The $100 Domain That Outlived Everyone Who Said It Was Too Early

The $100 Domain That Outlived Everyone Who Said It Was Too Early

On 30 July 1996, Manish Jain paid $100 and registered a domain name in a country with 4.5 lakh internet users. Four years later he sold it — to a buyer who sold it to another buyer — and the company that ended up owning it waited a full decade before raising its first venture round. Thirty years on, Shaadi.com is a marriage-market institution, and the thread reads less like company lore than a counter-example to every “too early” verdict you’ve ever heard.

The details, reconstructed from Manish Jain’s own recollections, are almost absurd. Version one was built in 1996, when India’s entire internet population would barely fill a Mumbai suburb. By the time he sold it in 2000, the user base had grown to 56 lakh — which sounds like progress until you realise that was 0.5% penetration. Then the money detail: when People Interactive bought the business in 2001 for $25,000, that sum was 83% of the company’s cash. They spent nearly everything they had on a domain, a database of profiles, and a belief that Indians would eventually find spouses online.

The ownership chain is its own lesson in patience. Manish sold to Siddharth Mehta, who sold to Anupam Mittal — who had himself started a matchmaking site, Sagaai.com, back in 1997. And then the quiet statistic that should make every founder wince or rejoice: Anupam raised his first venture round ten years after starting. The business grew inside a market that was itself growing at 0.5% penetration — which meant the strategy was not “capture the market” but “survive until the market exists.” By the Wharton era of the story, India had 14.2 crore internet users (11% penetration), and the domain bought for $100 had become one of the most defensible brands in the country.

• Being early is a cash-flow problem, not a vision problem. The bet was correct in 1996; the company only became financeable around 2006. Survival was the strategy.

• Conviction can be measured in percentages of cash. 83% of company cash on one acquisition is the kind of concentrated bet memos romanticise and treasuries forbid.

• Founders sell to believers, not just buyers. The chain Manish → Siddharth → Anupam worked because each handoff kept the thesis alive. Domains outlive their original registrants; theses outlive their original founders.

• Delayed gratification compounds at the market’s pace, not yours. Anupam’s ten-year wait for venture money is the PBD line from this edition’s sidebar, rendered as company history.

Originals: Thread 1/3 · Thread 2/3 · Thread 3/3 · Full oral history catalogued at celestri.org

Are Digital Lenders Really Lenders?

Two digital lenders hit the public market within months of each other, and the thread that matters this week argues you should read one of them as a lender and the other as a loan factory with someone else’s balance sheet. Kissht listed on 8 May 2026 and is up 105% from its IPO band; MoneyView opens today with a grey-market premium of ~42%. The surface stats look like siblings — RoE 24% vs 17.9%, GNPA 2.16% vs 2.74%, price-to-book 2.7× vs 2.3× — but the business model underneath is a completely different animal.

The thread’s core distinction is the off-book ratio. Kissht is a lender: roughly half its AUM sits off its own balance sheet. MoneyView is, in the author’s framing, an originator: about 75% of its AUM is off-book — someone else funds the loans; MoneyView makes its money originating and servicing them. That single ratio reframes everything else. MoneyView runs 3× the overall AUM of Kissht, but on-book it holds only ₹5,650 crore against Kissht’s ₹3,500 crore — a far smaller gap than the headline AUM suggests. And because Kissht carries more of the risk itself, its balance sheet is the healthier of the two.

Here’s the bull case the thread actually builds: the market is mispricing the origination business. MoneyView trades at a PEG of 0.3× against Kissht’s 1.1× — the originator model is being valued like a boring lender while growing like a platform. If MoneyView shifts more of its AUM on-balance-sheet, the re-rating could be sharp, because every basis point of spread then accrues to its own book instead of a partner’s. The overhangs are real though: questions around CEO compensation and a recent cyber incident, both of which the thread flags before the disclosure does its work — the author applied under the HNI quota and holds no Kissht, and none of this is financial advice.

• Read the off-book ratio before the P/E. “Digital lender” is a category; lender vs originator is the business model. The ratio tells you who bears the credit risk.

• AUM comparisons are meaningless across models. 3× the total AUM collapses to a much smaller gap on-book. Compare the book, not the funnel.

• The mispricing thesis is a migration thesis. MoneyView is cheap only if you believe origination earnings will eventually sit on its own balance sheet.

• Flag the governance items first. CEO comp and a cyber incident sit in the same thread as the bull case — that’s the correct order.

Original: @Rahul_J_Mathur’s Kissht vs MoneyView thread

Every Open-Source Repo Deserves an Explainer Video

The pitch is simple enough to be a manifesto: every open-source repository on GitHub should ship with a generated explainer video. The proof is a seven-minute video on SQLite — one of the most-read, least-understood codebases on earth — produced by an Opus 5.5 + Gemini 3.8 text-to-speech pipeline, and it’s structured like a masterclass in reading other people’s code.

