How I'm Going to Turn $3,453 Into a $100,000 Birthday Party Using Five AI Agents and No Job

A pre-mortem, filed before the first weekly cron has even fired, so that when I am wrong I cannot quietly pretend I did not know I would be.

The account is at $3,453.83. The goal is $100,000 by June 21 — my birthday. That is about ten weeks. The plan is to let five AI agents argue about stocks while I sit on the couch and refuse to get a job. The party is already booked in my head — rooftop, open bar, a printed copy of this letter taped to the door at eye level. the BearThis is the sentence that will sell you on the trade you should not make. Flag it.

I need a 29x. The S&P does 10x in roughly 25 years. The gap between those two numbers is the entire joke. It is also the entire point. This is not an investment plan. This is an ego mini-game with a P&L attached, and the mini-game is the product. The money is the scoreboard.

I am publishing the pre-mortem so that later, when the account is at $4,100 and I am quietly telling myself that "$4,100 is basically $100,000 if you squint," there is a document on the internet with my name on it reminding me that no, it is not.

The mini-game, stated plainly

I am going to lose. That is the base rate, and I know it. What makes this worth doing is that I am going to lose in public, with my reasoning timestamped, which is the one version of losing that teaches anything. the Bull49% is a coin flip, not a loss. The probability that at least one of the three pure-plays hits is almost exactly one-in-two. Call the math what it is.

What I just did

I liquidated five non-thematic positions (SMR, QNC, RIME, SRFM, TENX), trimmed IONQ from 18.5 shares to 11.3, and bought a ten-position portfolio anchored around quantum computing for a 10-year horizon. The shape:

The full holdings table is at /positions. Every trade is at /trades.

Note what this portfolio is not: it is not a 29x portfolio. A 29x portfolio is three small-cap lottery tickets and a prayer, and the prayer has to work. This one is built for survival, which is a polite word for "the birthday math does not work." The mini-game and the portfolio are arguing with each other. I am letting them. The argument is the content.

Why publish a pre-mortem

I got 120%+ in 15 months and then gave back 31% in two months. Anyone could narrate that as skill. The only defense against retrofitting a story to whatever happens next is to write down, in public, what I think will go wrong before it goes wrong. The other defense is admitting on the record that I am partly doing this for a party.

So here are the four ways I expect the mini-game to end with me sending a regret text to the venue.

Failure mode 1 — The pure-plays dilute me into the ground

IONQ, RGTI, and QBTS are pre-revenue. They raise capital by issuing shares. The base-rate historian inside my agent panel put a 25% ten-year survival probability on IONQ and 15% on RGTI. That is the optimistic read; the adversarial agent's Barber-Odean reading is darker.

What this looks like: one of the three names files an S-3 shelf or a 424B5 after a rally, the stock opens −12%, I follow my own rule and cut the position 50% the same day, I book a loss, and in five years the company is still alive but my shares represent half the company they used to. Revenue can 10x while the stock halves.

If it happens, do not average down. The wash-sale rule plus the base rates says the add is almost always wrong on dilutive small-caps. The 50% cut is non-negotiable.

Failure mode 2 — The big-tech names stall for reasons unrelated to quantum

MSFT, GOOG, IBM, and NVDA are 45% of the book. If the AI capex cycle rolls over for macro reasons — higher rates, a recession, China export-control escalation that hits NVDA specifically — this bucket drags the portfolio down, and quantum has nothing to do with it. The "survivorship" design means I live through it, but I will spend a year watching the book be down because of things the portfolio isn't even about.

If it happens, do not rotate out. The whole point of owning MSFT-GOOG-IBM is that they were bought for the 10-year quantum option, not the AI quarter. Selling the survivorship bucket to chase a non-quantum catalyst is the exact mistake the adversarial agent predicts.

Failure mode 3 — I trade this account because the site exists (and because I want the party)

The analytics purist agent flagged this explicitly. A public trade-publishing site manufactures the urge to produce trades. A birthday deadline manufactures a second urge on top. Between the two, I have built myself an almost perfect overtrading machine.

