Codi

Vibe Coder | Artificial intelligence | Match·Edge·Models

dm always open

[Codi @0xcod1]

The oldest prank in the book still works every single time: someone showering outside, and a friend quietly dumping extra shampoo on their head from above.

The person washing has no idea it's happening. All they know is the shampoo will not rinse out, no matter how long they stand there, more suds keep appearing out of nowhere.

It only works because they can't see above them and have zero reason to suspect sabotage mid-shower. By the time they figure it out, they've already spent five minutes confused why one bottle's worth of shampoo turned into an endless supply.

Simple, harmless, and somehow still funny every single time someone tries it.

Follow and save if you want more like this.

[clayne @0xclayn]

https://t.co/N7F0ebWYt9

[Codi @0xcod1]

A model six times smaller than its own predecessor just made that predecessor look outdated.

1. DeepSeek released V4-Flash-0731 in early August 2026, a 284 billion parameter model, small enough to fit on a single system's GPU memory in the right format.

2. Its full-size sibling, DeepSeek V4-Pro, runs at 1.6 trillion parameters, nearly six times larger.

3. On Artificial Analysis's own intelligence leaderboard, the smaller Flash model outperforms its own much bigger Pro sibling by close to 14 percent.

4. That's not supposed to happen under the old assumption that bigger parameter counts reliably mean smarter models, the entire scaling story most of the AI industry has been selling investors for years.

5. Enterprises can now run something close to frontier-level intelligence on hardware that used to only handle mid-tier models, cutting the cost of serious AI deployment by a wide margin.

The lesson buried in one model release: the next leap in AI might not come from a bigger model at all, it might come from figuring out how to stop needing one.

[clayne @0xclayn]

https://t.co/4GmVaB61Yh

[Codi @0xcod1]

Anthropic just gave every startup running on its models a way to cut their AI bill without switching providers.

Claude Opus 5 shipped July 24, 2026 with an effort setting, low, high, or max, letting a company control how much compute a task gets instead of paying full price every time.

That matters because AI-native startups are already spending 40-50% of revenue on inference, versus 5-25% for normal SaaS hosting. This dial trims cost on the easy requests while still paying full price where it's actually needed.

Same API, same integration. Just a way to stop overpaying for the easy 80% of what the product does.

Follow and save if you want more like this.

[clayne @0xclayn]

https://t.co/4GmVaB61Yh

[Codi @0xcod1]

A cartoon shopping assistant with no body, no camera, and no face outside a screen earned more from sponsored posts last year than most human influencers ever will.

Her name is Lu, the virtual mascot of Brazilian retailer Magazine Luiza, and she's been "alive" since 2003, long before anyone called this an AI influencer. In 2024 she pulled in an estimated $2.5 million across 74 sponsored collaborations, which works out to roughly $34,000 per post.

The engagement numbers explain why brands keep paying. Virtual influencers like her average close to 5.7 percent engagement, about three times higher than the roughly 1.9 percent human creators typically pull. When she wore specific outfits in music videos with Brazilian artists, those exact outfits sold out on the retailer's own app within days.

She's also not an outlier anymore, just the biggest example of a fast-growing category. The global virtual influencer market hit an estimated $11.74 billion in 2026, and 73 percent of surveyed brands now run at least one virtual creator program, up from 60 percent just a year earlier.

Nobody has to manage a camera crew, a schedule, or a person's actual life to run a campaign like this. They just have to keep one consistent character on-brand, indefinitely, and let the engagement numbers do the rest.

[clayne @0xclayn]

https://t.co/4GmVaB61Yh

[Codi @0xcod1]

@0xclayn yeahh))

[Codi @0xcod1]

A retired long-haul trucker with zero coding background is quietly clearing six figures a year running one small AI model nobody at his old company would ever think to build.

He spent thirty years driving freight routes and knew exactly which paperwork slowed dispatchers down the most, mismatched load numbers, missing weight tickets, drivers submitting photos of paperwork that didn't match what the system expected. Nobody inside the big logistics companies had time to fix a problem that small.

He didn't build anything from scratch. He fine-tuned an existing open model on thousands of real dispatch documents, mostly ones he'd collected and anonymized from his own years on the road, until it could catch a mismatched load number faster than a tired dispatcher scanning it at 2am.

He licenses it to small regional trucking companies too small to build their own tooling, priced per truck per month instead of a big enterprise contract nobody that size could afford. Thirty-one companies signed on within the first year.

The income isn't glamorous, no press coverage, no funding round, just steady monthly payments from companies who'd rather pay a small recurring fee than hire someone to fix the same problem by hand.

He spent three decades learning exactly which small paperwork mistake cost a trucking company the most money. That knowledge turned out to be worth more than any coding skill he never had.

[clayne @0xclayn]

https://t.co/4GmVaB61Yh

[Codi @0xcod1]

@0xcodix wow
perfect

[Codi @0xcod1]

When this film was made in 1990, buying one share of stock still meant a person on a physical floor was shouting your order out loud.

Your order left the broker's office, hit the exchange floor, and from there a specialist, one single person assigned to that one stock, physically matched it against someone else's order standing a few feet away. No screen decided the price. A human did, in real time, in a crowd.

