Olivia
Online
kynexser
@kynexser
the math is free. the discipline isn't.
Joined July 2026
15 Following
7 Followers
59 Posts
Β Pinned Tweet
https://t.co/j5tWW57y59
FEYNMAN AND DIRAC BOTH FLEW TO NEW ZEALAND TO EXPLAIN THE SAME THEORY. THEY REACHED OPPOSITE CONCLUSIONS
dirac, christchurch, 1975. four lectures. he spends them arguing the fix that made quantum electrodynamics work isn't legitimate mathematics.
feynman, auckland, 1979. four lectures. he won the nobel prize for that exact fix.
four years apart. same country. same theory. two of the men who built it.
feynman didn't even trust this material at first. he refused to debut it at caltech and picked somewhere remote to test whether a normal audience could follow QED at all.
it worked so well the four chapters of his 1985 book are transcripts of these four talks.
part one is him showing why physics keeps eating itself. motion swallowed heat and sound. newton took gravity. maxwell in 1873 proved electricity, magnetism and light are one thing.
your eyes catch a sliver of it. radio, infrared, x-rays, identical, just stretched differently.
then 1929. first quantum theory of electrodynamics written down, and every precise calculation returns infinity or zero. that deadlock held for almost twenty years.
both sets of tapes have been free the whole time.
https://t.co/j5tWW57y59
GRACE HOPPER HANDED OUT PIECES OF WIRE TO EXPLAIN WHY YOUR CODE IS SLOW. THE NSA COULDN'T PLAY THE TAPE FOR 40 YEARS
each wire is about 30 centimeters. that's how far electricity travels in one nanosecond.
she started cutting them because engineers kept asking for satellite links and couldn't understand why the delay was unavoidable.
you can argue with a number. you can't argue with a piece of wire in your hand.
her own line on it: she didn't know what a billion was, and didn't think anyone in washington did either, so how would they know what a billionth is.
the rest of the lecture is 1982 predicting now.
at 8:22 she says the amount of data will grow faster than linearly, demand for instant access will grow with it, and those two are in direct conflict.
she also points out nobody was researching the value of information. no papers, no field.
then she passes around a chip with 8 processors on it, part of a NASA system running 128 by 128 of them. 16,384 processors, in 1982.
the recovery story is its own thing.
someone filed a FOIA request in 2021. NSA said no records exist. pushed further, they admitted the tapes were there, on 1-inch AMPEX reels, and said they aren't required to find obsolete equipment to read their own archive.
the National Archives eventually pulled the footage off both reels. NSA published it in august 2024.
forty-two years in a vault because nobody had the right machine.
https://t.co/j5tWW57y59
@HuySolo_BBW what gets me is that he had no physical reason to believe it. the math said there was a particle with opposite charge, and he trusted the equation over every instinct he had.
QUANTUM PHYSICS PREDICTED ANTIMATTER BEFORE ANYONE HAD SEEN A SINGLE PARTICLE OF IT
1928. Dirac writes down an equation for the electron.
the math spits out solutions with negative energy. everyone else would have called that a mistake and thrown it out.
he doesn't. he takes it literally.
the equation is saying there's another particle. same mass as the electron, opposite charge. nobody had ever observed such a thing.
he publishes it anyway.
four years later carl anderson finds it in cosmic ray tracks. exactly what the math described.
that's the whole story of dirac. he trusted the equation over his own intuition and the universe agreed with him.
this is him lecturing in christchurch, new zealand, in 1975. he's in his seventies. four lectures, walking through quantum mechanics himself.
the tapes were considered lost for decades before they surfaced online.
the picture breaks up in places. the audio on this upload was cleaned up.
it's one of very few recordings of a founder of quantum mechanics explaining the thing he built.
https://t.co/j5tWW57y59
@explosss1ve thats crazy, how mcp changed 3d industry
FEED BIFROST RAW FOOTAGE OF A JOB SITE AND IT HANDS BACK A FULL 3D SIMULATION YOUR ROBOT CAN FAIL IN SAFELY
point a camera at a real environment. bifrost turns the footage into a photorealistic, physically accurate 3d world, gpu-accelerated, in about 30 minutes.
no manual modeling. you don't touch a 3d tool.
describe the scenario in python or just tell a coding agent what you want:
world.spawn("container_ship", quantity=12)
world.spawn("buoy", quantity=30, scatter=True)
world.ocean(sea_state=4)
that's a storm-tossed harbor generated for testing, not filmed.
the actual point isn't the pretty render. it's running a robot's policy against thousands of variations, weather, lighting, rare hazards, and finding exactly where it breaks before it breaks near a person.
backed by airbus ventures, already used on maritime and aerospace problems where a real-world failure test isn't something you get to run twice.
