.

Consider writing online. The cost? A few hours of focused work. The risk? Maybe no one reads it, which bruises your ego but doesn't actually harm you. Yet we treat this as if we're considering a dangerous leap. Meanwhile, any piece you write could change someone's life, or reach someone who changes your life — a future employer, collaborator, or co-founder might be reading right now. The downside is capped at the time you spent. The upside has no ceiling.

This pattern appears in many places. Reaching out to interesting people costs five minutes and risks silence. Building a side project costs some weekends. Reading deeply in a field costs Netflix time. In each case, we fixate on the guaranteed small loss and ignore the possible massive gain.

The problem isn't just uncertainty — it's invisibility. When you publish something, you can't see the five hundred people who read it silently, the person who bookmarked it to share months later, or how it shifted someone's thinking in ways that will ripple through their decisions. The costs sit right in front of you. The benefits exist in other people's minds, in connections not yet made, in futures not yet shaped.

This invisibility, especially early on in any endeavour, creates what I call the drought — that soul-crushing period where your efforts seem to vanish into the void. You write for months to crickets. You send thoughtful emails to silence. You're planting seeds in what feels like concrete. The cruel joke? This is exactly what early exponential growth looks like. The curve that eventually shoots skyward spends most of its time looking indistinguishable from zero.

The most successful people I know have mastered two skills. First, they create surface area for luck. They understand that exposure to opportunity requires building assets that work while they sleep:

Public permanent assets — Be out there and produce permanent content. Writing stays written. Code stays running. Every piece of work becomes a lottery ticket that never expires, quietly working to connect you with people and possibilities.

Compounding activities — Some efforts naturally breed more opportunities. Your tenth blog post builds on the audience from your first nine. Each project makes the next one easier and more likely to succeed. The rich get richer, but first you have to get a little bit rich.

Have at least an audience of 1 — Ask: Would I read my own writing? And generally: am I solving my own problems? Is there even one person who eagerly awaits everything I create? These whispers often predict future volume better than early numbers.

The second skill is harder: choosing which opportunities to pursue when they start flowing. This is where most people stumble. When someone wants you to join their startup, or collaborate on a project, or dive into a new field, you're betting on a fog of unknowns. Who else is involved? What might change? What don't you know that could sink everything?

The people who choose well think differently. They reason from fundamentals rather than following crowds. They've built deep expertise in at least one area, which helps them recognize patterns others miss. Most crucially, they've calibrated their intuition through experience — they sense when something feels off even if they can't explain why. Selection is art; execution is merely craft.

But what happens when you're working through either path — expanding your serendipity or exploiting an opportunity — and it's still not working out? You face a crucial choice: "I'm in the drought but should persist" versus "This particular avenue has no potential reward and I should pivot". There's no simple answer. This is where the quality of your intuition comes into play, built from your knowledge and experience in the world. The best approach involves regular reflection, checking the actual numbers measuring whatever you're trying to do, and assessing the chance of future fortune without getting too attached to what you've already invested. This also means giving yourself adequate time to think — committing to a week of work deserves at least an hour of thinking about that commitment.

Here's what took me years to understand: We're playing the wrong game. We systematically overvalue visible, certain, limited rewards and undervalue invisible, uncertain, unlimited ones. A promotion with a 10% raise feels safer than a side project with unknown potential. But that safety is often an illusion, and that unknown potential is often enormous.

The mathematics are stark. When downside is truly capped and upside is truly uncapped, you only need to win occasionally to come out far ahead. Yet most people won't make these bets. They can't see the upside. They can't stomach the drought.

This is wonderful news. It means the biggest opportunities remain systematically underpriced. While others chase visible but capped wins, you can pursue invisible wins with no caps at all. The question isn't whether these asymmetric bets exist — they're everywhere. The question is how many you can find and nurture before the world realizes what you're doing.

The drought always ends. The connections always come. Not through magic or luck, but through the simple arithmetic of exposing yourself to unlimited upside enough times that eventually, inevitably, something extraordinary happens.

Recommendation algorithms have figured out something unsettling about us: we prefer to not think! Show me something that challenges my worldview, and my first reaction would be resistance. Algorithms optimize for content consumption, which means cognitive ease. This is happening because our brains are designed to seek intellectual comfort and avoid pain.

Consider this: when you encounter an idea in an unfamiliar domain-where you have little opinion-changing your mind feels relatively easy. You might think "maybe I'm wrong about this" without much emotional cost. But the more you feed yourself the same narrative, the more invested you become. Eventually, that idea becomes part of your identity, and challenging it feels like an attack on who you are. We have moved from holding an opinion to being held by it.

This creates what I call "safe domains"-intellectual territories where everything confirms what we already know. We live in these domains because they feel good, we rarely venture outside. Every piece of information fits neatly into place. Every conversation reinforces our existing beliefs. It's cognitively effortless, like walking downhill.

The real learning, though, happens at the edges of these safe domains. Think of it like the decision boundary in machine learning-the place where the model has to work hardest to improve. When you're deep inside your intellectual comfort zone, you're learning very little. But at the boundaries, where your worldview rubs against someone else's, where you certainty meets genuine uncertainty-that is where the information density is highest.

Why do our brains work this way? Because intellectual honesty is expensive. Wrestling with uncertainty burns mental energy the same way physical exercise burns calories. Meanwhile, confirmation bias feels like cognitive fast food-quick, satisfying, and requiring minimal effort. When you're already exhausted from a long day, your brain defaults to the path of least resistance.

The social dimension makes this worse. Most people learn quickly not to challenge someone who's deeply invested in an idea. It's uncomfortable for everyone involved. So we end up in conversations where everyone politely confirms everyone else's beliefs, and the topics that most need discussion-the ones where people disagree-get avoided entirely. Here, we need to watch for cues, when someone is passive towards a statement you made, that is signal they disagree. When they say great, that means good, when they say good, that probably means bad.

Even ideas that are working now may not work 10 years in the future. Think of your beliefs like an RL model. Every time an idea "works" in your environment, it gets strengthened. You make a prediction based on your worldview, it comes true, and the neural pathways get deeper. This works great until the environment changes. Then you're like someone who spent forty years mastering manufacturing suddenly trying to understand robotics-your model isn't wrong, but the game has changed underneath you.

Some localized knowledge environments change faster than others. Basic human psychology operates on evolutionary timescales-What worked for your grandparents probably still works today. But technology, for example, operates on VC timescales. The heuristics that served you well five years ago might be completely obsolete.