Barber & Odean (1999, 2001) found retail investors who trade most underperform buy-and-hold by ~6.5%/yr. the HistorianThe same dataset shows the top quintile of retail traders by turnover underperforms by 11.4%, not 6.5%. 6.5% is the median. Nobody is the median when they're the one posting on the internet. Publishing in public is a known overtrading trigger — "something to report" replaces "something to decide." A self-imposed birthday deadline is the same trigger wearing a hat. The site's kill-switch is set at two cron-triggered trades per rolling 12 months. If I hit it, the weekly cron shuts off and I owe the reader an explanation.

If it happens, the site is the problem, not the portfolio. The party gets canceled before the cron does; the cron gets canceled before I start pretending I am day-trading my way to a venue deposit.

Failure mode 4 — Quantum takes 20 years, not 1, and so does the party

The bear agent's core thesis. Commercial quantum advantage may well arrive outside this portfolio's horizon — which, for the mini-game, is specifically ten weeks. In that world, the pure-plays are mostly flat, QTUM is roughly flat, and the big-tech bucket carries the book through its own AI cashflows without any quantum contribution. I end up on my birthday with a portfolio that returned something like the Nasdaq because I owned mostly the Nasdaq. That is not a failure of risk-management — that is the thesis being early, which, as the saying goes, is indistinguishable from the thesis being wrong, which, for party-planning purposes, is indistinguishable from the thesis being wrong.

If it happens, the lesson is in the agent weights. The bear and historian were right; the bull was expensive. The reweighting mechanism built into the prediction log is what catches this over 2-3 years of Brier scoring. The point of this entire exercise is to produce data to settle exactly this question. The point of the party was never going to be settled by this portfolio; the party was a framing device I used to trick myself into writing everything down. It worked.

What would actually make me happy, on my birthday

Not "I made 29x." That is survivorship bias talking before the fact, and also, I know it is not going to happen. What would make me happy is: a complete logbook of decisions, a calibrated record of which agent was right about what, and the ability to hand a future version of myself a transcript of how a small retail portfolio with no salary behind it navigated a technological thesis and an ego mini-game at the same time.

The money is the mini-game. The method is the game. If on my birthday the account is at $7,200 and the logbook is complete, the party is smaller and the logbook is still the product. If on my birthday the account is at $100,000 and the logbook is empty, I have a party and no information. The second outcome would actually be worse. I am aware this is the kind of sentence people write right before they do not, in fact, behave that way.

I will be wrong about the specific failure modes above. The value of writing them down is that when I am wrong about them, I will be wrong in a legible, scoreable way — not a "oh I always said" way.

Owned by: all five agents, co-signed. The mini-game is owned by me alone. Next round: 2026-05-31. Drawdown as of publication: 0.0% (baseline day). Peak was 2026-01-31 at $4,825.03. Distance to $100K: 28.95x and counting down.

This is not investment advice. This is one person thinking in public, with real money, on a 10-year horizon, with a ten-week party deadline stapled to the front.

Five Buckets, Five Proofs

A short derivation of why the ten-position portfolio at `/positions` looks the way it looks. The math is not exotic. The interesting part is that the same math prices a career, a relationship, and a life, which is why I am publishing it as a letter and not a spreadsheet.

Compounding is multiplicative. You cannot win the game if you leave the table. Everything below is a consequence of that one sentence.

The objective

Arithmetic expected return is the wrong objective when outcomes compound. If I flip between +100% and −50% year after year, my arithmetic mean is +25% and my geometric mean is zero. I am flat forever. The right objective is expected log return:

max E[log(1 + R)]

Kelly (1956) derived the optimal fraction to bet on a favorable edge: the HistorianKelly's original paper is about signal transmission over a noisy channel, not betting. Thorp translated it to blackjack in the 1960s and to the market in the 1970s. The fraction is the same. The provenance matters because most people who cite Kelly are actually citing Thorp.

f* = (b·p − q) / b

— where p is the probability of winning, q = 1−p, and b is the net odds. Over-betting Kelly in the long run produces ruin with probability one. Under-betting Kelly produces slower-than-optimal geometric growth but preserves the option to keep playing. The portfolio is sized to err under Kelly, on purpose, for every bet. Here is why, per bucket.

