My ex-CTO used to say something I never forget.

Most of what we do day to day ends up being inconsequential. The real impact is in a handful of choices, but it’s not always clear which ones they are, but it's rarely obvious which ones they are until later.

That question — what makes a good decision — is what got me seriously into mental models. Charlie Munger formalised the approach: build a latticework of frameworks from different disciplines and reach for the right one when the moment calls. Not to sound smart. To actually use them. This is the first time I'm writing about them. I want to start with one of my favourites.

additive and multiplicative systems

Most people think about life additively. Do good things, accumulate good outcomes. One bad decision hurts, but you recover. Keep moving, net positive over time. That’s not always how it works.

There are two kinds of systems: additive and multiplicative. And knowing which one you’re operating in changes everything.

In an additive system, bad things hurt but they don't destroy you. Outcomes stack. Failures subtract. But the overall result survives individual mistakes. You're still in the game.

Multiplicative systems are different. Here, the components don’t add — they multiply. Which means one zero doesn’t just cost you. It wipes out everything that came before it. Doesn't matter how large the other numbers are.

Len Bias was a basketball prodigy. College Player of the Year. Second overall pick in the 1986 NBA Draft by the Boston Celtics. People who watched him play say he was the best basketball player who never played professionally. God-given talent, years of work, everything pointing up.

Two days after the draft, he died of a cocaine overdose.

All those multipliers: the talent, the work, and the draft position hit zero. Not set back. Gone.

That’s a multiplicative system. And it’s about as clean an example as you’ll find.

it’s rarely a full zero though

Len Bias is the extreme end. A true zero. Most of us won't experience that (hopefully). More often it's a 0.4 or a 0.6. A massive snake on the snakes and ladders board that doesn't end the game but sends you back years. The damage is real. The compounding is lost. But you're still in it.

I think about my own life this way. A decade of neglecting physical health: the brain fog, the lost energy, the productivity I’ll never fully account for. Years I didn’t use as well as I could have, and the ground I had to make up afterwards. Financial habits in my twenties that cost me compounding I can’t get back. None of these destroyed me. But each one set a clock back. And when I’m honest about it, I can feel the weight of what those partial zeros cost.

The model doesn’t just show up in dramatic, life-ending moments. It lives in the slow, quiet decisions too. The ones you don’t even clock as decisions at the time.

the question worth asking

What this model actually gives you is a two-step move before any significant decision.

First: is this domain additive or multiplicative?

Additive domains are skills, creative work, and most career bets. You can experiment, fail, lose ground, come back. A bad call stings but it doesn’t restructure everything downstream. Take the swing. Err on the side of more bets, not fewer.

Multiplicative domains are health, reputation, relationships, legal, and finances (kinda grey). In these, the default job isn’t to maximise upside. It’s to protect against zeros. One bad move doesn’t just cost you time, it can rewrite the whole equation.

Second: invert.

Munger’s most-used mental model wasn’t forward planning. It was asking:

What does a zero look like in this domain, and can I rule it out? Once you’ve protected against the possible zero, you can think about upside. But not before.

Most people optimise for gains when they should first be asking what game they’re playing. This model forces that shift and the inversion makes it actionable. You’re not just categorising your decisions. You’re stress-testing the right thing in each domain.

one caveat

This model is a lens, not a rulebook.

If you take it too literally you start running your life like an optimisation problem. Every decision gets stress-tested for its risk to your score. And that’s its own kind of trap.

Some decisions that look like 0.4s on paper are the ones that keep you sane. Or present. Or true to yourself. And I genuinely believe that happiness and peace of mind aren’t soft variables you can quietly deprioritise, they feed back into everything else.

The model helps you avoid catastrophic mistakes. It doesn’t tell you how to live.

