Ali Nawab

ChatGPT, Claude, Gemini. Every model you've used runs on a single paper published in 2017. I read it. This is what it actually says, and what I built once I understood it.

The paper

Before 2017, AI models processed text like a reader with amnesia. One word at a time, each step erasing what came before.

Eight researchers at Google Brain threw that out. Their paper, "Attention Is All You Need," introduced one idea: instead of reading sequentially, every word simultaneously asks every other word how much it matters. The answer reshapes the meaning.

Take: The bank by the river flooded after three days of heavy rain.

"Bank" has two meanings. The model looks at "river," "flooded," "rain." Those words vote. The financial meaning collapses. Riverbank survives.

Every word. Same time. Milliseconds. What comes out isn't your sentence. It's a compressed version: each word encoded by context.

Compressence

There's a concept for this: compressence. Compress an idea until only the essence remains. Any further and you lose the meaning.

Think of it as a bell curve. "Bank" in isolation spreads across a distribution: financial institution at the center, highest probability, riverbank further out. Add "river" and "flood" and the curve shifts. The financial meaning falls to the tail. Riverbank becomes the peak. Attention moves the curve. Compressence is the peak.

Applied to the 2017 paper, the smallest version that keeps the meaning:

Meaning lives in relationship, not in isolation.

Remove "simultaneously" and you're back to sequential models. Remove "relationships" and you have no attention. Remove "at scale" and you have a toy.

What nobody tells you

The model has no memory.

Every new chat starts from zero. What feels like continuity is the conversation history pasted back in. The model reads it fresh each time.

It doesn't think either. It predicts the most plausible next word given everything in its context window. Confidence is a statistical property of training data, not a signal of truth.

Precision: how often it's right when it answers. Recall: how many right answers it captures. A model can fail at both while sounding completely certain.

No memory. No reasoning. No plan. Pattern matching at scale.

Most users cluster at the center of a bell curve: same prompts, same outputs. The 100x engineers are in the tail. Not because they have better models. Because they understand one thing: the model samples; it doesn't think. What it samples from depends entirely on what you give it.

What you can build from that

If the model has no memory, engineer one. If it only knows what's in the context window, design that window deliberately. If it's pattern matching on text, your text has to be precise.

I keep years of notes in Obsidian: journals, book highlights, article clips, half-finished ideas. I connected Claude directly to that folder and wrote CLAUDE.md: a file that tells it how to operate on my notes. My tracks, my rules, my structure. It never touches raw notes. It only writes to the wiki layer it maintains.

I built 22 commands. Each one is a markdown file, a reusable procedure that teaches the model how to do something specific. In my old setup, this would be a Python script. Now it's a markdown file. Fat with context. The system running them is deliberately thin: reads files, manages context, calls the model, returns output.

/today reads my notes, calendar, and open threads. Writes a daily plan based on what's actually happening, not a template someone else designed.

/context builds a snapshot of who I am right now: role, projects, open decisions. Every new session starts with Claude reading it. The model never asks me to catch it up.

/emerge surfaces ideas my notes imply but I've never written down. First run found a pattern across 18 months I hadn't consciously noticed.

Intelligence lives in the commands. The harness stays thin.

Compressence test on what I built: The model only knows what's in the window. Fill the window with everything that matters.

Five years of notes. 22 commands. One file that tells the model who I am. The architecture told me how to build it.

Vague prompts produce vague answers. Not because the model is dumb. Because you gave it nothing to compound.

The new job

Every enterprise needs someone to own this. Not centrally. At the team level.

The job: find workflows where compute changes the math completely. Not 10% faster. 100x more volume. Every inbound lead processed, not just the top 10%. Every contract reviewed, not just the flagged ones. Every customer onboarded at the same quality, not just whoever gets the senior rep.

The hard part isn't the agents. It's the context. What does the agent need to know, in what form, at what step? Where does the human stay in the loop?

Two categories: latent and deterministic. Judgment, synthesis, pattern recognition: these belong in the model's context space. Numbers, facts, SQL, contract clauses: these must stay deterministic, outside the model's reach. The job is knowing which is which and building the boundary. Blur it and you get a system nobody trusts.

The skill is the same whether you're building for yourself or a team of 500: the model reads, not thinks. Design around the constraint.

It's already a job. Most companies just haven't written the description yet.

What to do with this

The prompt isn't a question. It's the only document the model will ever read about you. Write it like one.

Most people type a question and wait. They're using a search engine that writes sentences instead of returning links. The ones getting real leverage treat the prompt as a brief: full context, clear constraints. Give it nothing and it invents.

Stop treating this like a conversation. Start treating it like a spec.