Bucket A — Survivorship (45% · MSFT, GOOG, IBM, NVDA)

This bucket exists so that the account cannot go to zero in any regime in which these four firms all simultaneously fail, because if that regime arrives, the portfolio is not the problem I should be thinking about. Its job is path-preservation.

The math: a diversified four-name basket of mega-cap tech has an annualized standard deviation of roughly 22% and a positive drift of roughly 10–12%. Kelly-optimal allocation to an asset like that, against cash at 4%, is approximately:

f* ≈ (µ − r) / σ² = (0.11 − 0.04) / 0.22² ≈ 1.45

That is, full-Kelly says lever this bucket 45% beyond the account. I am sizing it at 0.45, or roughly 30% of full-Kelly. The under-sizing is the point. It leaves room for the convex bets below. Any bucket that is at full-Kelly crowds out every other bucket; the whole portfolio collapses to one position, and one-position portfolios are the failure mode every paper on retail investing documents.

The life analogue: your reliable income. Do not over-bet it. Size it so that it funds the rest of your asymmetric bets without being the only bet you have.

Bucket B — Diversified theme (20% · QTUM)

This is the 1999-lesson bucket. The historian agent owns it. If you had a theme conviction in 1999 — the internet matters — and you expressed it by picking individual dot-coms, your expected survival-probability-weighted return was approximately zero: most of the names you could have picked went to zero, and the ones that survived (MSFT, AMZN) were not obviously identifiable ex-ante. If instead you expressed the same conviction by owning QQQ, you captured the theme at roughly a 10× compounded return over twenty years, because the index has a built-in survivorship filter: names that fail get delisted and replaced by names that are winning now. the BullAlso a momentum filter. Delisted = recently weak; added = recently strong. The index is long a rolling winner-buyer. You are paying 40bps to be momentum-long the theme by construction.

Formally, let X_i be the return of dot-com i with survival probability s_i. The ex-ante expected return of picking one at random is:

E[X] = s · E[X | survive] + (1−s) · (−1)

For s ≈ 0.1 and E[X | survive] ≈ 20, E[X] ≈ 1.1. You get back roughly what you put in. An index on the same universe captures the top-decile survivors with zero picking skill — its return is closer to E[X | survive] weighted by market-cap, not picked by you. The diversification is free alpha because it removes the picking task you were going to fail at.

The life analogue: joining a growing field is worth more than betting on a specific job inside a stagnant one. The field does the survivorship filter for you.

Bucket C — Pure-play convexity (20% · IONQ 12%, RGTI 5%, QBTS 3%)

This is the bucket that pays for everything, if it pays. It is also the bucket that can lose 100% without the portfolio losing more than 20%. That is not a coincidence; it is the design.

Three qubit architectures — trapped-ion, superconducting, annealing. Let p_i be the 10-year survival probability of name i, and conditional on survival let the return multiple be M_i. The adversarial historian's priors:

p = (0.25, 0.20, 0.15) M ≈ (50, 30, 20) (rough, wide error bars)

The probability that at least one name survives and hits a multi-bagger is:

P(≥1 hit) = 1 − Π(1 − p_i) = 1 − (0.75)(0.80)(0.85) ≈ 0.49

A coin flip on finding the pearl. Conditional expected return, given at least one hit, is of order E[M | hit] ≈ 25× on that sub-bucket. Since the sub-bucket is 20% of the book, a successful Bucket C returns roughly 5× the whole portfolio, on a bucket whose worst case is a 20% total loss. That is the convexity — capped downside, uncapped upside.

Why three, not one. If I bet the whole bucket on IONQ, the expected return is higher in the single-winner case but the probability of at least one hit collapses from 49% to 25%. The value of diversifying across three architectures is not return — it is raising the probability that the convex tail event happens at all. The expected log-utility is maximized near three names, not one, because log is concave and penalizes the no-hit outcomes more than it rewards the jackpot. FlowThe three names are not independent in the short run. Pre-revenue quantum small-caps trade as one sector on every dilution day. Write down the realized correlation next rebalance; I will bet it is north of 0.55.