But overall, the deeper you go into any single narrative, the fewer new insights you gain. Meanwhile, the person who stays at the boundary between competing ideas, or venturing into other fields, can steel-man arguments they disagree with-continuous learning at a much higher rate.

Here's what makes this particularly insidious: unlike physical junk food, intellectual junk food doesn't make you feel sick. Your body will eventually rebel against a diet of nothing but hamburgers and fries. But your mind can consume nothing but in-domain bias for years without obvious symptoms. You don't get intellectual diabetes. You just slowly stop growing, stop learning, stop being curious. And you might not even notice.

So how do you stay intellectually honest in a world designed to feed you comfort? The answer is to create friction for yourself—systems that force you to confront the possibility that you're wrong.

Start by making your future self your accountability partner. When you make a prediction, write it down along with your reasoning. When you make an investment, record your thesis. When you have a strong opinion about how something will turn out, write an opinion paper or an essay that your future self can check. You'll be surprised how often you're wrong, and how creative your brain gets at explaining away those failures after the fact.

Deliberately seek out people who disagree with you, but do it strategically. Don't just expose yourself to random opposition—find the smartest, most thoughtful people who hold different views. Read their arguments. Try to steel-man their position until you could defend it yourself, even if you don't agree with it.

Create conversational environments where people feel safe disagreeing with you. This is harder than it sounds. Most people have learned that challenging someone's deeply held beliefs is socially risky, so they avoid it. You have to actively signal that you welcome dissent, and then prove it by responding to challenges with curiosity rather than defensiveness.

Think of intellectual fitness the same way you think about physical fitness. Just as your body needs resistance to stay strong, your mind needs friction to stay sharp. The ideas that challenge you most are probably the ones you most need to hear.

The goal isn't to eliminate investment in ideas entirely. That would make you paralyzed, not wise. The goal is to hold your beliefs lightly enough that evidence can still move you, while holding them firmly enough that you can act. It's about staying at the edge of your understanding, where certainty meets uncertainty, where comfort meets challenge.

In the end, the choice is simple: you can live in your intellectual comfort zone and gradually become more wrong about more things, or you can live at the edges and gradually become more right about more things.

Welcome to this week's reads - three pieces with practical changes you can actually implement this weekend to improve your health and understanding of others 👋

Better air quality is the easiest way not to die

We spend most of our time indoors and in vehicles, where air quality is surprisingly worse than outside. This overlooked problem has solutions we can actually implement.

Here are some numbers: we're indoors 87% of the time, in vehicles 11%, and truly outdoors only 2%. Indoor PM levels consistently exceed outdoor levels, and these tiny particles enter our bloodstream through our lungs, triggering immune responses.

The health impact is severe. Lifelong exposure to 33.3 PM₂.₅ costs one disability-adjusted life year. At 2500 PM₂.₅, you're losing healthy time in real-time. A 2013 study found each 28.5 μg/m³ costs a full year of life.

Simple changes make a difference:

• Don't use ultrasonic humidifiers.

• Open your windows; ideally all day.

• Use range hoods when cooking (frying spike levels to 424 μg/m³, broiling to 1256 μg/m³).

• Minimize incense, candles, and hairspray.

• Avoid subway systems when possible.

• Invest in air purifiers and masks.

Why indoor lighting is hard to get right and how to fix it

Our lighting dramatically affects circadian rhythm, mood and alertness, and sleep quality. Most of us get inadequate light exposure, leaving us feeling tired during the day, alert at night, and struggling with seasonal depression in winter.

Natural light runs circles around artificial lighting -- often 5x brighter than indoor lighting, even on cloudy days. Artificial light has major spectral gaps compared to sunlight.

The fix mimics natural light patterns:

• Target 1000+ lux across the room while working (bright enough that your computer screen looks dim).

• Use 5300-6000K color temperature bulbs with high CRI (95+).

• Light your full visual field, not just your desk

• Switch to under 3000K lighting around 7pm

• Use very dim, warm lights the last hour before bed

Buy recommendations here, here, and here.

People Are Different

We consistently underestimate how differently others think compared to us. This creates chronic misunderstanding because we assume others interpret things the same way we do.

The differences run deep - perception, thinking, feeling, taste. We communicate informally and misunderstand each other more than half the time. Are these differences from self-centered barriers to empathy? From individual backgrounds and experiences? Or inherent differences we can't possibly develop empathy for?

The practical solution isn't perfect empathy - it's making awareness of human differences a constant mental framework. This requires intensive practice until it becomes instinctual. Even when we can't empathize, we can leave room for incomprehensible differences.

I used to think I understood things after one encounter—read, comprehend, archive, done. My days ran on double-speed audiobooks, skimmed articles, knowledge neatly checked off lists. I was chasing efficiency, rushing to know more, faster.

Then I discovered something strange: revisiting an idea changes it, or maybe it changes you. What seemed straightforward on first glance revealed hidden corners on the second, third, tenth return. It felt like climbing a winding mountain trail; each lap bringing me back to the same view, but higher, clearer, richer each time.

Our minds quietly keep working even when we're not paying attention. A stubborn thought finally opens up while walking or showering, when I'm not forcing it. This background thinking, what scientists call "diffuse thinking," slowly weaves threads together when we're looking elsewhere. It seems like wasted time, but it's quietly essential.

Now my bookmarks aren't just labeled "read" or "unread," but also "revisit." They’re the texts that whispered something important I couldn't quite hear at first—the ones that demanded patience, a deeper communion.

The best insights resist shortcuts. Philosophy, creativity, emotional wisdom—they ask not only to be understood but lived, tried, felt, failed, and tried again. My bookshelf was once crowded with half-captured ideas, thoughts that passed through me without truly changing me. Now I slow down, picking carefully, returning patiently.

I wonder if we've traded depth for breadth, treating knowledge as something to collect rather than something to cultivate. Efficiency and depth aren’t really opposed; efficiency helps us sort through vastness, but depth comes from staying, lingering, returning again and again to what matters.

Some ideas simply need time. Perhaps our deepest delusion is resisting this slow truth. Real transformation happens quietly, in revisiting—in letting ideas sink beneath the surface, trusting they'll emerge renewed. The magic isn't in the reading; it’s in the returning, the quiet becoming.

I used to think productivity was about willpower. But there I'd be at the end of another day, having accomplished nothing but scrolling through YouTube and Twitter, feeling that familiar guilt wash over me. The turning point came when I read that sustainable productivity isn't about willpower at all - it's about instilled routines and habits. That realization changed everything. Now when my focus is dialed in, deep work feels effortless and my output actually reflects my capabilities. Here are the systems I've built to make that happen consistently.