A FLOATING MUSCLE CAR IN THE CLOUDS LOOKS EASY. THE CLOUDS ARE THE HARD PART
the car is a standard hard-surface model. easy either way.
the volume around it is what eats hours.
clouds like this are a density texture inside a volume domain. getting them soft instead of smoke-like takes real tuning.
kimi k3 through blender mcp writes the python to set that domain up and place the car inside it.
the headlights are the harder part. they need to cast light into the fog, not just glow, or the effect falls apart.
that's light interacting with a volume. slow by hand, easy to describe to a model that knows the settings.
water droplets on the body are a small bump pass, easy once everything else is in place.
what it can't decide is whether the mood actually reads as cinematic or just murky. that's still a person's call.
usual note: raw python straight into blender, zero sandbox. keep it in a throwaway file.
A FLOATING MUSCLE CAR IN THE CLOUDS LOOKS EASY. THE CLOUDS ARE THE HARD PART
the car is a standard hard-surface model. easy either way.
the volume around it is what eats hours.
clouds like this are a density texture inside a volume domain. getting them soft instead of smoke-like takes real tuning.
kimi k3 through blender mcp writes the python to set that domain up and place the car inside it.
the headlights are the harder part. they need to cast light into the fog, not just glow, or the effect falls apart.
that's light interacting with a volume. slow by hand, easy to describe to a model that knows the settings.
water droplets on the body are a small bump pass, easy once everything else is in place.
what it can't decide is whether the mood actually reads as cinematic or just murky. that's still a person's call.
usual note: raw python straight into blender, zero sandbox. keep it in a throwaway file.
https://t.co/YDY1uveJ2m
KIMI K3 BUILT A LIMINAL SPACE MALL IN BLENDER, THEN FLOODED IT FROM THE CEILING
the setup: an empty two-floor mall, neon pink and blue, mannequins behind fencing, wet floor signs out before any water even shows up.
kimi k3 wrote the python through blender mcp. blender executed it directly, nothing modeled by hand.
the escalator ride up is a simple camera animation.
the flood on floor one is a fluid sim, keyed to rise right as the camera clears the top step.
the hard part wasn't the water itself.
it was the black hole that opens in the ceiling and pours water upward before gravity flips and the whole space goes under.
that's not one effect. it's a fluid sim, a particle system for the mannequins and debris tumbling through it, and a force field doing the inversion, all timed to one continuous camera path.
worth saying every time: this is raw code running straight into blender with zero sandbox. throwaway file, not your real project.
the last shot is the same escalator again, riding up into pure black.
nothing dramatic in the code for that one. just an empty render and a light that never turns on.
https://t.co/YDY1uveJ2m
a nobel laureate spent an hour on camera arguing the universe itself doesn't know what happens next.
cornell, 1964. bbc filmed it. feynman was forty-six, two years before his nobel prize.
he won it in 1965 for reworking quantum electrodynamics, the math behind how light and matter interact.
fire one electron at two slits.
nobody alive can tell you which slit it goes through. only the odds of where it lands.
einstein hated this. he said god does not play dice.
feynman's actual answer, on camera: "i think i can safely say that nobody understands quantum mechanics."
not a joke. the real thesis.
nature isn't hiding the outcome from you. it doesn't have one yet.
that's basically what a casino is too. the house doesn't know which number hits. it just prices the odds correctly and lets everyone else keep chasing a certainty that was never on the table.
the lecture's been free since cornell posted the bbc tapes in 2015.
nobody's figured out how to give away the part where you actually act on odds instead of waiting for a certainty that isn't coming.
https://t.co/JzbEJ0onc0
mit's most famous programming lecture opens with the professor trashing the name of his own field.