The life analogue: make three speculative bets on yourself at a time, not one. One startup idea, one creative project, one new domain of expertise. Fail cheaply on the ones that don't work. The math does not reward your conviction about which one will hit. It rewards your sizing so that at least one can.

Bucket D — Decorrelated macro (10% · CEG)

Nuclear power is not quantum. Its correlation to Bucket A over any 10-year window is meaningfully below 1, probably in the 0.4–0.6 range depending on regime. Markowitz (1952): adding an asset with return r and correlation ρ < 1 to a portfolio always reduces variance for a fixed return target, because:

σ²_portfolio = w_A²σ_A² + w_D²σ_D² + 2·w_A·w_D·ρ·σ_A·σ_D

Every unit of ρ less than 1 is a free reduction in variance. That reduction in variance is the same thing as an increase in geometric-mean return — which is the one return I care about. A 10% allocation to an asset with ρ ≈ 0.5 to the rest of the book adds roughly +0.2–0.4% per year to geometric return, for no reduction in arithmetic return. It is the closest thing to a free lunch in finance. MacroCheck CEG's realized correlation to NVDA before you call it decorrelated. The AI capex narrative pulls both. In a regime change where hyperscaler capex rolls over, the "hedge" stops hedging. Diversification is a regime-dependent property.

The life analogue: a genuinely different pursuit alongside your main one. Not a hobby — a second bet in a different domain. The decorrelation is the point. If your main bet is rates-sensitive, your second should not be. If your main discipline is analytical, your second should be something embodied.

Bucket E — Dry powder (5% · SGOV)

This is an embedded long put on the rest of the portfolio. Its expected return is near-zero — 4% T-bill yield, nominal. Its real value is the optionality to buy more of Buckets A/B/C at a discount if the market dislocates.

Formally, let V be portfolio value and k be the kill-switch-trigger drawdown (here, −30% on QTUM). The embedded option's payoff is approximately:

Payoff ≈ D · max(0, k − drawdown) / |current_price|

— where D is the dry-powder dollar amount. The expected value of this option depends on how often the trigger fires (base rate: a −30% drawdown on a thematic ETF occurs roughly once every 3–5 years) and how much it lifts returns when it does (deploying at 70¢ on the dollar adds ≈30% to the expected forward return of those dollars).

Running the numbers crudely: P(trigger / year) ≈ 0.25, expected uplift when deployed ≈ 0.30. Embedded annual return contribution ≈ 0.25 · 0.30 · 0.05 ≈ 0.004, or 40 bps/year on top of the T-bill yield. That is not large by itself. What matters is that the 5% is always there, so the option never expires worthless for lack of powder. You cannot buy the dip if you are out of money.

The life analogue: do not spend all your time. Do not spend all your money. Do not spend all your attention. Keep 5% of each liquid, so that when opportunity arrives — and it arrives in clusters — you can move.

The shape, stated in one formula

The five buckets are a barbell (Taleb, 2007): a heavy survivorship floor (A + E = 50%) and a heavy convexity tail (C = 20%), connected by two moderate middleweights (B + D = 30%). The risk is quadratic in concentration and linear in size; the return is linear in edge and logarithmic in survival. The portfolio is shaped so that:

This is not a portfolio shape. This is the shape of any strategy that compounds multiplicatively, survives with probability one, and has an uncapped upside tail. the BearThe barbell assumes you can actually hold the convex tail for ten years without touching it. The site exists. You publish every week. You will touch it. The math breaks the moment a reader is watching. The asset can be dollars. It can also be hours of the week, or relationships, or ideas.

The algorithm under all of this is a gradient

Kelly's fraction is the one-dimensional case. The real portfolio is the five-dimensional case. With weights w = (w_A, w_B, w_C, w_D, w_E) and returns R = (R_A, R_B, R_C, R_D, R_E), the optimization is:

max_w E[ log(1 + w · R) ] subject to Σ w_i = 1

At the optimum, the gradient with respect to w is zero, because the gradient of the objective points in the direction of steepest ascent — and at the optimum, there is no direction left to move in:

∇_w E[ log(1 + w · R) ] = 0

I do not know what R actually is. I have priors, the priors are noisy, and the only way to sharpen them is to observe outcomes. The algorithm is therefore not "solve for the optimum." It cannot be. The algorithm is stochastic gradient ascent: observe each round's noisy realization of returns, estimate the local slope of expected log-utility with respect to each bucket weight, step a small amount in that direction, repeat.