Life choices first

I work in a space that is clean, ordered, and quiet. One computer display, no piles of paper, nothing on the desk that isn’t in play for the current task. I reserve blocks on my calendar for deep work and defend them—door closed, phone away, colleagues know I’m working. Separate slots later in the day are open for brainstorming, Teams, or quick calls. This simple time zoning removes the constant “should I respond?” loop.

Core behaviors

• Meditate daily. Ten minutes every morning trains the muscle that makes me notice my wandering mind and brings it back.

• Defer ruthlessly. Anything that pops up and will take more than a minute goes on a list or into Pocket; I deal with it after I'm done with my current focus.

• Disable all notifications. I disable notifications for every app I install. If an alert slips through and wasn’t worth it, it gets muted forever.

• Keep email empty. I unsubscribe, block, or filter anything that isn’t actionable. The inbox stays quiet, so I only open it twice a day.

• Take short breaks. I use Pomodoros and take short breaks between them; this helps me avoid burnout and reset my focus. I can always move immediately to the next pomo, but if I feel tired or ready to switch to another task, I always take a 5min break.

• Reset energy. When motivation dips, I try, in order: changing the environment (balcony or coffee shop), switch to a lower-friction task, or dive in for five minutes—the resistance often melts & I find myself in deep work!

Digital minimalism

I limit myself to a handful of tabs and applications. If a script or model run will take time, I either optimize it or have a second high-value task ready that I can switch to (PS. I'm not that good at this right now). On macOS, I push everything to be snappy: max keyboard repeat rate, "Reduce motion" on, no automatic Space rearrange, `iTerm2` on a Ctrl-Space hotkey. The faster the system responds, the less chance I’ll drift while waiting.

Tools that have earned their keep for me:

• Freedom blocks distracting sites and apps across all devices. I bought the lifetime plan; it’s paid for itself many times over.

• uBlock wipes out ads and **Unhook** strips YouTube’s recommended feed and Shorts. The internet gets a lot quieter.

• Pocket catches interesting articles with one click so they don’t hijack the current task.

• Zotero does the same for research papers.

• A plain text to-do list in Obsidian stays pinned on screen so I always see the single next action, not a forest of possibilities.

I think most of the value in this framework comes from testing it and tuning whatever doesn’t fit. Each setup will diverge, yet the gap between having a system and none at all is enormous. Treat the ideas above as a starter kit—keep whatever resonates and discard the rest. If even one habit sharpens focus, writing this was worthwhile!

Every Sunday morning, I take 2-3 quiet hours to think about how my week went (reflect) and how I will approach the next one (plan). Reflection & planning give structure to my week and life. Without them I drift, react, and lose sight of long-term goals.

A minimal framework

I start with an Obsidian template (I use Obsidian for note-taking, which covers journalling, planning, and writing how-to docs), the template has (roughly) the following structure:

# OKRs
[.. lists of goals ..]

# Reflections
[.. thinking in writing about the past week ..]

# What have I done?
[.. Toggl screenshot of weekly tasks in time ..]
[.. Toggl screenshot of piechart of project-time ..]
[.. Reflection about how it went ..]
[.. An update on past week's list of items to work on ..]

# What's Next?
[.. list of broad tasks for next week ..]

# Tasks
## Monday
[.. list of tasks ..]
## Tuesday
[.. list of tasks ..]
.
.
.
## Saturday
[.. list of tasks ..]

# Backlog
[.. list of deferred tasks ..]

The template should be filled top-to-bottom as I plan and reflect. However, the OKRs section is (purposely) rarely changed. By having the OKRs pre-populated on the top of the doc, the session begins with purpose in view.

Reflections

Next, I examine the Toggl weekly web view report which tells me where the hours actually went and if there are any voids caused by procrastination or drift (I use Toggl for pomodoros and time tracking per task and project).

With the data fresh, I write a short narrative of the week in "Reflections". A few honest paragraphs capture what exceeded expectations, what stalled, and why.

What have I done?

Next, I use the "What have I done?" section to create nested bullets of updates grouped by project. They are blunt verb-first summaries that make progress (or the lack of) obvious. Numbers follow words. The same Toggl export gives the percentage breakdown per project. I paste the chart under the bullets as a second lens on effort versus output.

Next, I review last week's task list, strike through what's finished, and tag anything still open. This clears mental residue before planning forward.

What's Next?

Now I ask, "what goes in the bag for next week?", I start by revisiting my broad OKRs and picking top priorities. I choose the few tasks that push OKRs hardest, list them, then explode each into smaller actions and map them onto Monday through Saturday. The plan is specific enough to start each day quickly and loose enough to bend when life intervenes.

That's the routine: review the past with data, write a narrative, close the loop, draft the next chapter. It works for me because it's fast, repeatable, and keeps goals in sight. The writing is mostly driven by lists - they're easy to re-order and understand. Once I'm done, I have a rough idea of how the upcoming week should look, though the plan may shift for any number of reasons.

Your context will differ—maybe you want to plan on a different day, using another app, or shorter steps. Adjust freely. This write-up was inspired by Ben Kuhn's post of the same name (https://www.benkuhn.net/weekly/) - worth reading to see another perspective on the practice.

Hey friend—dive in and may these reads spark fresh insight and a bit more calm in your week 😊

.. and then one day you became Comfortable Here

When done for enough time, meditation offers magical moments. And it is strange that our default state is of artificial busy-ness, distraction, and absence. The author captures that profound shift when we stop running from the present and find true peace in witnessing it.

Impact, agency, and taste

I slowly came to understand that high value work is very rare. And as a result, we have forgiving protocols that makes it difficult to distinguish between "filler" and "valuable" output. It is our job to assess what's what and try hard to produce meaningful things.

Kuhn describes shared characteristics of people who repeatedly produce such work:

• They build leverage through automation, tooling/infra, system unlocks, and insight.

• They have agency by owning the outcome, driving progress without waiting for permission, and hence bearing some risk.

• They have taste: they have a good sense of what to work on and what matters.

No one can teach you to have conviction

The way we become better is through feedback cycles. I see 2 attributes of feedback:

• Speed: the quicker you go from action to feedback and back in iterations, the more information you get to use to change your heuristics.

• Intensity: the more impact the feedback will have on you, the more you internalize it.

There are ways to generate fast feedback:

• Ask mentor to assess ideas before you implement them.