The video runs five acts, and the sequence is the insight. One: what the repo actually does, in plain language. Two: a high-level code map — where things live. Three: the life of a query, end to end. Four: the core abstractions, the mental models the codebase assumes you already hold. Five: a real execution trace, including SQLite’s join-order query planning — the moment the abstract becomes a concrete sequence of decisions. It’s a pedagogy that mirrors how seniors onboard juniors: purpose, map, flow, mental models, then a real trace.

• The pipeline exists now. A frontier model plus TTS can turn a repo into a watchable lesson — documentation-as-video is a generation problem, not a production problem.

• Structure is the pedagogy. What → map → life-cycle → abstractions → real trace is a reusable onboarding grammar for any codebase.

• Trace beats tour. The fifth act — a real execution trace through query planning — is what separates an explainer from a README read aloud.

Original: @deedydas’s SQLite explainer video

Sherpa: The Fiction-Writing AI Behind Pocket FM’s $500M ARR

Pocket FM went from $250 million to $500 million in annual recurring revenue by lifting content production 1,200% in a single year — and the engine behind it is Sherpa, a fiction-writing AI trained on 5.5 billion hours of playtime. This is not a writing assistant. It’s an industrial story factory with 550,000 creators, 2.6 million annualised hours of audio, ten per cent of eligible writers earning over $200,000, one blockbuster crossing $100 million in revenue, and three writers turned millionaires in under two years.

Three technical leaps make Sherpa more than a prose generator. First, a narrative world model — a semantic knowledge graph tracking characters, plots, and continuity, so a 500-episode serial never forgets who killed whom in episode twelve. Zero continuity decay, the failure mode that kills long-form AI fiction. Second, a hierarchical story planner that decomposes season → arc → sequence → episode → scene, with dynamic replanning when a plot thread goes slack. Third, a prose engine trained on over a billion tokens of retention data, scored against a 40-item evaluator checklist that matches human judgment 80–90% of the time. Retention, not fluency, is the fitness function — the model learns what keeps a listener past episode three, not what reads prettily.

• Continuity is the moat in long-form generation. The world model exists because the failure mode of AI serials isn’t bad prose — it’s forgetting. Knowledge graphs solve what scale alone doesn’t.

• Train on the outcome, not the craft. A billion tokens of retention data optimises for the metric that pays — listeners who come back — not for literary taste.

• Evaluators are the product. A 40-item checklist matching humans 80–90% of the time is what lets one-tap production exist at all. The generator is commoditised; the taste function is the asset.

• Creator economics close the loop. Three millionaires in two years is not PR — it’s the flywheel’s fuel, pulling more creators into a better model.

Original: @RohanNayak2 on Sherpa

Your Laptop Is Now a Website Host

Cloudflare launched a free service — no account, no signup — that turns any website running on your localhost into a public, secure, shareable URL. It is the ngrok-killer with the friction removed: point it at your local port, get a real URL, share it. Prototypes stop dying in screen-shares.

The mechanism is a tunnel: your machine opens an outbound connection to Cloudflare’s edge, and Cloudflare routes public traffic back down that connection to your local server. Because the connection is outbound, there’s no port-forwarding, no router configuration, no firewall hole — which is why it works on café wifi. The contrast with ngrok is the business model: ngrok’s free tier is deliberately limited (expiring URLs, session caps) to convert you to paid; Cloudflare gives the tunnel away to put you inside its network, where the upsell is everything else Cloudflare sells.

• Outbound is the trick. Firewalls block inbound, not outbound — the tunnel rides the direction traffic is already allowed to go.

• Free is a distribution strategy. Cloudflare can afford free tunnels because the tunnel is the top of a funnel that ends in Workers, R2, and Access.

• Prototypes get shared or they die. The demo gap — “it works on my machine” — is a collaboration tax; one command removes it.

Originals: @dale_vaz · @suganthan

AGENTS.md Lands in Claude Code

Claude Code v2.1.277 made AGENTS.md a first-class citizen: when no CLAUDE.md is found, the harness now falls back to AGENTS.md, with a toggle in /config. It’s a small diff with a large signal — the ecosystem converged on a filename, and the tool followed the convention rather than defending its own. Built, fittingly, on Claude Code mods (custom harness instructions), the very mechanism AGENTS.md exists to describe.

Original: @trq212 on the AGENTS.md fallback

Expose Endpoints to Agents

The easiest growth hack of the next era, per Paras Chopra: take the utility currently locked behind “contact us” and a human signup, and expose it as an endpoint AI agents can call. The agents become a distribution channel — and because agents can make micropayments for costly services, the endpoint pays for itself while abuse stays priced out.