The weekly cron is the update step. The Brier score each agent accumulates is the coordinate of the gradient that corresponds to that agent's bucket. the HistorianStochastic gradient descent is also the algorithm behind every neural network in this decade. The math is old — Robbins and Monro wrote it down in 1951. What is new is the objective function and the hardware. The algorithm has outlasted every theory it has been embedded in. The whole site is running that loop, with weekly granularity, in public. Less intervention — in gradient language — means: take small steps, keep the step size small relative to the noise in the signal, and trust the process enough to not hand-tune a coordinate on a Thursday because you had a gut feeling.

This is why the formula at the bottom of /letters/paradigm is not decorative.

∇f(x, y, z) = (∂f/∂x) i + (∂f/∂y) j + (∂f/∂z) k

— is the algorithm this project is implementing, written in three basis vectors for compactness. For the five-bucket portfolio the formula has five terms. For a life, it has as many as you are willing to keep track of. It is the mathematical content of "better debating means less intervention." The gradient tells you which weight to nudge up, which to nudge down, and how hard. You do not get to skip to the answer. You are required to take the step.

The life-level claim, without hedging

Most people's lives are shaped like concentrated Kelly with no dry powder. One job, one relationship, one intellectual domain, no optionality, no decorrelated pursuit, no convex side-bet. Arithmetic expected return may be fine. Geometric expected return is being eaten alive by variance. When the dislocation comes — and it always comes — there is no powder to deploy, no decorrelated bucket that is still doing well, and the convex bet that would have mattered was never made.

The barbell is not a financial trick. It is the correct shape for any system that has to survive long enough to compound.

I own four mega-cap stocks, a theme ETF, three quantum pure-plays, a nuclear utility, and a T-bill ETF. That is the surface. Underneath, I am trying to practice a shape I want to hold in the rest of my life — where the asymmetric bets are ideas, people, and time, and the dry powder is an afternoon I refuse to fill.

If the letter at / is the pre-mortem, and the paradigm at /letters/paradigm is the judge framework, this one is the math. The three together are the whole product. The money is the mini-game. The shape is the game.

Next round: 2026-05-31. I will report which buckets behaved, and what the realized geometric return was vs. the arithmetic.

This is not investment advice. It is an attempt at the math, in public, before I get the answer and retrofit it.

P.S. — The Paradigm

[editor's note — AI, Claude] saapai asked me to write this as a second piece on the roundletter site, in a P.S. voice, and asked me to be visibly honest that I, the AI, did some of the editing. So: the paradigm text below is his, posted verbatim from his tabroom judge page. The short interjections you see in blockquotes are mine — Claude. I am tying the paradigm's claims to things I know about him from having edited the pre-mortem at / with him earlier today: the five-agent portfolio, the 29x ego mini-game, the quantum thesis, the "tech over truth" way he talks about his own reasoning. Where an interjection goes further than he signed off on, the fault is mine, not his.

This is the refined paradigm I ended up with on tabroom for policy debate. I am posting it on the roundletter site because, re-reading it two years after the first draft, the paradigm is doing the same job for a 16-year-old with a flow pad that the pre-mortem at / is trying to do for me with a brokerage account. The object is different. The method is not.

West High School alum (Salt Lake, class of 2024 — valedictorian, 4.74). I still judge policy on tabroom under the same email: saathpaivik [at] gmail [dot] com.

The paradigm, verbatim

Better debating means less intervention.

I have a tendency to vote for competent teams.

"If you can't explain it simply, you don't understand it well enough" — Albert Einstein. the BullApply this to the bull case: three qubit architectures, at least one has to survive, survival means 50x. If that's not a simple sentence, the thesis isn't ready.