Notes

1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762
2. The benchmark was WMT 2014 English-to-German translation. The model scored 28.4 BLEU, beating the previous state of the art by more than 2 points at a fraction of the training cost.
3. Several of the eight authors went on to found the labs now competing with Google. Vaswani co-founded Adept (acquired by Amazon in 2024). Gomez co-founded Cohere. Uszkoreit co-founded Inceptive. The paper that gave Google's competitors their foundation was written by Google employees on Google time.
4. Precision and recall are borrowed from information retrieval, where they measure search quality. The framing is deliberate: the model has the same failure modes as a search index, just less visible.
5. The term comes from Chris Begg, via a profile at whatgotyouthere.com. Begg uses it to describe compression to essence: remove any further and meaning breaks.
6. The pattern: more time spent building infrastructure than writing. Eighteen months of daily notes made it visible.

I have been taking notes on everything since the early 2000s. OneNote. Windows CE with a stylus. Notability with Apple Pencil on the first iPad. Roam Research when it launched. My tools changed. The habit didn't.

I now write notes on clean A4 paper in meetings, with a Blackwing Pearl Graphite if I can find one. Everything digital lives in Obsidian. A daily note, a weekly recap, essay drafts, a 7-day plan.

I came to Obsidian from Roam and didn't like it at first. Too polished. I also kept hoping Roam would become way better. Obsidian won me over slowly: plain text files, I own the data, and I can extend it when something bothers me. Small Canadian team with strong opinions about what makes a great product.

My laptop broke last week. I went back to taking notes by hand, writing "/tomorrow" next to tasks and picking them up the next day. Roam had a plugin that did this automatically. Obsidian didn't. Getting a task into tomorrow's note required clicking around, a modal popup, retyping. Three steps for something that should be zero.

I built Chrono Blaster to remove those steps. Type do something /tomorrow, select it from the slash menu, done.

The task lands in tomorrow's note as a checkbox. The current line keeps a wikilink so you know where it went.

I built this during Ramadan. For some reason, fasting brings me back to building things. Fasting until sundown gives me about two hours a day. When time is that scarce, you only build what actually irritates you. No research sprints, no roadmaps. Just friction, then fix.

This is the first Obsidian plugin I've built. Audience of one. I'll keep adding to it and submit it to the plugin directory once it's stable.

It's at github.com/alinawab/chrono-blaster if you use Obsidian and want to try it. 

Four months after selling Agentnoon for a life changing outcome, I still haven't written about it.

I know what the piece is. The real version of why founders sell. Not the LinkedIn version. 80% financial, 10% exhaustion, 10% strategic at signing. After close it moved 40-30-30. That shift is the whole story. Nobody writes about the shift.

Instead I've been decompressing.

But I recently had my mind blown with Anthropic and Obsidian.

I had five years of notes in Roam Research, Apple Notes, voice memos, email drafts. I moved everything into Obsidian. Vin Verma had shown his setup on Greg Isenberg's podcast. I built on top of it, connecting Claude Code to the vault with slash commands that let me query my notes the way I think. He's been one of the most interesting people I've started following on X in the last year.

It took a while to start. Giving an AI access to five years of private notes is not a small decision. I haven't fully resolved that question. I built it anyway. Some of us have to take the risks early so everyone else doesn't have to.

Over 48 hours in Ramadan, between Sehri and Iftar, I built 21 of them.

/emerge reads everything in the vault and surfaces ideas I've never explicitly stated. Conclusions the notes imply but that I haven't said yet.

I ran /emerge on my own writing plans. $50 in API costs and a lot of reasoning and back-and-forth later, I had answers I wasn't ready for.

The writing infrastructure is complete. The writing itself doesn't exist.

It also found something I hadn't noticed. I imported all my old blog posts. 40-plus pieces from 2018 to 2020. The longest is maybe 200 words. Most are 50. I was posting five times in two weeks in January 2018. The writing plan I've built for 2026 calls for one essay every two weeks, each drawing from an intellectual library that spans Ibn Khaldun to Nassim Taleb.

Not the same writer.

The old posts were fast and unguarded. The 2026 plan is careful, positioned, aware of its audience. I don't know which mode is right. Probably some of both.

The exit essay is about a decision that's already been made. This one is about something happening right now: building a system for thinking in public, one command at a time, in a month where I'm fasting until sundown and have limited extra time. 

The constraint is the advantage. You find out what's actually worth doing when you only have two hours.

Started writing this in December. It sat until now.

Four months since we exited. I'm a "serial entrepreneur" now. Feels strange but it's facts.

For the last four years I was focused on winning the month, then sending out an investor update: revenue, wins, losses, learnings, where we needed help. That chapter is closed. This is what comes next.

Highlights

Four months in, I'm happier than I was when we made the decision to exit.

Spending more time with kids, family, friends, dentist. Several people have said I seem more relaxed, even happy. I've been told I'm hard to read, so that's more significant than it sounds.

Lowlights

I was so focused that people were sometimes afraid to talk to me.

I'm realizing how much of my life was on hold while I was busy building. Makes me wonder if I'll ever start another company again.

Big learning

When we sold Kiwi in 2017, I decided not to make any splurge purchases for 12 months. Best decision I made that year. Doing it again.