Collect it. Use it where it fits. Don’t let it use you.

see ya.

i started explaining it and halfway through i thought, wait, this is actually worth writing down.

so here it is. this is what i told him.

for a while, my process was: come up with a rough brief, dump it into AI, get an output, start editing. repeat.

and honestly? it was exhausting in a way i couldn’t explain at first. it felt like correcting a student’s notes. you’re not making anything, you’re just reacting to what someone else made. there’s a ceiling on how original that can ever be, because you’re bounded by what the AI inferred from your brief at that particular moment.

more than that, it just didn’t feel like mine. the sense of ownership wasn’t there. and without that, the work is just... fine. never more than that.

so i blew it up and started over.

my process now is split into two phases. the most important thing about them is that they don’t talk to each other.

phase one: creator mode.

no judgment. no constraints. no thinking about what’s realistic or what fits the brief or what the brand voice says. i’m just exploring, pulling from whatever i’ve been reading, watching, listening to. letting connections form. jotting things down without filtering them.

the whole point of this phase is to create a space where the inner editor isn’t allowed in yet. because the moment you start judging an idea, you stop generating them. and the ideas you kill early are sometimes the ones that would’ve been the most interesting.

i also deliberately limit AI here. the moment i open that tab, i’m working within someone else’s frame. i’m reacting instead of discovering. it creates a ceiling before i’ve even started climbing.

and honestly? this phase just makes me happy. there’s something about having a space where you can just go crazy, think of things, explore whatever you want, no one grading it, no output expected yet. i didn’t have that before.

phase two: editor mode.

this is where i come back to earth. what’s realistic? what fits the voice? what’s actually good versus what just felt exciting at 2am? this is the crafting and execution phase, ruthless, practical, structured.

and this is where AI can go completely crazy. everything i’ve already decided about direction and substance, AI helps me execute it faster, cleaner, at scale. that’s the right use of it. not as a creative partner in the beginning, but as an execution engine at the end.

the more you can get AI to handle in the execution phase, the more time and energy you free up for the generative one. that feels like the right expedition to be on.

i didn’t arrive at this on my own.

i’ve been into hip hop my whole life. big kanye fan. so i’d known who rick rubin was for years, produced albums for everyone from jay-z to johnny cash to kendrick. famously doesn’t play an instrument. just listens.

i’d seen clips here and there but never really went deep. then one youtube rabbit hole later, i ended up reading the creative act. for someone who genuinely struggles to finish books, i finished it in less than a week. audible gave me a badge for it.

what rubin talks about, this idea of art as a process of honest reception, of removing yourself as a filter and letting the work emerge, it rewired something for me. it’s not just about music. it’s about how you show up to the act of making anything.

i started thinking about creativity differently after that. less as something you produce and more as something you tune into.

the process i have now isn’t final. it’s just current.

creativity is a river. you keep learning things, get inspired by different people, stumble onto new ways of seeing. the process i follow today is not the same as the one i followed two years ago. and i’m genuinely excited to see what i pick up this year that shifts it again.

what i know for now: protect the generative phase. keep the editor out until it’s time. use AI where it actually belongs, in execution, not in exploration.

and make sure whatever you’re making still feels like yours.

I’m an engineer turned marketer. Haven’t coded in years. But the recent AI updates—Opus 4.6, GPT 5.3 Codex, all the hype made me restless enough to dive in.

Turns out, building stuff with AI is actually doable if you’re non-tech. But the landscape is confusing as hell, and there’s a real learning curve.

If you’re lazy like me, someone who hates configuring tools more than learning new tech, this is for you.

Here’s what I figured out.

too many options

If you’ve been paying attention to AI coding tools, you know the paradox: there are so many options that it’s hard to know where to start.

Claude Code, ChatGPT Codex, Computer Use agents (Clawdbot), Lovable, Antigravity, Cursor, the list keeps growing. Each one promises to make coding accessible. None of them tell you which one you actually need.

