Benny

404 is just the beginning • money math for the AI era

[Benny @0xBenjamine]

https://t.co/VIBl2zlsSS

[Benny @0xBenjamine]

Two Nobel Prize winners stacked a $1.25 trillion bet on just $4.7 billion of their own money, lost $4.6 billion in under four months, and when Warren Buffett offered $250 million for the wreck they turned him down

The fund was Long-Term Capital Management, run by Salomon Brothers legend John Meriwether with Myron Scholes and Robert Merton, the Black-Scholes guys, and it borrowed $124.5 billion while printing 40%+ two years in a row

Then Russia defaulted in August 1998, $1.85 billion vanished in one month, and the New York Fed had to drag Goldman Sachs, JPMorgan, Merrill Lynch and 11 more banks into one room to put up $3.6 billion before it took Wall Street down with it

Under 5 minutes of math to watch before you put leverage on any trade or let an AI bot do it for you, so bookmark it and watch today because Bear Stearns refused to chip in and 10 years later Bear Stearns was gone too

[Roan @RohOnChain]

https://t.co/ylkXptDWTw

[Benny @0xBenjamine]

@sama hard to avoid when the industry pitch is that these models will cure disease and solve physics, and a lot of people believe it

[Benny @0xBenjamine]

@mathemetica works for sin. e^(-1/x²) is smooth too, but every derivative at 0 is 0, so its taylor series there is just 0 while the function isn't

[Benny @0xBenjamine]

@Math_files and the fundamental theorem says they undo each other. take the slope of the area under the curve up to x and you get the curve back

[Benny @0xBenjamine]

@matheorems particle detectors measure momentum this way. a charged track curves in the magnetic field and the fitted radius gives you p = qBr

[Benny @0xBenjamine]

@NightSkyToday in quantum eraser experiments, if you erase the which-path info later, the interference shows up again once you sort the data

[Benny @0xBenjamine]

@konstructivizm proxima is the nearest dot on this map and voyager 1 would still need around 75,000 years to get there

[Benny @0xBenjamine]

@konstructivizm if horizon-agn can't produce the low-mass black holes in this sample, how much of the merger story was built on simulations like it?

[Benny @0xBenjamine]

@skdh their 2018 nature paper on majorana signals was retracted in 2021, and the 2025 one came with an editor's note saying it didn't prove majorana modes. glad someone outside microsoft finally gets to test the chip

[Benny @0xBenjamine]

Jensen Huang lost about $20 billion of his own fortune in a single day as Nvidia dropped $589 billion on January 27, 2025, the biggest one-day wipeout for any company in US stock market history, the world's richest people including Larry Ellison lost $108 billion combined, Marc Andreessen called it AI's Sputnik moment, and the trigger was a PDF from a small Chinese lab called DeepSeek claiming it trained a frontier model for about $5.6 million

OpenAI, Oracle and SoftBank had just announced a $500 billion Stargate plan, so a $5.6 million model looked like the end of the chip boom, but that number was one line of math, 2,048 Nvidia H800 GPUs running for 2.788 million GPU hours at an assumed $2 an hour, and the same DeepSeek paragraph admitted it left out all prior research and experiments, SemiAnalysis estimated DeepSeek's real hardware spending at well over $500 million, about a hundred times the headline that crashed the market

The model has 671 billion parameters split into 256 experts, and a tiny router wakes only 8 of them per word, so just 37 billion ever do any work, which is why it really was cheap to run, and Microsoft's Satya Nadella answered the panic that same day with Jevons paradox, cheaper AI means the world burns way more of it, by July 2025 Nvidia became the first $4 trillion company and on October 29 it passed $5 trillion

5 minutes of math to watch before the next AI headline makes you panic sell, so bookmark it and watch today because one sentence most headlines skipped cost investors $589 billion in a day and Nvidia almost doubled from there

[Matt Shumer @mattshumer_]

https://t.co/ivXRKXJvQg

[Benny @0xBenjamine]

@Math_files today that letter lands in spam. and a clerk with no affiliation would need an endorser before arxiv takes his paper

[Benny @0xBenjamine]

@PeterSchiff tuesday it was 5.24 on the 10y and 5.58 on the 30y. now 5.27 and 5.63, the gap between them went from 34 to 36bp

[Benny @0xBenjamine]

@Math_files ran it with gmpy2 and it holds, the first prime is at n = 616,980 and it's a 70 digit number. the whole check took under a second

[Benny @0xBenjamine]

@thsottiaux long sol runs in codex were crawling for me yesterday. if it's really 2x by morning i'm rerunning everything

[Benny @0xBenjamine]

@sundarpichai 19.6% on harvey's legal benchmark, triple anyone else, and it still fails four cases out of five. lawyers have a couple more releases left

[Benny @0xBenjamine]

@JensenHuang if the "meet regularly to establish standards" line at the bottom turns into one shared eval standard, that helps everyone building on top of these models way more than five separate policies

[Benny @0xBenjamine]

@michaeljburry ballard still had years of input before words, riding around with his dad, watching trees grow, staring at a map. feels like an argument against text-only training more than against language. would models trained on video and robot data pass his test?

[Benny @0xBenjamine]

@paulg the first time it's just one risk out of a hundred. once it's cost you something, you have an actual memory to match against and you spot it way earlier

[Benny @0xBenjamine]

@wandermist ok "nobody noticed" is a stretch, n-gram models ran on this for decades. the surprise was how far it goes once you scale it to billions of parameters