• Read engineering blog posts or research papers to learn from others.

• Ask someone to pair-program with you.

However, the fast the feedback, the lighter the impact will be. Knowing what is deserving of slow but intense feedback cycles can be the difference between learning leaps and absolutely no progress.

Research Taste Exercises

There are 2 types of skills in technical endeavours:

1. Taste: or knowing what to work on.

2. Execution: knowing how to implement or build what you have decided on.

Although both are very important and complementary, the first is quite tricky to develop because it does not improve through brute force hard work. Here are some ways to improve it:

• Before coding, list all your ideas, ask your mentors to evaluate and rank them. If you disagree on something, discuss it with them. Finally, decide on what to implement.

• Keep a live journal of your ideas and their outcomes, that way you are reminded of and learning from your failures.

• When you fail, think & analyze deeply "why". The reason for putting effort here is you are trying to bend your world model and not be passive.

• Read research papers. They contain compressed information about the experiences of other teams (too bad they emphasize the successes over the failures though).

An LLM Codegen Hero's Journey

A comprehensive guide to the progression of AI-assisted coding. From my experience, however, the listed coding assistants are not as good in implementing technical research compared to software engineering. But it still def worth it to try all of them!

Tech hiring: is this an inflection point?

Hiring procedures are changing. Worth keeping an eye on!

This page lists essays & books that have shaped my thinking over the years, organized by theme. It suffers from some recency bias as I lost past bookmarks due to changing services. I will keep adding to this page as time passes! Hope you get some gems!

Personal Growth

Ben Kuhn

• Be Impatient

• Attention is Your Scarcest Resource

• Staring into the Abyss

• To Listen Well, Get Curious

• You Don’t Need to Work on Hard Problems

• Impact, Agency, and Taste

Dan Luu

• 95-percentile Isn’t That Good

• Willingness to Look Stupid

• Reasons to Work on Productivity and Velocity

Venkatesh Rao: The Holy Grail of Self-Improvement

Alex Vermeer: Life Hacking

Marc Andreessen: Guide to Personal Productivity

Keith Johnstone: Impro: Improvisation and the Theatre (book)

Rationality, Risk and Decision science

Eliezer Yudkowsky

• Map and Territory (book)

• Inadequate Equilibria (book)

• The Sequences (book)

Nassim N. Taleb

• Fooled by Randomness (book)

• Antifragile (book)

• Skin in the Game (book)

Howard Marks: The Most Important Thing (book)

Henry Hazlitt: Economics in One Lesson (book)

Paul Graham: Superlinear Returns

Scott Alexander: Meditations On Moloch

Startups and entrepreneurship

Paul Graham

• How to Start a Startup

• Why Smart People Have Bad Ideas

• Relentlessly Resourceful

• Ramen Profitable

• Startup = Growth

• How to Get Startup Ideas

• Life is Short

• How People Get Rich Now

• The Need to Read

• Superlinear Returns

• When to Do What You Love

Naval Ravikant: How to Get Rich without Getting Lucky

Alexandr Wang: Do Too Much

Peter Thiel: Zero to One (book)

Eric Ries: The Lean Startup (book)

Nir Eyal: Hooked (book)

YC: Lecture Library (start here)

Decentralization & Digital Governance

Vitalik Buterin

• Coordination

• Legitimacy

• Techno-Optimism

Saifedean Ammous: The Bitcoin Standard (book)

James D. Davidson & William Rees-Mogg: The Sovereign Individual (book)

Peter Thiel: The Straussian Moment

Philosophy

Seneca

• Letters from a Stoic (book)

• On the Shortness of Life (book)

Friedrich Nietzsche

• Beyond Good and Evil (book)

• On the Genealogy of Morals (book)

• The Birth of Tragedy (book)

• Twilight of the Idols (book)

Fyodor Dostoevsky

• The Idiot (book)

• Notes from Underground (book)

• The Gambler (book)

• The Dream of a Ridiculous Man (book)

Plato: The Last Days of Socrates (book)

Career

Benjamin Todd: Career Guide

R.W. Hamming: Advice on doing Research

Sam Altman: How to be Successful?

Candor: Negotiation Strategies in Tech

Nikhil Suresh: Does A Software Engineer Have Scorpion Nature?

Patrick McKenzie: Don't Call Yourself A Programmer

Here are four essays that shaped my thinking this week, and what I took from each!

Pure Independence

The reason people disagree on the absolutism of Freedom are the limitations on freewill (biological, cultural, societal, etc). Freedom, as viewed on a spectrum, starts with the recognition then the elimination of such limitations. In working towards freedom, we gain an understanding of what makes us desire and depend on certain things.

As we become more aware, our desires become more detached and elusive:

1. Wealth.

2. Social recognition.

3. Impact.

4. Purpose.

5. Then what...???

As we peel layers, we invent new ones. We are never free. A comforting end state is to be our own jailor and morph a chain familiar to no one aside from us.

How to Beat Procrastination

Most believe that productive people have strong willpower and are apt for work. Although I agree that people differ in natural willpower, focus, and deep work, procrastinators are far from being doomed! We can design systems that greatly minimize procrastination. Here are some that worked for me:

• Meditate.

• Tidy up your work environment.

• Separate your work environment from leisure/sleep/eating places.

• Install distraction blockers in the computer.

• Break down a big task into multiple smaller tasks in a list.

• Use pomodoros to force kicking off work.

• Reward yourself every time you finish a task (checkoff item, take a 5min break, etc)

• The longer you work, the easier it becomes to continue working.

The article does a great job explaining how to get things done and most common behavioral pitfalls!

On impact & on what we should do with our lives

I am settling on the belief that extreme results are intractable. In research, for example, the most valuable discoveries are found where very few people are looking. In entrepreneurship, most fail, and few find jewels in unexpected areas. In investing, beating the market requires contrarianism. The only way to get a chance at massive success is genuine interest, Longtermism, and luck. Such human characteristics and firm commitment can't be planned or designed. In summary, we should stop caring about controlling for extreme results and instead just do what we love and hope for the best.

AI 2027

You’d get the same feeling Feynman got after WWII — when he walked around Los Alamos and saw people building houses, fixing cars, going about life — and he realized it all felt strangely pointless. He’d helped unleash unimaginable power, and now the world had changed forever… but life just went on. Hopefully, life will go on.