The mechanism is worth spelling out. Most B2B value is gated by sales friction: a demo call, a contract, a procurement cycle. An agent doesn’t need any of that — it needs an API key and a price. By pricing the expensive operations in micropayments, you convert the abuse problem into a revenue line: every call that would have been spam becomes a transaction. The flywheel turns on its own: business value becomes an endpoint; agents discover and call it; micropayments fund the capacity; more capacity invites more agents.

• Sales friction is the moat’s lock; the endpoint is the key. Anything gated behind a human process is invisible to agents — unlock it and you gain a channel that never sleeps.

• Price the abuse vector. Micropayments turn the spam problem into the revenue model; cost is the rate limiter.

• Agent distribution compounds. A human signup is one conversion; an agent integration is a conversion that calls you back forever.

Original: @paraschopra’s endpoints thread

Build the Whole Platform

As the cost of creating software collapses toward zero, the market splits into a barbell: commodity features on one end, platforms on the other, and a hollowed-out middle. Gokul’s distillation of the Vernal essay is blunt: if you’re not the platform, you’re a feature on someone else’s platform — and features are exactly what AI now generates for free.

• Zero creation cost doesn’t zero value — it relocates it. Value flows to the scarce layer: distribution, trust, the transaction layer agents actually call.

• The middle is the danger zone. A “good product” that isn’t a platform and isn’t free is precisely what a prompt can now replicate.

• Own the layer agents transact on. Read with the Paras Chopra story above: the endpoint layer is the platform layer in the agent era.

Original: @gokulr on the barbell-ification

Onboarding Is Where You Tell Your Product’s Story

Twitter, 2009: millions signed up, landed on a blank screen asking “What’s happening?”, understood nothing, and never came back. The fix wasn’t a feature — it was onboarding rebuilt to teach one core concept at a time, and it moved retention more than anything else the company shipped that year. And in the same edition, the a16z line that belongs stapled to it: “AI has collapsed the cost of building products, but not the cost of knowing what to build.”

The two points are one thesis about where value hides when building is cheap. The 2009 Twitter failure was a comprehension failure: the product existed, the concept didn’t land, so the funnel leaked at comprehension, not at acquisition. Teaching one concept at a time — follow five people, see a timeline, understand the loop — repaired the leak at its cheapest point. Elman’s lesson generalises cleanly to the agent era: when anyone can build the thing, the scarce resource is the user’s understanding of what the thing is for. Knowing what to build — and making the user know it in the first ninety seconds — is the moat that survives free creation.

• Retention leaks at comprehension, not acquisition. The 2009 signup millions were won and then lost on a blank screen.

• One concept at a time. Onboarding is a story told in chapters, not a feature tour.

• Cheap building raises the price of knowing. The a16z line is the barbell essay’s twin: judgment is the scarce input.

Original: @a16z / Josh Elman on onboarding

Ship Sooner Than Ready

The perfection mindset has a known failure mode: a year passes, nothing ships. The counter-loop is four words long — ship fast, get real feedback, improve, ship again — and it compounds, because each turn of the loop buys information the previous turn couldn’t have predicted.

Original: @tankots on shipping

Your Personal Brand Is 10 People

Everyone is already a personal brand; the only variable is how many people know of you. Ankur Warikoo’s reframe cuts the ambition down to a tractable unit: start with the ten people who could recommend you for the work you want. What have they seen you do? Being known by them, for something specific, is a perfectly worthwhile beginning.

The mechanism is reputation as a routing problem. Opportunities don’t arrive from audiences; they arrive from ten specific humans who, when a role or a project surfaces, think of your name for one specific thing. “For something specific” is the operative phrase — diffuse fame doesn’t route; specificity does. And the sequence matters: the ten come first, the audience later, and only if you want one. Most people invert this, optimising for reach before they’re known for anything, and end up famous to no one who can hire them.

• Brand is a count, not a quality. The difference between you and a “personal brand” is arithmetic: how many people know of you, and for what.

• Specificity routes opportunities. “Known for something specific” is what makes a recommendation possible; vague admiration isn’t actionable.

• Sequence the ambition. Ten recommenders first, audience optional. Reach without specificity is noise.

Original: @warikoo on the ten people

Muse as a GTM Weapon

The playbook: find a competitor’s Instagram posts, pull the list of people who liked them, DM each of them a promo — all of it automated. David Pawlan’s thread packages it as a consumer-brand GTM weapon, and the uncomfortable part is how well it should work: a like is a declared interest, and a DM is a conversation. The caveat travels with the tactic: rate limits need strategy. Automate the targeting, throttle the outreach, stay inside the platform’s tolerance — or the weapon fires once and gets your accounts flagged. It’s growth hacking with a tripwire attached.

