The Unread Pile

Edition No. 9Saturday, 3 October 2026Compiled from @RAKSHITLODHA’s likes & bookmarks

Lead Story · Business

The $43 Billion Lesson in Holding: Eduardo Saverin’s Revenge

By @aakashgupta, quoting @Kalshi_Film · posted Oct 2 · liked

The most brutal scene in The Social Network — sixteen years old this year — is Eduardo Saverin watching his 34% of Facebook get diluted down to 0.03% while everyone else’s holdings are protected. Real life wrote a different ending. Aakash Gupta’s retelling lands the punchline first: Saverin is now worth $43 billion, the richest person in Singapore and the richest Brazilian alive. The most famous victim in startup history won by refusing to sell.

Editorial illustration of a share certificate being diluted

From 34% to 0.03% — and then, slowly, back. The dilution scene was real; the ending wasn’t in the movie.

The mechanics of the scene, Gupta notes, are mostly accurate: Facebook reincorporated in Delaware, the new company issued 24 million fresh shares, and Eduardo’s stake collapsed. The detail the film glosses over is the one that decided the real case — Facebook’s own law firm prepared the papers he signed. He never brought his own lawyer, and that conflict of interest became the heart of his lawsuit.

Chart: the dilution and the recovery

Source: @aakashgupta’s post; settlement stake widely reported, never officially confirmed.

Then came the part the movie couldn’t show, because it hadn’t happened yet. Saverin sued, and the sealed settlement reportedly restored a stake of around 5% plus his co-founder title. Then he did the thing almost nobody manages to do: he held. Through the 2012 IPO. Through Cambridge Analytica. Through every Meta crash. He even renounced his US citizenship right before the IPO and moved to Singapore — saving an estimated several hundred million in taxes on the way.

“The most famous victim in startup history won by refusing to sell.”— @aakashgupta

Timeline: twenty-two years of holding

Source: @aakashgupta’s post.

Everyone else in the saga settled for less. The Winklevoss twins took $65 million. Sean Parker was out of the company within a year of the events on screen. Eduardo took the stock and waited — and the waiting did the compounding. It’s a story about the difference between being right and being rich: Saverin was arguably wronged and lucky and disciplined, and only the discipline was entirely his.

34%

Saverin’s founding stake, 2004

0.03%

After the Delaware dilution

$43B

His net worth today (Forbes)

$65M

What the Winklevoss twins took

Main learnings

The dilution was legal mechanics plus an unrepresented signer — always bring your own lawyer is not folk wisdom, it’s the whole case.
The settlement restored him to ~5%; the fortune came from holding for 15+ years through every crisis, not from the lawsuit itself.
Optionality compounds: the citizenship move before the IPO was a second, quieter fortune made of tax structure.
Being diluted, suing, and winning the lawsuit were events; refusing to sell was the strategy.

Practical takeaway — In any cap-table fight, the paperwork matters once and the holding matters forever. And in your own startup dealings: never sign the other side’s lawyer’s papers.

Read the original post →

Technology · AI

Essay · Machine Intelligence

Stop Reading LLM Output: The Ladder of Better Formats

By @karpathy · posted Oct 2 · liked & bookmarked

Karpathy’s thesis in one line: we’ll spend a lot more time trying to understand the outputs of language models. As LLMs do more of the legwork autonomously, human work rises up the abstraction stack into oversight and understanding — and luckily, the models can help with that part too. His post is a practical ladder of output formats, each better than the last for actually getting something.

From text to video: ask for the format your brain parses best, not the format the model defaults to.

Rung one: better writing. Ask the LLM to explain in ASD-STE100 — a controlled-language specification originally developed for aerospace maintenance documentation. Models are well-versed in it, and its heavy constraints produce a clean writing style he finds far more readable. The spec is stringent, so he sometimes softens it: ask for “80% of the way to ASD-STE100.” The insight is general — constraining the model’s style is a feature, not a limitation.

Rung two: diagrams. Instead of writing, ask for a diagram. They’re easier to process and parse than paragraphs for anything structural. Rung three: web pages. Ask for output “in HTML” and get a beautiful, interactive page — models have gotten genuinely good at frontend, animations included.