β hal abelson starts 6.001 by saying "computer science is a terrible name for this business." turns out it's not really a science, and it was never really about computers.
β filmed in july 1986 for hewlett-packard employees, all 20 lectures have been free on mit ocw ever since.
β the textbook, sicp, the "wizard book," has also been free online for decades.
β the actual subject was controlling complexity. writing instructions precise enough for a machine to run and clean enough for a person to still follow ten years later.
anyone can write code that runs once. the part that actually gets paid for is the part mit filmed for free almost 40 years ago. nobody's figured out how to hand you the taste yet.
https://t.co/JzbEJ0onc0
Kimi CEO Zhilin Yang says the most valuable skill he learned before building one of the world's strongest AI models wasn't programming.
It was mathematics.
Years before Kimi started competing with models like Claude, Yang was already known for attacking the hardest math and optimization problems he could find instead of chasing grades. One of his professors described him as "the most talented student I've seen in years."
While other students were trying to finish assignments, he was building graph algorithms that reduced labeling costs by more than 90%, work that later influenced real systems used in industry.
He has also talked about spending days stuck on a single proof or equation, only to solve it while walking home with music in his headphones.
People see the model.
They rarely see the years spent learning how to think mathematically before writing a single line of AI code.
The code built Kimi.
The math made the code possible.
https://t.co/JzbEJ0onc0
One of the best lectures about AI wasn't about AI at all.
It was recorded more than 40 years ago.
Long before GPUs, transformers, or ChatGPT, one computer scientist explained why hardware keeps getting faster while software keeps getting harder.
The recording sat in an archive for decades.
Now it's free.
You don't watch it for history.
You watch it because half the problems AI engineers are arguing about today are already in that lecture.
Some videos get old.
This one just caught up.
https://t.co/JzbEJ0onc0
TERENCE TAO SPENT HIS LIFE ON THE ONE SKILL HEDGE FUNDS PAY $500K FOR, AND MOST PEOPLE GET IT EXACTLY BACKWARDS
telling a real pattern from noise. that's the whole job those firms are buying.
Tao's lifelong theme is that almost nothing is purely one thing.
there's perfect structure, like a clock. perfect randomness, like a coin. everything real is a mix of the two.
the work is separating them.
the primes are the perfect test. they look scattered, lawless, random.
yet Tao and Ben Green proved they contain arbitrarily long evenly-spaced runs. order was hiding inside the chaos the whole time.
that's signal detection, minus the marketing. and he gave the lecture at UCLA for free.
here's the part no fund can buy.
the math is public. the judgment to know when a pattern is real, and when your own eyes invented it, isn't.
that judgment is the entire job. the lecture is free. the years it takes to trust it are not.
https://t.co/JzbEJ0onc0
Richard Feynman explained the biggest AI problem in 1964.
Most people skipped the lecture.
Every major AI lab is trying to make language models reason.
Feynman spent an hour explaining why reasoning breaks the moment you stop respecting mathematics.
No GPUs.
No transformers.
Just chalk and a blackboard.
Halfway through the lecture, he rebuilds Kepler's laws from first principles so clearly it almost feels unfair.
You start watching a physics lecture.
You end up understanding why AI still struggles with reasoning in 2026.
The video has fewer views than random tech demos.
It probably shouldn't.
https://t.co/JzbEJ0onc0
https://t.co/JzbEJ0onc0
WILLIAM SHARPE WON A NOBEL FOR ONE IDEA MOST TRADERS STILL IGNORE, RETURN MEANS NOTHING WITHOUT THE RISK THAT BOUGHT IT
lose 50% and you're not down 50%. you now need a 100% gain just to get back to zero. the drawdown isn't symmetric, and that asymmetry is the whole game.