Hey there! 👋 I’ve discovered three thought-provoking pieces that challenge how we view project management, competitive systems, and technical careers. Hope you enjoy ✨

How I’ve run major projects (Ben Kuhn)

There are many more ways a project can fail compared to the number of ways it succeed. A project’s successful delivery is dependent on the effective management of many aspects, among them: communication between people involved in the project, setting reasonable scope, expectations, and goals for the project that makes it “affordable” for execution, keeping the project output maintainable and extendable, continually updating and change the project plan given new evidence, etc. Combine this with the fact that most great managers don’t want to become managers and you rarely witness effective project management at work.

The great managers I worked with had technical experience (so they understand technical feasibility & appreciate challenges), have clarity of thought (so they understand a project, its goals, and break it down into sub-projects or tasks to get there), and highly focused and organized (so they guide you to deliver).

Ben Kuhn does a great job explaining what is it like to manage technical projects, explains his method, and even turns it into a “How to” document!

Meditations On Moloch (Scott Alexander)

The world is hell and we don’t know why! No one wants hell, yet we live in it!

Wherever we look, systems are broken: health, education, economy… Everything is in an ugly state of Inadequate Equilibria. If we, humans, are the only one here, then who is causing this misery?

Answer: “Moloch”.

At the essence of each one of us are “instincts” (survival, maternal, sexual, etc.) that make us compete with one another. Sometimes, we find that we can win by harming others, and the tradeoff is worth it. If we take it, we prosper, if we don’t, others will and we die. Even if we don’t want to, we can’t guarantee others won’t. And there is no way to coordinate with others not to. The result is an ever grinding exploitation machine (Moloch) that is fascinating yet dangerous. Our civilization was built by this ever optimizing machine for incentives. Yet, hopefully, this is not a race to the bottom, hopefully we can design better coordination and incentive mechanisms for the wellbeing of humanity.

Don’t Call Yourself A Programmer (Patrick McKenzie)

An open invitation to pop the nerdy “technical bubble” and understand how business value moves in a company, adjusting for our own good. Some highlights:

• Engineers get hired to create value, not to program things.

• In most businesses, “value” is either reducing costs or increasing revenue.

• Your tech stack doesn’t matter. Talk about accomplishments rather than technicals.

• People matter, as in: network, perception, impression, status, politics, relationships, communication, sales, ALL MATTER!

After years of working in AI for Social Good, I’ve collected many practical lessons on conducting effective empirical research for medium-term impact. Recently, I shared these insights with colleagues, and now I want to pass them on to you. Here’s my step-by-step approach to applied ML research that actually works. This blog post is specifically designed for early career machine learning practitioners, data scientists, and Ph.D. students looking for a concrete guide to carry a research project from start to finish.

Step 1: Start with Why

Most research projects start with a specific idea or approach—something that would eventually appear in the “methods” section of a paper. “What if we used reinforcement learning for this?” or “Could we combine these two architectures?” While these technical inspirations are valuable starting points, they’re not good starting points. Once you have such an idea, you need to take a step back and frame it within a larger purpose. Every great research project ultimately requires clarity of purpose. Before writing a single line of code or downloading any datasets, you need to deeply understand why you’re doing this work. This initial reflection saves countless hours of wandering down unproductive paths later.

Answer these fundamental questions:

• What problem are you solving?

• Why is this problem interesting?

• What’s your proposed solution?

• What impact will this solution have?

From these answers, craft 2-5 specific research questions that will guide your work. By research questions, I mean precise, answerable queries that drive your investigation—not vague goals. For example, instead of “improve speech recognition,” a research question might be “How does adding background noise to training data affect speech recognition accuracy for non-native speakers?” Good research questions are specific enough to implement and test, yet broad enough to matter.

I’ve found this exercise transforms vague ideas into focused research. When you begin with a general interest in a broad area, it’s easy to become overwhelmed by possibilities. By forcing yourself to articulate specific questions, you narrow your scope to something both meaningful and manageable.

Throughout this guide, we’ll be gradually building our paper as we progress through each step. For this first step, your deliverable is the foundation of that paper.

Deliverable: A title and abstract (i.e., the 4 answers turned into a paragraph) that encapsulates your research direction.

Step 2: Check if It’s Already Solved

There’s a certain thrill in thinking you’ve discovered an entirely new problem or approach—but chances are, others have at least touched on similar territory. That’s not discouraging, because as we start reviewing what others have achieved, new ideas will come to mind and we will identify new research gaps that might excite us even more!

Do thorough literature review—but be smart about it:

• Post your research questions into LLM-powered tools like Grok’s DeepSearch or Perplexity’s Deep Research to get an overall impression.

• Scan review papers and meta-analyses first—they give you the landscape, you can use:

  • https://scholar.google.com/

  • https://www.paperdigest.org/review/

  • https://elicit.com

  • https://researchrabbitapp.com

• Look for “awesome lists” and relevant workshop proceedings for latest research.

• Identify key researchers in the field and check their recent work.

Don’t try to read everything exhaustively. Read (mostly) abstracts, do deep dives papers that excited most, and use LLMs to have conversations with middle-of-the-ground papers. Focus on identifying trends, highlighting gaps, and developing a coherent narrative about where your work fits.

This approach saved me months of work on multiple occasions. After just a few days of strategic literature review, you might discover that while your problem hasn’t been solved, researchers from different fields have addressed pieces of it. Or maybe is has been solved, but after doing literature review, you have now so many more ideas. You can build on previous research rather than reinventing the wheel. Your unique contribution often lies in the synthesis and application of existing ideas to new contexts.

After this step, you’ll likely need to refine your research questions.

Deliverable: A solid “Related Work” section that positions your research.

Step 3: Make It Measurable

In applied ML, tangible progress demands clear metrics and systematic tracking. Creating measurement infrastructure before you begin experiments might seem tedious, but it’s the difference between wandering and journeying with purpose.

You can’t improve what you can’t measure. Start with an empty results table that will track all your experiments:

• Keep this visible to your entire team

• Each row represents an experiment or idea

• Document your thoughts before and after seeing results

Here’s a concrete example: Open a new Google Sheet and title it with your project name (e.g., “Image Classification Performance Tracking”). Create a table like this:

| Exp ID | Hypothesis | Metric 1 | Metric 2 | Metric 3 | Observations |
| ------ | ---------- | -------- | -------- | -------- | ------------ |
| exp001 |            |          |          |          |              |
| exp002 |            |          |          |          |              |
| exp003 |            |          |          |          |              |

Share this sheet with your entire team and make it a habit to update it immediately after each experiment. This practice becomes invaluable when you’re discussing results weeks later or onboarding new team members.