I spent a good chunk of time just learning what each tool does and where it fits. Here’s the landscape:

Claude Code: Fast processing, great for first iteration and developing your prompt as you can interact with the application side by side as you iterate, excellent for finalizing your spec

ChatGPT Codex: The actual coding workhorse. The macOS app makes it accessible for non-tech people, thinks closest to how a developer would, thinks deeper and longer, lets you run multiple instances on a project—it’s like spawning multiple devs for different features

Lovable: Rapid website builder, $25/month, good if you need to ship something fast and don’t mind paying

Agents (Clawdbot): Full system control, could be a security risk unless you’re running them on a dedicated machine. Unless you’re rich enough to buy a Mac Mini just for your AI agent, maybe avoid these. It also needs some terminal usage, not necessarily non-tech friendly.

Codex skills and Claude skills: Does what clawd agents do but easier, doesn’t need full system access. Better option for most people.

After analyzing all of them, my conclusion: Claude Code and Codex are both excellent. You can’t go wrong between them.

Claude Code processes faster and renders interactive applications easily. Great for spec work.

Codex builds like a developer thinks. Multiple parallel threads, macOS app for non-tech people, better for actual development.

Cost consideration: Codex requires ChatGPT Plus ($20/month). Claude Code is part of Claude Pro (also ~$20/month). Factor this into your decision.

I chose Codex for building. I still use Claude for copywriting and creative work—it feels more human.

what i built: a fitness tracker

I’m turning 30 this year. Read that your fitness at 30 sets you up for your 40s and beyond. I’d lost mine grinding at startups for four years.

I had goals and KPIs defined, but tracking them meant configuring Notion or a spreadsheet. I’m lazy. Building something seemed easier than setting up another tool.

Perfect use case for AI.

the process: spec first, build next

Here’s what actually worked:

step 1: build the spec with ChatGPT

I spent time with ChatGPT defining exactly what I wanted: how data should be structured, what happens with edge cases, the color scheme, even how calculations should work.

I wasn’t learning to code. I was learning to write a product spec, the kind you’d give a developer on your team. That’s what Codex essentially is.

This is critical. A solid spec saves you dozens of iterations later.

The spec process isn’t coding, it’s product management. You’re defining what the product should do, how it should behave, what edge cases to handle. It flexed design thinking, product thinking, and critical thinking muscles.

Honestly the most enjoyable part. Creativity is the difference maker here.

step 2: hand it to codex

Once I had a spec that felt good enough for MVP, I gave it to Codex on high thinking mode. It took 30 minutes to create the app. Here’s the exact prompt I used.

From idea to working MVP in under an hour. That’s such an unlock for builders.

the reality: iteration required

Here’s what the first version looked like:

Functional. Basic. It worked, but it wasn’t refined.

After using it for a bit, I iterated. Here’s the current version:

Better data organization, cleaner dashboard, more useful “at a glance” view.

This is the reality: AI gives you an MVP. It’s not a replacement for product development and testing.

what’s hard (still)

After building and iterating, here’s what I learned AI tools are still not great at:

design and aesthetics

Colors, visual hierarchy, taste—these don’t come out of the box. Prompt engineering gets you maybe 30-40% there. The rest requires iteration and your own sense of what looks good.

ux thinking

The tool won’t intuitively know how users should flow through your app. You have to test, identify friction, and refine. This is still a human skill.

deployment

Getting your app on the web is harder than it should be for non-tech people. Codex has Vercel automation, but it’s limited unless you have a Vercel account (which no marketer has). This gap needs better solutions. I made mine fully local, that boots up and shuts down through simple commands.

iteration loop

AI builds fast. But you still need to use the thing, find what’s wrong, articulate improvements, and run it back through. The tool doesn’t replace judgment or taste.

cost

Basic AI usage isn’t expensive, but once you start pushing it hard, costs add up. I experimented with building my own virtual team using agents in heavy coding mode for an afternoon. Great results, but they burned through credits fast. That afternoon alone cost me $13. Scale that across multiple projects and you’re looking at real money.

what’s excellent (surprisingly!)

What AI tools absolutely nail:

Project architecture: How your app should be structured—nailed it.

Data architecture: How information flows and gets stored—nailed it.

System logic: The actual functionality and development—nailed it.

These are the hard technical parts. The fact that AI handles them reliably means non-tech people can actually build functional tools.