Sharpe built the ratio Wall Street still runs on. it asks the question every screenshot hides.
not "how much did it make." "how much risk did it take to make it."
a 70% win rate means nothing on its own. win 7 trades, lose 3, and still bleed money if the winners pay $1 and the losers cost $3.
variance controls the path. a real edge can look broken for months. a broken strategy can look brilliant right up until the streak that ends it.
and the perfect backtest is often just the luckiest survivor out of 10,000 versions that died.
the dangerous question is "how often am I right." the useful one is "can this survive its worst losing run."
return gets the attention. risk-adjusted return tells you whether the result was even worth surviving for. Sharpe proved that half a century ago and the screenshots still haven't caught up.
https://t.co/0TRQle9rd2
MIT STUDENTS BEAT THE BIGGEST CASINOS FOR MILLIONS USING MATH FROM A BOOK ANYONE COULD BUY
no cheating. no marked cards. just a $2 paperback and the discipline almost nobody has.
it started with Edward Thorp. in 1962 a math professor proved that a blackjack player tracking the cards can flip the odds in his own favor. most people read it as a party trick. a few MIT students read it as a business plan.
they didn't organize like gamblers. they organized like a hedge fund.
> outside investors, a pooled bankroll
> spotters at the tables betting tiny, silently counting, waiting
> a "big player" who strolled over and bet huge only when the deck turned hot, then walked
it looked like luck. it was a machine.
the casinos couldn't call it cheating, because it wasn't. so they hired detectives, built face databases, and banned anyone they could identify. the house had met people running the house's own math against it.
but the edge was never the counting. the edge was in the book, and the book was public.
the moat was the part nobody copies. bet the minimum until the math is actually on your side, then bet the maximum, and never break that rule for a feeling.
a one-percent edge plus a bankroll deep enough to survive the losing runs isn't gambling. it's being the casino. several of them walked off the floor straight onto Wall Street, where the same idea is worth far more than millions.
the math was free. the discipline was the whole fortune.
https://t.co/0TRQle9rd2
WARREN BUFFETT IS WORTH $140 BILLION AND STILL LIVES IN THE HOUSE HE BOUGHT FOR $31,500 IN 1958
it's the same five-bedroom place in Omaha. worth about $1.4 million now. he never moved.
he still does his own taxes. drives himself to work. the salary he draws from Berkshire has been $100,000 for decades, unchanged.
the frugality isn't the lesson though. it's the tell.
the man who compounds better than anyone alive spends nothing on signaling wealth. every dollar he doesn't burn on a bigger house stays in the machine that makes more dollars.
his edge was never a stock pick. it was refusing to let the money change how he lives, so the money could keep working.
and the final move is the sharpest one. 99% of that $140 billion is pledged away. he isn't handing his kids a fortune, because he thinks a head start that big ruins the person who gets it.
most people chase the number to spend it. he chased it to give it away. the discipline that built the pile is the same discipline that lets him let go of it.
https://t.co/0TRQle9rd2
PERSI DIACONIS SPENT 40 YEARS PROVING WHEN A SIMULATION IS LYING TO YOU, AND WALL STREET NEVER LISTENED
every quant fund since Simons has chased the same answer. more data, more compute, more simulations.
Diaconis, the Stanford authority on this exact question, says that's the wrong knob.
his whole field is mixing times. how long a random process has to run before its output actually means anything.
the uncomfortable finding: most simulations run in practice never prove they've converged. they just stop when the deadline hits, and people read the result as truth.
a chain that hasn't mixed gives you a confident, precise, completely wrong number. no GPU fixes that. throwing compute at an unconverged simulation just gets you to the wrong answer faster.
Simons required a proof before deploying a strategy. that's the part nobody copied.
more compute makes a wrong model wrong at scale. the thing that separates Renaissance was never the horsepower. it was refusing to trust a number they couldn't prove.
https://t.co/0TRQle9rd2
Last Seen Users on Sotwe
Most Popular Users
Elon Musk
@elonmusk
241.1M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
112.3M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.1M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.3M followers
KATY PERRY
@katyperry
88.7M followers
Taylor Swift
@taylorswift13
82.5M followers
Lady Gaga
@ladygaga
74.1M followers
Virat Kohli
@imvkohli
71.4M followers
Kim Kardashian
@kimkardashian
70.3M followers
YouTube
@youtube
68.7M followers
Bill Gates
@billgates
64.4M followers
Neymar Jr
@neymarjr
64.3M followers
The Ellen Show
@theellenshow
62.4M followers
CNN
@cnn
61.8M followers
Selena Gomez
@selenagomez
61.7M followers
X
@x
60.8M followers