Diagram: Karpathy’s ladder of output formats

as code gets abundant, custom discardable artifacts (web apps, video explainers) become worth making for one person, one question, one afternoon. Hand-drawn for The Unread Pile · from @karpathy’s post

Source: @karpathy’s post.

Rung four — the one he’s most bullish on: explainer videos. Fully custom, bespoke videos generated on any arbitrary topic. His suggested experiment: “Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration” — and he reports this is actually starting to work. Free local-compute alternatives exist if you’d rather not hand over an API key.

“As intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts that would have never made sense to create before.”— @karpathy

The deeper point is economic: when intelligence and code are abundant, the cost of a custom artifact collapses, so the rational thing is to commission bespoke, disposable tools for one question, one afternoon, one person — a web app, a video explainer — things that “would have never made sense to create before.” Push the boundaries here, he says, and you’ll be surprised.

Main learnings

The bottleneck is shifting from producing answers to understanding them — format is the lever.
Constrained writing (ASD-STE100, or 80% of the way there) beats free-form prose for readability.
Diagrams > paragraphs for structure; HTML pages > both for anything interactive; bespoke video is the frontier.
Abundant intelligence makes one-off, discardable software artifacts rational — build for the question, not for reuse.

Practical takeaway — Next time you ask an LLM something hard, don’t accept the paragraph: ask for ASD-STE100, a diagram, an HTML page — or try the 3b1b-style explainer video experiment.

Read the original post →

Work · Careers

Playbook · Careers

The First 90 Days: Two Threads That Compound

By @spakhm · posted Aug 11 · liked

Slava Akhmechet’s guide to starting a new job — “surprisingly scales from intern to director, maybe further” — runs on two parallel threads. Thread A: find one valuable problem you can solve in about a day, solve it fast and well, talk to people about it, then find the next one 10% bigger. Thread B: talk to everybody, journalist-style, and really understand what they’re working on and what’s blocking them.

Illustration of two parallel threads at a new job

Solve small things fast, and interview everyone like a journalist. Do both at once.

Flowchart: the two-thread onboarding system

Source: @spakhm’s post.

Do this for three to six months and the compounding does the rest: you’re immediately useful (slightly, but that beats not at all), you build momentum, and — the key arithmetic — if each problem is 10% bigger than the last, in a few months you’re solving big, genuinely valuable problems. Along the way you build credibility, real knowledge of how things work, and a network of people whose problems you understand — which is exactly what lets you pick better problems next.

Chart: 10% bigger, every time

Source: @spakhm’s post; the curve is the point.

The failure modes are the mirror image, and each is a recognisable archetype: getting distracted by dozens of projects (shallow everywhere, trusted nowhere); reaching for bigger work too soon because the small stuff feels beneath you (nobody trusts you yet, and the impression solidifies); picking one big vague amorphous problem and making no progress in three months; contributing to conversations before you understand anything; going into pure “learning mode” and solving nothing; or pure “solving mode” without talking to anyone. His kicker, in a follow-up: “Also finish things repeatedly.”

“No one trusts you yet, and since you keep not solving anything, that impression solidifies. You need to earn people’s trust. Start small, finish it.”— @spakhm

Parallel threads: solve & learn

1 day

Size of your first problem

10%

Bigger each time — exponential

3–6

Months to big, valuable problems

Main learnings

Solve and learn in parallel — learning mode without shipping earns no trust; shipping without talking picks the wrong problems.
Trust is earned in small finished increments; 10% compounding turns tiny wins into a reputation in months.
Talking about your work is part of the work — invisible solving doesn’t build credibility.
The journalist questions (what’s valuable, what’s blocking, what would you do instead) are also the best problem-discovery tool you have.

Practical takeaway — Starting anything new — job, team, project — install the two threads this week: one shippable win per day-ish, and five journalist conversations.

Read the original post →

The $43 Billion Lesson in Holding: Eduardo Saverin’s Revenge

Main learnings

Inside this edition

Quote of the day

Stop Reading LLM Output: The Ladder of Better Formats

Main learnings

The First 90 Days: Two Threads That Compound

Main learnings