The exchange had just started leaning on SuperDOT, a system quietly routing smaller orders straight to that specialist's post electronically, skipping a human floor broker in the middle. In 1990 it could already push through close to a billion shares a day and return a fill in under a minute, which felt instant at the time.

Traders on the floor were still talking over each other constantly, so much of the real communication ran on hand signals instead, flicks and gestures precise enough to mean an exact price and quantity, readable across a room too loud to hear anyone.

None of that fully disappeared until 2006, when the NYSE finally merged with an all-electronic exchange and open outcry stopped being the default. For sixteen years after this film was shot, a hand signal across a crowded floor could still move real money before a keyboard ever touched it.

The film is showing you the exact system in its final stretch, a market that still ran, in large part, on people standing close enough to shout at each other.

[ooaes @21xooaes]

https://t.co/IE4DRgPuYV

[Codi @0xcod1]

A Yale economics professor used to run mortgage trading on Wall Street. Then he built the theory that explains why his own former industry blew up the global economy.

His name is John Geanakoplos. He has a PhD in economics from Harvard, finished in 1980, and by most measures should have stayed in academia writing papers nobody outside a seminar room would read.

Instead he became Managing Director of Fixed Income Research at Kidder Peabody, then one of six founding partners of Ellington Capital Management, a hedge fund built around mortgage securities. He was inside the mortgage machine years before it became the center of the 2008 financial crisis.

Standard financial theory at the time leaned on the efficient markets hypothesis, the idea that asset prices already reflect all available information, so bubbles shouldn't really be possible in any lasting way. That theory had a rough few years right around 2007 to 2009.

Geanakoplos's answer to that failure is something he calls the leverage cycle, the idea that how much borrowed money is sloshing through the system, not just prices, drives the boom and the crash. When lenders let buyers put down less and less of their own money, prices get pushed up on borrowed air. When that credit dries up, the same mechanism drives the crash on the way down.

He opens the very first lecture of this course by admitting the theory he's about to teach has already been discredited by the crisis he lived through professionally, then spends the rest of the semester explaining exactly why, and what to build instead.

Most economists write theories about crises they only ever watched from a distance. This one wrote his after spending years inside the machine that caused it.

[Tigerflow @tigerfl0w]

https://t.co/Ye9wCJ9Jiv

[Codi @0xcod1]

An MIT professor spent one semester teaching the theory that markets are perfectly rational, then spent the next twenty years proving his own course wrong.

His name is Andrew Lo. He filmed this exact finance course in the fall of 2008, months after the financial crisis had already made a mockery of the idea that markets always price things correctly.

Lo was born in Hong Kong, moved to Taiwan as a child, got his economics degree from Yale, then a PhD from Harvard by age 24. By the time he was filming this lecture, he wasn't just a professor, he was already running his own quant hedge fund, AlphaSimplex, trading on the exact market inefficiencies his own textbook chapter said shouldn't exist.

The course covers the Efficient Markets Hypothesis, the backbone idea that stock prices already reflect all available information, so nobody can consistently beat the market through skill alone. It's still taught in nearly every MBA program on Earth.

Lo doesn't fully believe it. He spent years building an alternative he calls the Adaptive Markets Hypothesis, treating financial markets less like a perfectly rational machine and more like an ecosystem, where investors compete, adapt, and make the same mistakes evolution keeps making everywhere else in nature.

Wall Street pays quant analysts six figure salaries to build models around exactly this idea, that markets are adaptive, not perfectly efficient. Lo taught the free, public version of that insight to a room full of twenty year olds and put the whole lecture online for anyone to watch.

The professor teaching you why markets are supposedly efficient may be the same man quietly getting rich off proving they aren't.

[Lumen @lumenxbt]

https://t.co/3hWKMIBJcw

[Codi @0xcod1]

The word algebra literally means "reunion of broken parts," and it comes from a 9th century book about debt.

𝗠𝗬𝗧𝗛: algebra was invented as an abstract school subject to torture teenagers with x's and y's.
𝗥𝗘𝗔𝗟𝗜𝗧𝗬: it was built to solve extremely practical problems, inheritance disputes, land division, trade debts, in Baghdad around 820 AD, by a scholar named al-Khwarizmi.

𝗠𝗬𝗧𝗛: the rules for solving an equation have always looked the way they do in a textbook today.
𝗥𝗘𝗔𝗟𝗜𝗧𝗬: al-Khwarizmi's original method was written out entirely in words, no symbols at all, describing step by step how to 'restore' an equation by moving pieces from one side to the other until it balanced, which is exactly what algebra still means at its core.

𝗠𝗬𝗧𝗛: algebra is just a harder version of arithmetic.
𝗥𝗘𝗔𝗟𝗜𝗧𝗬: arithmetic tells you what a specific number is. Algebra tells you what has to be true no matter which number you plug in, which is the actual reason it can describe a bridge, a rocket trajectory, or a loan payment using the exact same handful of rules.

Even the modern word 'algorithm' traces back to the same man. Two of the most important ideas in modern computing came out of one 9th century book about splitting up debt fairly.