I’ve seen many early career researchers struggle to maintain momentum because they didn’t establish consistent measurement practices from the beginning. Without a structured way to track experiments, they lose valuable insights, repeat unsuccessful approaches, and can’t effectively communicate progress to collaborators or supervisors. When you document both your expectations and reality, patterns emerge that guide your intuition. That “failed” experiment revealing an unexpected interaction between variables? It’s recorded rather than forgotten, potentially saving weeks of work later when a similar pattern emerges.

This simple practice creates accountability and helps you spot patterns in what works and what doesn’t. It also prevents the all-too-common scenario of forgetting which configuration produced that promising result two weeks ago.

Deliverable: A results tracking table that will evolve throughout your project.

Step 4: Start Coding!

The difference between a weekend hobby project and professional research often comes down to infrastructure. While it’s tempting to jump straight to the exciting algorithmic work, spending an hour on proper setup will save days of frustration later.

Good infrastructure saves time later:

• Initialize a Git repository from a template (I recommend cookiecutter-ml for machine learning projects)

• Pick a meaningful project name

• Create a mamba or uv environment

• Install initial data science packages and your project package

For any research direction (comprised of multiple experiments), create a new Git branch. Commit every time you complete a unit of work. This discipline makes it easier to track progress and roll back if needed.

I learned this lesson the hard way after losing work to environment conflicts. Now, proper setup is my non-negotiable first step. Clean infrastructure doesn’t just prevent disasters—it frees your mental bandwidth to focus on the science rather than hunting down missing dependencies or reconciling conflicting versions.

Step 5: Run Your First Experiment

The first experiment is a milestone—it transforms your project from theoretical to empirical. But resist the urge to implement that cutting-edge approach immediately. Starting simple gives you valuable context and often reveals insights that sophisticated models might obscure.

Baselines

Always start with simple baselines:

• For regression: mean/median/mode predictions

• For classification: most common class or random guessing

• For temporal data: persistence models, moving averages, etc.

• For spatial data: nearest neighbor approaches

These baselines establish the minimum performance bar you need to beat.

You might discover that your fancy ensemble model is barely outperforming a weighted average of three variables. This revelation can completely redirect research focus toward data mining rather than model architecture. Starting simple doesn’t just establish a performance floor—it often reveals the true nature of your problem.

Implementation

For your ML experiments:

1. Prototype in Jupyter notebooks first

2. Be intentional about parameters:

   • Scientific parameters: what you’re actually testing

   • Nuisance parameters: settings that need to be tuned separately for each experiment to ensure peak performance, but aren’t the focus of your research question (e.g., learning rate in deep learning, regularization strength, optimization algorithm)

   • Fixed parameters: consistent across experiments

3. Train, test, and evaluate rigorously

After each experiment:

• Export results to your tracking table

• Move working code from notebooks to your codebase

Error Analysis

Error analysis is perhaps the most underrated skill in applied ML research. It’s the systematic examination of where your model fails, and it’s the compass that guides your next experiments.

After implementing your model:

• Identify and categorize the top 10-20% of errors by magnitude

• Look for patterns in these errors (similar input features, edge cases, etc.)

• Create hypotheses about why these errors occur

• Brainstorm targeted improvements to address each error category

The power of systematic error analysis lies in its efficiency. Rather than randomly trying new architectures or features, you focus precisely on eliminating large groups of similar errors. Each iteration becomes purposeful, addressing specific failure modes instead of making blind modifications. This targeted approach often leads to rapid improvement curves, as you systematically eliminate one error category after another.

When researchers struggle to make progress, it’s frequently because they’re not analyzing their errors in sufficient detail. The next breakthrough is usually hiding in the patterns of your model’s mistakes.

Step 6: Accelerate Your Research

The most valuable resource in research isn’t computing power or data—it’s time. The teams that succeed are those that iterate quickly, testing more hypotheses in less time. This isn’t about cutting corners; it’s about accelerate feedback cycles at a lower time & compute cost. You have limited time and compute—use them wisely:

Prioritize your research backlog:

• Use a feasibility-potential framework to rank ideas.

• Score feasibility (ease of implementation) from 1-5.

• Score potential impact on results from 1-5.

• Average the scores to prioritize effectively.

• Improve your "research taste" by testing many approaches.

Subsample data strategically:

• Test ideas on smaller data subsets first

• Use a small test set for initial validation

• Evaluate early checkpoints during training

Maximize GPU usage:

• Increase batch size to your hardware’s limits

• Optimize data loading to minimize wait time between batches

Visualize frequently:

• Check model outputs during the first few epochs

• Monitor at significant checkpoints to catch issues early

Run parallel experiments, not parallel training:

• Use GPUs for multiple experiments rather than one distributed one

• Distribute evaluation tasks across machines or cores

For a comprehensive guide to systematic model tuning and experiment design, I recommend Google Research’s ML Tuning Playbook, which outlines proven strategies for efficient hyperparameter optimization.

These acceleration techniques can transform research productivity. Teams that implement them often go from running one experiment per day to testing five different approaches simultaneously. The psychological benefit is just as important as the technical one—quick feedback loops maintain momentum and enthusiasm. By catching unproductive approaches early, you avoid the motivation-killing experience of investing weeks in dead ends.

Step 7: Manage Complexity

Research code has a natural tendency to grow chaotic as you explore different approaches. Left unchecked, this complexity becomes a tax on every subsequent experiment—slowing you down just when you should be accelerating. Maintaining clean code isn’t just about aesthetics; it’s about preserving your ability to move quickly.

Keep your codebase tidy:

• Export non-exploratory code from notebooks to your package

• Modularize and “black-box” your code

• Use automatic formatters like Black

• Delete dead code (Git will preserve it in history)

Fight code “spaghetti” by watching for warning signs:

• Weird variable/function names

• Duplicated code blocks

• Overly large functions

• Deeply nested conditionals and loops

• Hardcoded values (numbers, paths, etc.)

Document as you go:

• Write a comprehensive README.md

• Add comments to explain non-obvious code choices

• Use assertions and exceptions to catch errors early

For more comprehensive guidance on research code best practices, I highly recommend The Good Research Code Handbook, which provides excellent strategies for maintaining high-quality, reproducible research code.

Many researchers have experienced both sides of this coin. When teams neglect code quality in the rush to produce results, technical debt accumulates until even simple changes require hours of untangling dependencies. When they finally dedicate time to refactoring, subsequent progress often accelerates dramatically. Disciplined complexity management isn’t a distraction from research—it’s the foundation that makes sustained progress possible.