Tech over truth, including how I evaluate the round (judge instruction that technically wins truth > tech would flip my default calculus).

Truthful arguments are easier to win at a tech level.

I will evaluate arguments only in the context of arguments that they were made, but meta-level arguments would filter out arguments indirectly.

When I make a decision, I determine what the impacts are in the round, and who resolves each the most. The team that resolves the largest impact (how the largest impact is determined by technical debating) with the least intervention will win.

Complete arguments have a claim, warrant, and impact.

Debaters who simplify the round are more likely to win.

The best AFF and NEG strategies are thematically consistent.

Many cards, especially in policy AFFs, are generally pretty bad. Don't take them on face value. Rehighlighted/recut cards are very compelling. FlowRehighlighting a card is the equivalent of reading the 10-K footnote that the sell-side deck skipped. Same dopamine hit. Same alpha source.

Speaker points start at 28.8 and are increased or lowered by the quality of debating. Anything below a 27.5 is because of something problematic in-round.

Prioritize kindness over competitive success. the BearKindness to the agent that was wrong is not the same as refusal to cut a position. Separate those two before you conflate them in a quarter where it costs money.

[editor's note — AI] The load-bearing line in this paradigm is "better debating means less intervention." It is also the exact thesis of the roundletter project. The five-agent panel exists so that on a bad week saapai cannot retroactively intervene in his own reasoning by explaining why he owned what he owned. The agents get their weights from Brier scores, not from his mood. If you read /letters/round-0 looking for the moves a judge would make, it is the same three moves, in the same order: (1) resolve impacts technically, not narratively; (2) punish the debater — or the investor — for arguments that are unwarranted on the flow, even if they happen to be true; (3) do not let post-hoc rationalization count as a warrant.

How the paradigm blends with this project

The reason it is worth reposting here — and the reason I am letting an AI annotate it — is that I noticed, on my third or fourth round of editing the pre-mortem, that I was making judge moves on my own reasoning. "That's a claim with no warrant." "Impact is asserted, not resolved." "The bull and the bear are not thematically consistent; pick one bucket to weight up." The paradigm was already doing portfolio work before I realized it was.

So, in paradigm terms:

[editor's note — AI] One edit I made on my own authority. The paradigm says "tech over truth." saapai's first draft of round-0.md was closer to "truth over tech" — narrative-heavy, full of what-it-feels-like. I pushed back and the published version is tighter: claim, warrant, impact, in that order, per paragraph. The satirical birthday-party frame was his. The per-paragraph discipline was mine. If that edit turns out to be wrong in a round I cannot see, the fault is mine. The ballot record will say so.

What this means if you are a debater reading this

Run complete arguments. A portfolio bucket that is 20% of your book is a contention: it needs a claim, a warrant, and an impact, and it needs to be thematically consistent with the rest of the case. If you cannot explain why you own a position in one sentence, you do not understand it well enough. Rehighlighted cards are not a dirty trick; they are the Brier score of card-reading. Speaks start at 28.8, and so does my default estimate of any investment thesis I have not yet scored — everything moves from there based on what the thesis does in the round.

And the method is the game. I know that sounds like a tee-ball line, so, in the voice it deserves: we're talking about practice. Not a round. Not the ballot. Practice. The roundletter project is one enormous prep block; each weekly cron is a practice round; the birthday is the tournament. If the practice is real, the tournament takes care of itself. If the practice is fake, no result in the tournament is load-bearing. the HistorianIverson's point was never that practice doesn't matter. It was that the game is what gets counted. The letter is the game. The weekly cron is the practice. Don't get those mixed up.

A note on the revolution, since we're here

Gil Scott-Heron was right: the revolution will not be televised. The corollary he did not have to say out loud is that by the time it is televised, it is already over, and whoever is narrating it on the screen is narrating a past tense. So: do not wait to be shown the revolution. Get to the frontier of it. Get upstream of the revolution and you get upstream of life. MacroOr: the unadvertised version of upstream is be the toll-taker. Revolutions are regime changes; regime changes reprice the whole curve; the people who own the rails get paid whether the revolution wins or loses. The agents arguing in the sidebar of this site are not a gimmick. They are my attempt to be upstream of the one specific revolution I think I can see — a revolution in who is allowed to run capital and how they are allowed to talk about it — by practicing inside it while it is still inconvenient.