The design and UX refinement? That’s still on you. But that’s the creative part anyway.

what can you actually build with this?

People say “your imagination is the limit” about AI tools. Usually that’s hype. But with these coding and agent tools, it’s closer to true than it’s ever been.

The unlock isn’t just building apps for yourself—it’s spawning capability you couldn’t afford to hire.

virtual teams, not just tools

Crypto moves fast. There’s endless data but terrible signal-to-noise ratio. Blockchain transparency means even more to analyze. You can build agents that help you keep up:

Market analyst agent: Monitors industry data and your product analytics. Feeds you insights for strategy without drowning you in dashboards.

SEO/GEO agent: Integrates with Ahrefs or SEMrush, generates content based on keyword opportunities, runs technical SEO audits on your site. Handles the grunt work while you focus on strategy.

Content creative agent: Connects to your X analytics (via API or regular Excel dumps). Becomes your brainstorming partner—helps identify what’s working, suggests new themes, tests content pillars. Constant experimentation without the overhead.

These aren’t hypothetical. People are building these now.

the efficiency shift

A prominent VC I spoke with recently shared something striking: they used to exit investments when team size hit 14 people (11 devs, 3 marketers). Now they’ve revised that down to 6 people total—3 devs, 3 marketers.

That’s not about doing less. It’s about AI absorbing the busy work that used to require bodies.

Marketing has endless grunt work: data analysis, content production, technical SEO, performance tracking, competitive research. Every function does. If you can leverage AI well, you automate that layer and keep the creative, strategic work for humans.

my workflow now

After experimenting, here’s what I’ve settled on:

1. Spec with ChatGPT - Define exactly what I want, iterate on the requirements

2. Build with Codex - Let it handle the technical implementation

3. Write/create with Claude - For copy, content, anything requiring human voice

Different tools for different jobs. None of them do everything.

I’m just scratching the surface of what’s possible. I’ll keep building apps, iterating on the ones I’ve made, improving them. It’s a great way to develop product thinking. I started work on building my own virtual marketing team with openclaw, and will report results.

My take: this skill will be critical in 2026. Not because everyone needs to code, but because the ability to turn ideas into functional tools—fast—is what separates the best operators from good ones.

The tools are accessible now. The learning curve exists, but it’s not insurmountable. You just have to start.

resources that helped

If you want to learn this yourself, here’s what helped me:

Great walkthrough on the Codex macOS app:

Good intro to the Codex app:

@tushaarmehtaa

on X - His content motivated me to explore this deeply, they also do coding hackathons on the weekends in BLR.

Lay of the land on AI models:

Best no-BS explainer of clawd agents:

People talk about “10,000 hours” like it’s a motivational milestone. Hit the number, unlock mastery. In reality, it’s less about the hours and more about what happens to your brain somewhere in the middle of them.

At some point, a problem stops being something you actively reason through. It becomes something you recognize. Decisions that used to take days compress into minutes. You stop evaluating every option from first principles because you’ve watched it play over so many times.

But here’s what nobody tells you: the hours alone don’t get you there.

What five years of marketing actually taught me

I’ve been in marketing for almost five years. I’d put in the hours: campaigns, messaging, strategy, writing. But I didn’t feel like I’d truly mastered anything until a few months ago.

For the first time in five years, I consciously decided to step back. Not because I was burned out, but because I wanted to actually think about what I’d been doing. At first, I just rested. Then I started reflecting on the work itself—what engaged me creatively, what felt mechanical, where the creative process actually lived.

Then I went back to reading. Marketing books, books on creativity, whatever pulled me in. The clarity that emerged was hard to describe. All those years of work suddenly organized themselves differently in my head. I could see patterns I’d missed while I was in the middle of it.

My output since then has felt fundamentally different. It was more creative, more confident, and more satisfying. Not because I learned new tactics, but because I finally understood what all those hours had been teaching me.

The 10,000 hours didn’t unlock mastery on their own. The reflection did.