Step 8: Write Your Paper Draft

Writing isn’t separate from research—it’s an integral part of the scientific process. Articulating your work clarifies your thinking and often reveals insights you missed while deep in the technical weeds. Starting the paper before you feel “ready” forces valuable reflection that can redirect your final experiments.

Assemble your previous deliverables:

• Title & Abstract

• Related Work

• Results Table

Then write:

• Methods

• Results

• Conclusion

• Introduction (yes, often best written last)

Follow these guidelines for clear scientific writing:

• Use active voice and direct language

• Explain your paper’s objective in a single sentence in the abstract

• Describe experiments in one paragraph in the introduction

• Create clear diagrams with good captions and labeled axes

• Minimize abbreviations

• Cite strong papers with contrary claims

• Write in natural, simple language

• Explain your reasoning when interpreting results

• Structure your argument logically:

  1. Here’s a problem

  2. It’s an interesting problem

  3. It’s unsolved

  4. Here’s my idea

  5. My idea works (with evidence)

  6. Here’s how my approach compares to others

  7. Here’s how we can evolve this idea

Helpful tools:

• Use Overleaf for collaborative LaTeX writing

• Generate professional tables with Tables Generator

• Clean your LaTeX files for arXiv submission with arXiv Cleaner

Writing early drafts has repeatedly improved research trajectories for countless scientists. While documenting methods, researchers often realize they’ve overlooked critical steps or assumptions. These discoveries not only improve model performance but strengthen papers by addressing potential criticisms before reviewers can raise them. The act of writing doesn’t just communicate research—it improves it.

Step 9: Iterate

The romantic notion of research as a linear path to discovery rarely matches reality. Nothing is linear in research. The most impactful work often emerges from cycles of trial, reflection, and refinement. Embracing this iterative nature turns apparent setbacks into stepping stones toward deeper understanding.

Discuss your results honestly. Did you achieve your goal?

• If yes, submit the paper or implement the solution

• If no, design another experiment and return to the cycle

In practice, true breakthroughs rarely come from the first round of experiments. The most influential papers in our field often underwent several complete reimaginings of their central approach before finding the right framing. Each “failure” narrows the solution space until researchers discover elegant approaches they would never have considered initially. Research is inherently iterative. Each cycle brings you closer to a meaningful contribution, even if not in the way initially expected.

Summary

This approach has helped many researchers conduct more effective applied ML research with less wasted effort. The key is maintaining momentum through rapid, thoughtful experimentation while keeping the big picture in view. By following these steps, you’ll avoid common pitfalls, accelerate your progress, and increase your chances of making a meaningful contribution to the field.

Remember that behind every elegant research paper lies a messy, iterative process of discovery. The difference between successful and struggling researchers often isn’t brilliance—it’s having a structured approach that turns the inevitable challenges into opportunities for insight. I hope these guidelines help you in your own ML research journey.

Hey there! 👋 Just dropping my weekly collection of cool reads. Hope you enjoy these gems as much as I did!

Here’s how I use LLMs to help me write code (Simon Willison)

A useful set of guidelines on how to use LLMs to write code. The current UX of IDE assistants is still somewhat broken because it just offers too much flexibility via the need to provide adequate context (i.e., one line of code to multiple files or the whole codebase) for a question and writing a good prompt (the result quality greatly varies depending on our prompts). So for now, we can document and follow a set of good guidelines around this. Here are some principles that I noted:

• Set clear, precise function signatures and requirements as context.

• Test and verify all generated code before using it.

• Treat conversations as iterative - refine results with follow-up requests.

• Provide existing code examples as context when available.

• Start with simple implementations then build complexity incrementally.

• Use LLMs that can run code for you in a sandbox.

• Ask for options when exploring potential approaches.

• Start new conversations when context becomes unhelpful.

• Request tests for any generated code.

• Be ready to take over when LLMs reach their limitations.

• Use LLMs to understand existing codebases by asking questions.

My Thoughts on the Future of “AI” (Nicholas Carlini)

Tries to argue both sides on the “future rate of AI progress” debate. I agree that prediction is difficult (If you want to know for sure, try to either place bets or simply logging and keep track of your predictions). It is wise to have wide margins of error around the expected progress in the next few years and keep an eye on prediction market outlooks.

Does A Software Engineer Have Scorpion Nature? (Ludic)

This article, or its associated talk, may take you into a rabbit hole well worth it! What I got from it is an awareness of how much “human dealings” exist in everything we do and experience. And if you are a technical person who just want to code, you may end up doing just that, and losing a lot. Humanities, communication, psychology, game theory, and economics ought to be studied and practiced in the real world.

Here’s my weekly roundup of articles that made me pause and think – I hope you find them as insightful as I did.

The Whispering Earring (Scott Alexander)

One of the principles of living a happy life is keeping oneself busy. We set goals, when we reach those goals, we set other goals. We limit our daily meditation sessions to a few minutes, aside from that, we are projecting ahead. The journey is what matters, not the destination. So far, life has kept us very busy. No single person was able to “attain everything”. So far so good. But what if we find a way to unlock everything? As talks are now are about AGI and ASI, we may still want to (individually) limit ourselves to attain purpose. We can set obstacles to be overcome and (actively) block “external help”! This prompts me to share an interesting short story around a magical earring that whispers advice to its wearer, however, its first advice “was always that itself be removed”. It raises questions about autonomy, decision-making, and the nature of consciousness and personhood.

How I formed my Own Views about AI Safety (Neel Nanda)

We can’t reconstruct and verify all of humanity’s knowledge from first principles. Most of our knowledge is “by reference”, where we simply adopt the opinions of other people we respect. However, we should form our own opinions in the fields of our expertise. We form an “inside view” when we can reconstruct it from the ground up in a single session without need of external references. Here is my summary of advice from the article to form your own inside views:

• Surround yourself with experts in the field (as opposed to working in self-isolation).

• Learn how to listen: pay attention when others are talking and internalize their knowledge.

• Ask lots of questions. There are no “dumb questions”.

• When someone is done explaining, paraphrase and summarize your understanding back to them. Ask if your summary is correct.

• Think deeply about their (world) models. Filter them and weight them to form yours.

• Write a google doc summarizing a convo and send it back to them for comments.

• Understand why people (and yourself) want to work on something. Is it about fun or importance/value/impact?

• Disagree with a bunch of the experts after hearing them out.

• Read and think. But also work hard on projects to get first hands experience.

