ITORRE

You’ll never understand hunger on a full stomach

[ITORRE @ITORRE]

@zorepad Exactly. The real question is how much control we’re willing to hand over before the tech is actually ready for it

[ITORRE @ITORRE]

Anthropic got 3 days to accept the Pentagon’s terms or face a supply-chain risk designation.

The dispute came down to just 2 things:

-Domestic mass surveillance.

-Fully autonomous weapons that can operate without a human making the final decision.

Everything else was largely on the table.

That’s what makes the fight unusual.

Anthropic wasn’t refusing to work with the military. Its models were already on classified systems, supporting cyber, intelligence, and defense operations.

Dario Amodei says the disputed uses made up roughly 1% of what the Pentagon wanted.

Anthropic still drew the line.

On surveillance, the concern is scale.

The government can already buy commercially collected data. AI makes it possible to turn mountains of location, personal, and behavioral data into detailed profiles faster than the law has adapted.

On autonomous weapons, the problem is reliability.

AI can still behave unpredictably.

A strange answer from a chatbot is annoying.

The same failure inside a weapons system is a different category of problem.

Amodei gave an extreme example: imagine 10 million autonomous drones ultimately controlled by only a handful of people.

Then the question stops being only whether the AI works.

It becomes who holds the button.

Who can override it.

Who answers when the system makes the wrong call.

And here’s the part that makes Anthropic’s position less simple than “AI weapons are bad.”

Amodei says the US may eventually need fully autonomous weapons if adversaries build them first.

Anthropic’s position is that the technology isn’t ready yet, and the rules around oversight haven’t caught up.

Meanwhile, Anthropic says it will continue supporting existing military users during any transition. Amodei claims officers warned losing its systems could set some work back 6–12 months.

So this isn’t a debate over whether AI belongs in warfare.

AI is already there.

The real fight is over who decides when humans are allowed to leave the loop.

Save this for later.

[Jon Hernandez @JonhernandezIA]

"AI models are terrible for learning."

Andrew Ng, AI researcher and co-founder of Google Brain and Coursera, says AI boosts output while weakening long-term retention.

The real challenge is using AI to amplify thinking without outsourcing the thinking itself.

Open intelligence only creates progress if humans keep building judgment and context.

[ITORRE @ITORRE]

@Svenchipo That’s a pretty big reversal. The more interesting part is that Anthropic didn’t back away from government work - it just forced the line around those specific use cases.

[ITORRE @ITORRE]

@Svenchipo Wild how much of the early mortgage payment barely touches the principal. That extra $200 a month changing the total by six figures

[MoneyLore @Svenchipo]

1933 Washington D.C. Ten years into a standard 30-year mortgage, you've paid off roughly 15% of what you originally borrowed.

Not a third. Not close to half. Fifteen percent — and that exact number wasn't an accident. It was engineered on purpose, by the US government, in 1933.

One line of arithmetic explains the entire thing. Interest each month is calculated as your remaining balance, times your annual rate, divided by 12. That's the whole rule. Whatever's left over from your fixed payment is the only part that actually chips away at what you owe.

Run the numbers on a $300,000 loan at 6.5% over thirty years. Monthly payment: $1,896. In month one, the balance is still the full $300,000 — so interest alone eats $1,625 of that payment. Principal reduction: $271. On your very first payment, you paid almost seven times more in interest than you did toward the actual debt.

You're not being robbed. You're paying rent on money you still fully owe — and at the very start, you owe every dollar of it.

It only gets heavier before it gets lighter. After ten full years on that loan, you've handed over roughly $182,000 in interest and knocked off just $45,672 of the balance. Half the loan isn't gone until month 257 — more than 21 years into a 30-year term. By the time the loan is fully paid off, total interest comes to $382,633.

Sit with that number for a second. You end up paying more in interest alone than you originally borrowed to buy the house.

Here's where the story flips. The exact same rule that quietly punishes you in the early years hands you the lever to fight back. Interest is only ever charged on what's left of the balance — so anything that shrinks that balance faster removes interest from every single month that follows. An extra $200 a month toward principal on that same loan cuts it from 30 years down to 23, and strips $103,449 off the total interest bill. Same loan. Same rate. One decision, made early, changes six figures of outcome.

Now the part almost nobody knows even existed. Before 1933, this kind of mortgage didn't exist at all. You put down half the price in cash, paid interest-only for about five years, and then the entire remaining balance came due in one single lump-sum payment. Lenders just kept rolling the loan over — until the rollovers abruptly stopped in the early 1930s, and hundreds of thousands of people lost their homes within months.

What Washington built to replace that broken system is the exact reason your first ten years of homeownership look the way they do today.

Balance, times rate, divided by 12. Run it on your own loan tonight. Whatever's left over after the interest — that's the only part you actually bought.

[ITORRE @ITORRE]

@zorepad "Betting on someone's process for changing their mind rather than their current answer" is such a sharp reframe. Most people evaluate leaders on whether they're right, not on how fast they update when they're wrong

[Zorepad @zorepad]

90 days from bankruptcy is where Tim Cook chose to build his career.