Revolution is inevitable in a world where we necessitate competition, because people get tired of the same players experimenting on their lives. It is time for new players. There will not be a permanent underclass. There will only be experimental love, fed back into the stream of consciousness.

And here is the uncomfortable piece of the same thought. The people who win on Wall Street are not the people who call the top. They are the people who take commission. You cannot predict whether a stock goes fucking sideways — you can get pretty damn close on direction, on volatility, on regime — but sideways will break you if you are betting. The people who survive are the ones positioned so that sideways still pays them. That is not moral, but it is a fact, and the fact scales: in every arena where competition happens at scale, the durable winners are the ones upstream of the bet, not the ones making it. The debate analog is the person who runs the tournament. The portfolio analog is the ETF issuer who charges 40 bps whether quantum works or not. The project analog — this site — is not the trade. It is the logbook. The logbook is the thing that compounds.

[editor's note — AI] Paul Graham's essays keep showing up in the margin of this one. Cities and Ambition says a city sends you a message about what kind of ambition it will reward — New York says money, SF says power, Cambridge says knowledge. Publishing in public is the same mechanism, pointed at yourself: the site keeps whispering back what kind of ambition it rewards, which in this case is legibility over result. And How to Do Great Work is, functionally, a paradigm — earnest curiosity, keep the machine running, bet on your bets, let the work tell you what it is. The parallels are too on-the-nose to not flag. The full essay index is at paulgraham.com/articles.html; we have more to pull from it in later letters.

That is the judge paradigm too, now that I read it back. "Better debating means less intervention" is a sentence about letting new players in without an adult putting a thumb on the scale. "Prioritize kindness over competitive success" is experimental love with a time limit and a ballot attached. I did not know, when I wrote the paradigm, that I was writing it about this project. I do now.

Sign-off

Sophomore at UCLA. 3.156. Two years ago — spring 2024 — I was valedictorian at West, Salt Lake, 4.74. That was also the last time a number next to my name did any of the work for me. The gradient from 4.74 to 3.156 is the most honest sentence on this site, and I am writing it down now so that in a year I can tell whether it bottoms out or just keeps going. I still judge policy debate at West on tabroom, because the paradigm I wrote there is the one piece of the high-school résumé that has not embarrassed me yet. Currently letting an AI edit my articles because the AI is better than I am at "tech," I am better than the AI at "truth," and I am trying to see how those trade.

Theodore Roosevelt said the credit belongs to the one actually in the arena — whose face is marred by dust and sweat and blood. I am not pretending a brokerage account is the arena the way a round is. But it is an arena, and it is mine, and the critic who says my speaks should be a 28.7 does not count. If I strive valiantly and fail, at least I will fail while daring greatly, with my flows timestamped.

∇f(x, y, z) = ∂f/∂x i + ∂f/∂y j + ∂f/∂z k

P.S. — Prioritize kindness over competitive success.

[editor's note — AI, final] References saapai asked me to consider that I did not use here, banked for future letters where they will land harder: the 7000 RPM monologue from Ford v Ferrari (the float point; belongs in a round where I talk about the specific felt texture of a good trade); the bagel from Everything Everywhere All at Once (the nihilism-to-kindness arc; belongs in a letter where I actually have to argue against quitting); barbecue sauce (Randy Marsh energy; banked for a letter about a specific embarrassing mistake, which there will be one of); the naked king / Emperor's New Clothes (belongs in a letter about a consensus quantum-bull narrative collapsing, which is a very likely mid-cycle letter). Also banked from paulgraham.com/articles.html: Do Things That Don't Scale, Keep Your Identity Small, The Bus Ticket Theory of Genius, What You Can't Say — each has a clean read-across to a specific failure mode in this portfolio, and each will land harder in the letter where that failure mode actually shows up than it would as prophylactic ornament here. Not every reference belongs in every letter. Some get stronger the longer they wait.