The same pattern in a completely different domain

I’ve seen this same pattern play out with our team building Bitcoin infrastructure.

Over the past decade, they’ve spent well over 10,000 hours on a single problem: moving Bitcoin across different blockchain environments. They built and operated one of the first systems to do this at real scale, processing billions in volume.

After years in production, structural problems emerged—problems that seemed like reasonable tradeoffs early on became fundamental constraints they couldn’t build around.

So they rebuilt from scratch, designing around the failure modes they’d internalized from operating the first system.

The result? Major platforms now route their Bitcoin infrastructure through this new design. Not because of clever marketing, but because it solved problems those partners had experienced themselves. The operational experience translated directly into a solution that worked where others had struggled.

The breakthrough didn’t come from more hours of building. It came from stepping back with enough experience to see the patterns.

The 10,000 hour edge

Pattern recognition happens in the pauses, not just in the reps. You can grind for years and still miss what the work is teaching you. The insights compound when you create space to look back and see the patterns.

You stop asking “how” and start asking “where.” Early on, the question is “how do we do this better?” After enough cycles, it becomes “where does this reliably break?” That shift only happens when you’ve seen enough iterations to recognize the recurring weak points.

Decisions stop feeling effortful. You recognize what won’t scale. You see assumptions that will age poorly. You know which complexity is necessary and which is just accumulated cruft. The recognition comes early, almost automatically.

You can take new risks with more confidence. Not because you’ve eliminated uncertainty, but because you’ve internalized where things actually break. The edge isn’t invincibility—it’s earlier recognition of failure modes.

Surviving the middle

Most people leave during the middle part. The long stretch after novelty fades, when progress feels invisible, when the problem stops being interesting and starts being familiar.

The ones who stay are forced into a different kind of thinking. But staying alone isn’t enough. You also need to step back periodically and let the pattern recognition actually surface.

For me, that meant taking a real break after five years of marketing work. For the technical team, it meant watching a system operate for years, understanding its constraints, then having the clarity to rebuild rather than patch.

Progress shifts from inventing new approaches to eliminating the ones that introduce hidden fragility.

The people who solve hard problems aren’t the ones with the most novel ideas. They’re the ones who stayed with the problem long enough to internalize where it actually fails, and created the space to reflect on what all those hours taught them.

10,000 hours doesn’t guarantee you get there. But you definitely don’t get there without them.

In the immortal words of Kendrick Lamar,

Choosing a movie on Netflix is a gamble.

Picking a restaurant out of the gazillion options for date night with wifey is a gamble.

Wallowing through life just doomscrolling reels and shorts is also a gamble, one that doesn’t necessarily feel like one.

A classic checklist item for Indian middle class households for their kids’ summer vacation is chess classes.

Chess has somehow become synonymous with intelligence, and mastering the sport an indicator of long term success.

But, you see, the thing is life is not like chess, far far from it.

Let’s take a closer look.

A game of chess involves two players with each, a set of pieces.

Every time you make a move, you know exactly what you are dealing with, the pieces, and the position. If you feed this data to an AI model, it will spit out the perfect objective move you can make in that position, all you need is memory and computation time. There is no hidden information or luck at play.

You know what has a lot of that, LIFE

Let’s take an example,

Say you’re thinking of switching careers, you don’t know:

• how the people will turn out

• whether the opportunity is real or just well-marketed hype

• how long your motivation will last

• whether this move compounds or quietly goes nowhere

For most of the decisions you take in life, you are dealing with hidden information, and a whole lot of uncertainty.

That’s not chess.

That’s poker.

In poker,

• You don’t see your opponent’s cards.

• You don’t know what’s coming on the river.

• You’re constantly guessing, based on patterns, probabilities, and gut whether this hand is worth playing.

And the key thing poker teaches you (that life absolutely runs on):

Outcome ≠ decision quality.

Most people get this wrong in life.

They take one bad outcome as proof they were stupid.

Or one lucky break as proof they’re a genius.

Both are dangerous.

Anyways, thats the thought

Ps.

fuck chess

Subscribe now