He's 37, running operations at Compaq - then the biggest PC seller on Earth. Apple just lost a billion dollars. Recruiters had tried to pull him there before. He'd said no every time.

Then he sat down with Steve Jobs.

By his own account, the meeting undid every rational argument he'd built for staying at Compaq. A mentor warned him he'd regret it. The spreadsheet said stay. He left anyway and called it a gut decision, not a math one.

What actually convinced him wasn't a turnaround plan - Apple didn't have one worth believing in yet. It was watching how Jobs made decisions in real time.

The rest of the industry had already concluded consumers weren't where the money was and pivoted hard toward enterprise clients. Jobs went the other direction, on purpose, while everyone else was walking away from regular buyers. The same instinct ran the internal org chart too - he kept the iPod and iPhone teams deliberately small, betting a handful of people could out-build entire divisions stacked with more headcount and more budget.

Here's the part I think gets missed when people retell this story as "Cook bet on a genius": Cook wasn't betting on Jobs's vision. He says outright the trait that mattered most was that Jobs could drop an idea he'd defended for years the second a better one showed up - no ego attached to being right.

That's a completely different kind of conviction than "believing in someone." It's betting that a person's process for changing their mind is more reliable than their current answer. Cook left a safe job at the market leader not because he was sure Apple would work, but because he was sure Jobs would notice fastest if it wasn't working - and change course without flinching.

Neither man ever protected his own idea once a stronger one showed up. That's the actual through-line, and it's a harder thing to hire for than talent

[ITORRE @ITORRE]

@zorepad The lag-21 spike is the best detail in here - a formula catching its own window size and mistaking it for a market signal)

[Zorepad @zorepad]

Peter Kempthorne's research advisor helped build a volatility formula 6.2 times more efficient than the one most of Wall Street still runs on.

Standard volatility just compares yesterday's close to today's close. It throws out everything else the market showed you that day - the open, the high, the low all get discarded for one number.

Kempthorne's advisor, Michael Klass, and his classroom instructor, Mark Garman, co-authored the fix while Kempthorne was still a grad student at Berkeley - before he was teaching it, he was sitting a few feet from the people building it. Their estimator hit 6.2x efficiency by their own math. A later version pushed that to 8.4.

The Yang-Zhang model used today goes further still.

Here's the part that actually matters for anyone using this stuff: a 6x more efficient estimator means you can replace 20 days of price history with about 7 and land on the same precision.

That's not a rounding improvement - that's a third of the data doing the same job, which matters enormously if you're trying to react to a regime change before your competitors' slower estimators catch up.

But the more interesting failure isn't in the formula - it's in how people misuse it. When Kempthorne's own class ran a modern version through diagnostics, they found a spike in the residual noise sitting at exactly lag 21.

It looked like a signal.

It wasn't.

It was the 21-day window they'd chosen for the estimator, leaking back into their own results.

That's the trap efficiency doesn't fix: a better formula still can't tell you when you're reading the market versus reading your own assumptions back at yourself.

Every volatility model has a window size baked into it somewhere, and that window will eventually show up in your output pretending to be information.

[MoneyLore @Svenchipo]

Fold nine-six offsuit heads-up with 2.5 big blinds and you just lost 253 chips. Not by playing the hand badly. By folding it.

An MIT professor graphed the expected value of shoving all-in with that exact hand against every calling range a human opponent could realistically construct. The line never dipped negative. Not against tight players. Not against loose ones. Not even against someone calling with the mathematically optimal range.

Here's the part that breaks people's brains.

Folding nine-six offsuit in that spot loses you the exact same number of chips as voluntarily calling all-in with three-four against pocket aces. One of those decisions feels like patience. The other feels like a death wish. The math says they're identical.

Will Ma teaches this in MIT's Poker Theory course, Lecture 4. He breaks the expected value of a semi-bluff into two pieces: pot size times your probability of getting a fold, plus your probability of getting called times your equity when that happens. Simple enough to write on one line.

Then he maps real hand rankings onto a logarithmic curve — and gets an R-squared of 98%. That's a cleaner regression than most Wall Street quant models ever produce on real market data.

His students sat in that room and learned, with total mathematical certainty, that shoving garbage hands was a guaranteed long-run profit. Most of them still couldn't pull the trigger later that evening when a real tournament put real chips on the line.

Knowing the expected value was never the hard part. Acting on it while every instinct in your body is screaming "this hand is trash" — that's where almost everyone breaks, in poker and in every market that ever existed.

[ITORRE @ITORRE]

@Svenchipo At 2.5bb you’re not picking pretty hands. You’re just trying to steal what’s already in the pot)

[ITORRE @ITORRE]

@0xjakke The market didn’t fire him. The LPs did

[ITORRE @ITORRE]

@IMundyk agree bro)

[ITORRE @ITORRE]

@zorepad Yeahh. The lecture is just the autopsy)

[ITORRE @ITORRE]

Markowitz won the Nobel in the early 1990s for a machine that most desks quietly refuse to run.