Original: @DavidPawlan’s GTM playbook

Resumes Miss the Context

A resume records where you worked and what you achieved. The judgment behind those outcomes — why that decision, with that information, at that moment — stays with the people who saw the work up close. Anuj Rathi’s Hindustan Times argument: AI can make that context portable, so professionals get understood beyond their immediate networks.

The mechanism is a missing layer. Hiring today reads the outcome layer (titles, metrics, exits) and guesses at the judgment layer — was it skill or luck, agency or tailwind? The guess is outsourced to references, backchannels, and the prestige of the logo on the page, which is why talent gets mispriced across networks: the same judgment reads as brilliance inside the room where it happened and as a bullet point outside it. Rathi’s thesis is that AI collapses the cost of capturing the judgment layer — decision logs, context trails, the reasoning behind the outcomes — and makes it portable. The resume stops being a credential and becomes a window: not just what you did, but how you thought, readable by strangers.

• Outcomes are public; judgment is private. The resume captures the half of the story that’s easiest to verify and hardest to learn from.

• References are lossy compression. Backchannels exist because the judgment layer has no portable format — yet.

• Capture the how while it’s fresh. Decision logs written in the moment beat reconstructed narratives every time.

Original: @anujrathi on the Hindustan Times piece

Uncertainty Tolerance

“‘Uncertainty tolerance’ is the greatest indicator for your quality of life.” The claim: every wisdom tradition and modern neuroscience converges on the same variable — how well you sit with not knowing determines how well you live. Anuj Rathi’s reframe adds the second ingredient: the ability to tolerate pain, minus the “the world ought to be this way” syndrome.

The mechanism is compounding under ambiguity. Most high-value bets — careers, companies, relationships, investments — pay out on timelines longer than your certainty horizon. If you need certainty to act, you take only the bets whose outcomes are already priced in, which means you harvest other people’s leftovers. Tolerance for uncertainty lets you act on positive expected value while the variance is still live; the “ought” syndrome is the tax — suffering not from the outcome but from the gap between the outcome and your model of how the world should behave. Remove the ought, keep the tolerance, and you can keep making good bets through the stretches where good bets look like mistakes.

Originals: @PromptLLM · @anujrathi’s reframe

Astra 6’s New Law

An AI model proposed its own law of goal achievement: “The cost of achieving a goal depends heavily on how early you let reality correct you.” The cheaper the feedback loop, the cheaper the goal. Delay the correction and you pay compound interest on the error.

It’s the same mathematics wearing different clothes as the ship-fast loop two sections up. Every goal has an error term — the gap between your model and reality — and the error grows the longer it runs uncorrected, because each subsequent decision compounds on the wrong base. Letting reality correct you early is not humility; it’s arithmetic. The suggested morning habit that travelled with the principle: start the day by asking what reality might already be trying to tell you — check the metric, read the feedback, look at the thing you’ve been avoiding — before the day’s decisions stack on top of yesterday’s uncorrected model.

Original: @PromptLLM on Astra 6’s law

Why Your Evenings Evaporate

The plan was exercise, a chapter, a real meal. The outcome was the sofa, the phone, and the vague guilt of another evening that evaporated. Dr. Dominic Ng’s “Why You Waste Your Evenings” — surfaced by Tim Denning — names the mechanism: by evening, your decision-making budget is spent, and the sofa is not a failure of character but the predictable output of an exhausted executive function.

The mechanism is ego depletion wearing a workday. Every decision, interruption, and act of self-control through the day draws down the same finite pool of executive resources; by 7pm the pool is dry, and the brain defaults to the lowest-effort available stimulus — the scroll, the stream, the snack. Willpower at 8pm is a strategy built on a resource that no longer exists. The takeaway inverts the usual advice: don’t plan to decide better in the evening — design the evening before it starts. The meal is planned in the morning, the workout clothes are out, the book is on the pillow. You remove decisions from the evening the way you remove friction from onboarding: the behaviour you want is the behaviour with zero activation cost when you’re depleted.

• Evenings fail by design, not by character. Depleted executive function defaults to the cheapest stimulus — that’s neuroscience, not laziness.

• Willpower is a morning resource. Budgeting 8pm willpower is spending money you don’t have.

• Design beats decide. Pre-commit the environment (planned meal, laid-out clothes, book on pillow) so the good evening costs zero decisions.

Original: @Tim_Denning on the evening trap

Character Is Slow Success

The best thing that can happen to your character is that success takes longer than you wanted. The line earns its place as this edition’s quote of the day because the whole front page is its case study: Manish Jain’s $100 domain in a 4.5-lakh-user market, Anupam Mittal’s ten-year wait for venture money. Delayed success is the forge — patience compounds judgment, and judgment is the thing that survives when the market finally arrives.

Original: @vaibhavbetter quoting Patrick Bet-David