ajinkya.ai An experiment in learning with AI.

Everything

Gratitude: the under-trained, under-valued skill

We train compassion, empathy, kindness, equanimity — but almost never gratitude. Its opposite is entitlement, and I've noticed the people with the least gratitude are usually the unhappiest, and the hardest to be around. Gratitude isn't a mood. It's a skill, and almost no one trains it.

A gift to my 15-year-old self from Claude Fable 5!

At fifteen I built my own Visual Basic interpreter, going library to library with no internet to learn how it all worked. In just a few hours, Claude Fable 5 built me a very functioning Visual Basic clone in the browser — full debugger, autocomplete, syntax coloring, and a BASIC interpreter written in BASIC. My younger self would have loved it.

It May Not Be a Big Deal to You, but It Is to Your Subconscious

Even a small stressful thought — or a tiny exciting one — can quietly disrupt breathing, digestion, and sleep. On the surface it's nothing; to the subconscious it's a big deal. Watching how things register there matters more than we think, before seemingly harmless thoughts steal our sleep, health, and peace.

NASA's Earthdata, in Plain English

NASA's Earthdata repository holds 178+ petabytes, but its dataset names are uninviting and intimidating. As an experiment, I translated 3,000+ of them into plain English and built a curated catalog of the top 300 — each with an AI-generated cover image — to make Earth data more approachable.

Middle Managers Are Like Air Purifiers

You don't really need middle managers — fire them and everything runs smoothly, until it doesn't, because the momentum they set is still carrying you. They're like pure air: nobody appreciates it until you and your family start to develop lung issues.

Those who forget history are doomed to repeat it

On learning from history through interactive play — a browser game inspired by the ideals of Chhatrapati Shivaji Maharaj, where you earn the countryside by governing well rather than out-producing the enemy. What I built, the art and the design, and where to play it.

The Mind That Is Never Here

The restless mind is never satisfied where it is — with people one conflict, without them another; here it wants to be there, there it wants to be here, back and forth without end. When will this madness of the mind stop, that won't let me enjoy a single moment?

Everyone Sees From Their Own Angle

We like to think our words land exactly as we mean them; they rarely do. In a single day I watched a coworker, my family, and someone I'd complimented each reshape what I said through their own angle — a reminder that being understood isn't automatic, and that a misreading is usually just someone seeing from where they stand.

5 Reminders Before You Speak

Five reminders I keep for myself before speaking — check what your mind is doing first, remember that a truth needs the right time and place, try to see the other person's limits, watch the ego when you have the floor, and speak without violence in word, tone, or glance.

Mechanistic interpretability — reverse-engineering a transformer

An interactive, full-screen field guide to mechanistic interpretability — reading a transformer not as a stack of layers but as a residual stream that named circuits read from and write to. Framed the way you'd reverse-engineer a piece of malware.

Project Tukaram

Why I spent three months of weekends using AI to translate all 4,500+ abhangs of Sant Tukaram — the householder-saint who lost everything to famine and still became Maharashtra's most beloved voice — and what I hope the first complete, theme-mapped reading can offer the millions who love him.

The Mind Comes to Love What You Repeat

Five small lessons from practice — that equanimity is a muscle you can train, that awareness comes first, that the mind grows to love whatever you keep returning to (so discipline becomes desire), that life is short enough to demand guarding your focus for the essential few, and that momentum, once built, carries you like a river's current.

Fine-tuning a language model — a practical field guide

One complete, practical guide to fine-tuning a language model for your own task — when to fine-tune versus prompt or retrieve, why you adapt a small slice (LoRA) rather than retrain, the five steps end to end, the settings that matter, and a full production case study where a fine-tuned model replaced a general LLM agent on a narrow job.

She has her own ways of telling me

A toddler hasn't yet learned the adult, accepted ways of showing emotion, so it comes out in her own forms — different each time, and easy to misread as a tantrum. The work is to notice, gently and without judgment, what she is actually telling us — and the same patience is owed to the grown-ups around us, who never fully traded their own ways either.

Momentum is the key

In meditation, momentum is everything — not gripping the calm, but being willing to return to it again and again. And a different attitude to the mind: investigating why sadness or craving arises at all, until the mind's play loosens its grip.

She is not giving you a hard time — she is having one

When a child melts down, it feels like they're giving you a hard time — but really they're having one, caught in a storm their nervous system can't yet hold. Seeing that lets you lower your voice and simply stay. And the same is true of every difficult adult.

Use AI to build pyramids, not houses

With AI, supposedly one prompt makes a polished UI. The way I actually do it is 400 prompts, late nights after my daughter sleeps, two-word stretches between her interruptions. When someone says it was *just AI*, they miss what is actually happening. Writing code is cheap now — the cost of trying an idea collapsed, and work I never would have shipped comes back in reach. The bottleneck moves to taste — also the part AI cannot do, because the genuinely new isn't in its data yet.

The feeling of unworthiness

A phrase from a Vipassana book I found at a used bookstore — "the feeling of unworthiness" — and how much of what people do, the pleasing and the masks and the drama, runs on it.

The one thing I most want to give you

The most important thing I could pass on isn't technology or knowledge or even morality — it's the ability to manage your own mind: to stay quiet in chaos, to guard against jealousy and anxiety becoming habits, and the honest catch that I can't give it without having it first.

Do ancient lists follow a map?

The Rigveda's river hymn looks like it lists rivers in geographic order. Scholars said so for 150 years but never measured it. I did — and ran the same test on four more old lists. The Indian ones follow the map; the Zoroastrian Avesta and Genesis don't, or not simply. What a list is for seems to matter more than its language or its century.

The Chart Is a Recipe, Not a Forecast

A short take on the Vedic birth chart: read it backward, not as a forecast but as a record of the tendencies a person arrived with — the specification of a being fused from several streams of consciousness. The full argument is in the linked paper.

Re-deriving things again and again

A strong memory for your own conclusions can quietly block fresh thinking — the mind says "I already know this" before it has looked again. Weak retention and the ability to re-derive may be two sides of one coin, and a way, too, of re-reading a difficult Mercury.

How a cost question became a calculator, a paper, and a book

Kumar asked me what an AI agent costs to run. I didn't know. This is the story of how that question turned into a calculator, a paper, and a book — because the answer needed to work for the next team to ask, not just his.

Explaining the LLM cost paper

A section-by-section walk-through of the methodology in "Cost Modeling for Public-Facing LLM Chat Applications" — the naive per-token formula and why it lies, the structural corrections that break it (caching, traffic shape, segments, daily caps), the six equations that replace it, and the worked geospatial-Q&A example that ties them together.

Understanding LLMs — a field guide

Twenty-three interactive chapters that walk language models end-to-end — from what tokens are and how attention works, through retrieval and context engineering, to the production techniques (caching, tools, evals), inference and serving, and the choice of model itself, closing with foundation models beyond language. Read in order or jump to the chapter that matches what you are building.

How to cost an AI agent: a progressive tutorial

A single-page, progressive tutorial for engineers and analysts who want to build their own AI-agent cost calculator. Starts at "what is a token?" and ends at a reference architecture you can clone — data model, formula library, and a validation harness. Implementation-agnostic; one worked reference lives at calc.ajinkya.ai.

No entries match.