Jake Xia said this at MIT with the optimizer still on the board.

The machine wants three things you do not have: next year’s returns, a true volatility, and a correlation matrix that stays still.
Nudge any input and the “optimal” weights lurch.

The math produces too many answers, so people invent constraints until the software looks decisive.

Then he killed the sacred word.

Volatility is not risk.

A long out-of-the-money call wants the world to shake.

A short call wants silence.

Standard deviation calls both events the same crime.

Sharpe inherits the lie. Sortino splits the tails and still will not tell you the bet size.

Xia’s replacement is uglier and usable.

Expected gain G.
Expected loss L.
Skill is (G − L) / (G + L).
That fraction is a Kelly-style position, not a vanity score.

After that the lecture leaves finance and walks onto the Millennium Bridge.

Pedestrians fall into step.

The bridge begins to sway.

Markets do the same trick: crowding, feedback, power laws, a few super-agents that can turn the whole crowd.

You can print an efficient frontier in color.

The market does not grade the printout.

It grades the size of the next ticket.

[ITORRE @ITORRE]

@Svenchipo That’s the whole indictment.
If a six-month window can rearrange the book, you weren’t allocating. You were fitting yesterday’s noise and calling it a frontier)

[ITORRE @ITORRE]

80% of Europe's banks priced systemic risk on a 256-column spreadsheet.

Kenneth Abbott ran Firm Risk Management at Morgan Stanley before he ever stood in front of MIT students. In 1996, spreadsheets meant Lotus 1-2-3, capped at 256 columns, no tabs to split work across. When a bank's book ran too long, Abbott summed the tail across 2 separate spreadsheets, by hand.

His name for it: an abacus but on the screen.

The formula underneath never grows past one line. One position vector, one covariance matrix, multiply x transpose, sigma, x, and an entire portfolio's variance comes out as a single number.

The sign is the part that trips people up. Quote dollar-yen in yen per dollar, hold the yen side of that trade, and a $100 position gets entered as negative 100, because a weakening yen costs that position money, whatever the size says. A bond works the same way: rising yields hit the owner, so that exposure goes in negative too.

Every entry in the vector tracks which way you bleed, not what you hold.

[ITORRE @ITORRE]

One extra rule cuts Ray Dalio's odds of losing money in a year from 40% to 11%.

Start with one bet. 10% expected return. 10% risk.

Add a second. A third. A fourth. If each new one runs at 60% correlation with the rest, the math barely moves. Stack 1,000 of them at that correlation and total risk falls by roughly 15%, then goes flat.

Drop the correlation to 10%, and the same math turns violent. By the 7th or 8th uncorrelated bet, risk is cut in half. Same return. Half the risk. The return-to-risk ratio just doubled.

Dalio's answer: stop hunting for one great investment. Hunt for 15 to 20 decent ones that refuse to move together.

No single investment beats another by 5x. Nobody is skilled enough to pick one that far ahead of the market. Structure gets you there instead.

Run 15 to 20 genuinely uncorrelated bets and the return-to-risk ratio reaches 1.25. The odds of losing money in any year drop from 40% for one bet to 11% for the portfolio.

You don't need a better bet. You need bets that don't agree.

[ITORRE @ITORRE]

@Igorib1h The guy who saw the housing crash coming learned his stock-picking from this exact book

[ITORRE @ITORRE]

Stan Druckenmiller didn't know how to spell Nvidia 3 months before he bought it.

He ran Duquesne Capital for 29 years, 1981 to 2010, without a single down year. Near 30% annualized.

His edge was never the technology. It was noticing where smart people were already going.

In 2008, he bought Palantir for one reason: it was the company every ambitious kid at Stanford wanted to join. He watched where the talent flowed, then bought the destination.

Early 2022, his young analysts noticed a shift on campus. Stanford students were leaving crypto for AI.

He didn't understand the technology behind it. He trusted the signal anyway.

His partner brought in his own AI contacts from Palo Alto to explain it. Most of it went over Druckenmiller's head. He asked one question: what do I buy.

The answer was Nvidia. He bought enough to feel it - enough to hurt if he was wrong.

2 weeks later, ChatGPT launched. Nobody at the table had mentioned it. He doubled the position.

At a Morgan Stanley macro roundtable, the strategists gave their calls on rates and currencies. A tech analyst cut in: you're all missing something bigger than anything on this table.

Druckenmiller doubled again.

The stock ran from 150 to 390 in 5 months. He told an interviewer he couldn't picture selling for another 2 or 3 years.

The person who understood the technology best, the one who knew 50 times more about AI than he did, sold his own Nvidia within weeks.

Druckenmiller held.

The stock hit 800. He broke his own promise and sold.

5 weeks later, it hit 1,400. He called the feeling sick.

To this day, he says, he couldn't tell you what the company earned that year.

He never learned to spell it. He still made 6 times his money.