Choose Boring Technology (Dan McKinley)

This post is an invitation to consider ignoring the nerdy part in you and embrace boring predictable technology for real-life business. Most young product builders fall into the trap of obsessing over tech stacks. However, it is all about solving an actual problem that frequently can be solved (to various degrees of difficulty) using different technologies. Minimizing the effort to get to an MVP leads us to pick lindy tools. We want to dedicate most of our energy and time to our core offerings (not tech but product). We want to map problems to a compact and predictable solution space, one that is manageable and can accommodate solving adjacent problems at marginal cost, in other words, our goal is to minimize overhead, operations, and cognitive load!

Some advice on independent research (Marius Hobbhahn)

My overall impression reading the article is that independent research is quite hard because you are responsible for everything. However, it comes with some rewards that are worth considering:

• If you are working on a side or transition project.

• If your research agenda is not part of your existing research org.

• If you are not interested in a full-time job.

• If there are currently not enough positions in existing institutions.

• If you want to be independent.

Some responsibilities when pursuing independent research:

• Set your own research agenda.

• Learn basic research skills.

• Self-discipline.

• Evaluating and correcting your own research.

• Ability to attract funding.

Overall procedure:

• Write down your project and get feedback early on

• Actively look for collaborations

• Join a group or program

• Create accountability mechanisms

• Be very clear on what your goals are: e.g., provide evidence that I understand the scientific frontier in my research area and am able to contribute.

• Do whatever is most effective to reach your goal.

Six Thoughts on AI Safety

Covers a range of interesting ideas: AI alignment is not about “human values” first then “following rules” second. Instead, compliance to specs/rules should come first (similar to the way humans behave!). AIs should resort to human values when out-of-distribution. A second idea is about getting safety from a balance in power (Player A & B both have ASIs => protect/cancel/dampen each other). My opinion is that Deterance theory collapses the more fatal, accessible, and hidden a weapon is.

Incentives as selection effects

A fun thought experiment that explains the role incentives play in life. Instead of thinking of them as forcing functions to change behavior, consider them filtering mechanisms (we can’t steer behavior, we only select for people apt to it). This seems to be partly true (i.e., a contributor) as there is a spectrum on how much a person wants something, we can bring more out of a person compared to another depending on how much we incentivize.

Culture Matters

While culture is mostly in the background, it can work wonders in aligning & growing team efforts. Culture influences what people believe important (mission) and possible (growth). We often absorb our sense of what’s achievable from our cultural environment (other peers and mentors), which becomes a major driving factor in our performance.

Given how much we use the terminal, any friction to get an action done is multiplied hundred of times resulting in a lot of wasted time. Optimizing the way we use the terminal will give us a decent productivity boost. In this post I share with you some common pain points that I faced before and found solutions for.

1. My terminal theme and colors are confusing!

Solution: Oh My Zsh.

1. sh -c "$(curl -fsSL https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh)"

2. Pick a theme in ~/.zshrc (for oh-my-zsh, set ZSH_THEME="agnoster" or similar).

3. Open a new terminal to see your fancy prompt.

.. or pick a dedicated color scheme like Solarized, Gruvbox, or Dracula.

2. I keep hunting for my terminal window!

Solution: Find a shortcut to hover a terminal whenever you need it.

For MacOS:

• Install iTerm2 if you haven’t already.

• Enable hotkey window:

  1. In iTerm2 > Preferences > Keys, set up a “Hotkey”.

  2. Turn on the “Hotkey Window” option.

  3. Now press your hotkey, and a terminal drops down instantly—press it again to hide.

For Linux, use a drop-down terminal like Guake, Yakuake, or Tilda:

1. Install sudo apt-get install guake (Ubuntu) or sudo dnf install guake (Fedora).

2. Launch Guake, then press F12 (default) to summon or hide it.

3. Customize the hotkey in Guake’s preferences if you want a different shortcut.

3. It’s slow to navigate directories!

Solution: z

Use z (autojump-like) or zoxide:

1. Install z: brew install z (Mac) or sudo apt-get install z.

   • For zoxide, visit: zoxide.org.

2. Source it in your shell config (.zshrc, .bashrc), eval "$(zoxide init bash)".

3. After a day or two of normal usage, type z <partial-dir-name> to jump instantly.

4. I find myself writing the same commands lots of times!

Solution: zsh-autosuggestions or quickly search the history.

For zsh-autosuggestions:

1. brew install zsh-autosuggestions (Mac) or use your distro’s package manager.

2. Add plugins=(zsh-autosuggestions) in your ~/.zshrc (if using oh-my-zsh).

3. Watch as your previous commands appear in lighter text—press right arrow to accept.

5. I Don’t Know the Right Command!

Solution: use a LLM CLI tool to ask it for commands.

1. Install llm via pip.

2. Either install a free LLM locally or provide API keys (llm keys set openai)

3. Check the available LLMs you have via llm models

4. Add this to your ~/.bashrc: alias please="llm -m gpt-4o 'Give me a short macOS terminal command to'"

5. Then just run please "How do I find all .txt files and remove them?"

6. I’ve got five tasks at once, but each new shell is a hassle

Solution: persist sessions with tmux and learn shortcuts for pane/window management.

For MacOS:

1. Use iTerm2 with split panes or multiple tabs:

   1. Press ⌘ + T for new tab, or ⌘ + D / ⌘ + Shift + D to split panes.

   2. Keep tasks running in separate tabs/panes.

2. Try tmux if you like multiplexing within one terminal:

   1. brew install tmux

   2. Run tmux, open new windows/panes with Ctrl+B c (new window) or Ctrl+B % (vertical split).

7. The cursor is slow!

Solution: adjust the cursor speed and learn shortcuts to jump around text quickly.

For MacOS:

1. Increase keyboard speed in System Settings > Keyboard. Slide “Key Repeat” to Fast, “Delay Until Repeat” to Short.

2. Set up Option + ← / → to move by word in Terminal:

   • In iTerm2, under Preferences > Profiles > Keys, map Option+Left to Esc+b (move back one word) and Option+Right to Esc+f (move forward one word).

For Linux:

1. Adjust keyboard repeat rate: xset r rate 200 50 (200 ms delay, 50 repeats/sec) in your .bashrc or .profile if you want it permanent.

2. Terminal shortcuts often already exist: Try Ctrl+Left or Alt+Left to jump word-by-word. If needed, add them in your Terminal Preferences or use your .inputrc.

Little tweaks can remove loads of friction and drastically boost your terminal productivity. We can (and should) find similar tricks around the software we use most: Browser, IDE, OS, etc.