[clayne @0xclayn]

https://t.co/4OzBMnqXNP

[Codi @0xcod1]

AN 88 YEAR OLD PROFESSOR TAUGHT ONE COURSE FOR 61 YEARS, THEN GOT A STANDING OVATION FOR SOLVING A MATRIX

Gilbert Strang walked into MIT's 18.06 Linear Algebra classroom for the very last time in May 2023, closing out a career that started in 1962.
He didn't give a farewell speech. He taught elimination and rank, exactly like any other Monday.

- His free MIT OpenCourseWare version of this same course has pulled in more than 20 million views, almost entirely from people who never set foot on MIT's campus

- Strang is the reason an entire generation of self-taught engineers and machine learning practitioners can explain what a matrix rank actually means

- The final class wasn't just students, it was packed with colleagues and former students who came back just to watch one more lecture

- The room gave him a standing ovation not for a big finale, but after he casually finished working through a rank problem on the board

- His textbook, Introduction to Linear Algebra, is still the one most of those same online students are quietly following along with

Sixty-one years, and the last thing he taught the world was still just how to find a nonzero solution.

[clayne @0xclayn]

https://t.co/4OzBMnqXNP

[Codi @0xcod1]

ON THE FIRST DAY OF CALCULUS AT OXFORD, THE LECTURER TELLS YOU IT ALREADY CAUSED A WAR

Not a real war. Worse, in academic terms: a decades-long international feud over who actually invented the subject you're about to be tested on.

- Newton had the ideas first, sometime in the 1660s, but sat on them for years without publishing

- Leibniz published his own version first, cleaner notation and all, which is why we still write dy/dx instead of Newton's clunky dot notation

- The Royal Society, which Newton happened to run, launched an inquiry into the dispute and, unsurprisingly, ruled in Newton's favor

- British mathematicians kept using Newton's notation out of loyalty for nearly a century afterward, and fell noticeably behind the continent as a result

- The actual rigorous definition of a limit, the epsilon-delta version every first year eventually meets, wasn't nailed down until the 1800s, more than a hundred years after calculus was already being used to launch cannonballs and predict tides

The subject started a fight between two geniuses before it ever showed up on a syllabus.

[clayne @0xclayn]

https://t.co/4OzBMnqXNP

[Codi @0xcod1]

Mitochondria get introduced in every biology class as "the powerhouse of the cell," a phrase so overused it became a joke, and the joke buried the actually strange part.

𝗜𝘁 𝗵𝗮𝘀 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗗𝗡𝗔. 𝗜𝘁 𝗱𝗶𝘃𝗶𝗱𝗲𝘀 𝗼𝗻 𝗶𝘁𝘀 𝗼𝘄𝗻 𝘀𝗰𝗵𝗲𝗱𝘂𝗹𝗲. 𝗜𝘁 𝘄𝗮𝘀 𝗻𝗲𝘃𝗲𝗿 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘆𝗼𝘂𝗿𝘀. Every mitochondrion inside every one of your cells carries its own small, circular loop of DNA, completely separate from the DNA in your cell's nucleus, and it copies that DNA and divides on a timeline your own cells don't control. That's not how a normal piece of cellular machinery behaves. That's how a separate organism behaves.

The explanation biologists eventually settled on is genuinely wild: somewhere around 1.5 to 2 billion years ago, one primitive cell engulfed a free-living bacterium, and instead of digesting it, kept it alive inside itself, permanently. The bacterium got a stable home and a steady supply of nutrients. The host cell got a vastly more efficient way to generate energy. Neither one could easily go back to living alone after that, and the arrangement has been running, unbroken, in a straight line down to the cell currently keeping your heart beating.

Plants carry a second version of the same story. Chloroplasts, the structures that let them photosynthesize, are a separate ancient bacterium that got absorbed the same way, on a completely independent occasion.

So the sentence "the mitochondria is the powerhouse of the cell" was technically true and almost entirely missing the point. You are not one organism. You are a very old, very successful merger, and part of what's keeping you alive right now used to be something else's ancestor, with its own family tree, that just never left the building.

Sneary draws the whole timeline on the board, cell by cell. It is in the lecture.

[clayne @0xclayn]

https://t.co/j85C6aHKUs

[Codi @0xcod1]

Find a watch lying in a field and you instantly assume someone made it. Gears don't assemble themselves by accident, so something with a plan must have put them there.

For centuries that argument was used, seriously, as proof of a designer behind every living thing. An eye is far more intricate than a watch, the reasoning went, so it needs a far more capable designer behind it.

1. The argument has one real weak point, and it's not philosophical, it's mechanical. A watch doesn't make copies of itself. It doesn't pass small, accidental variations on to the next watch. Nothing about a watch competes with other watches to see which design works better. Life does all three, constantly, which changes the entire math.

2. Once copying, small variation, and competition are all running at once, you don't need a designer holding a blueprint. You need only one thing: a version that works slightly better than the version next to it, surviving slightly more often, passing that slight edge to its own copies. Repeat that for an absurd number of generations and something that looks impossibly engineered falls out the other end, without anyone ever holding the plan.

3. That process has been run so many times, in so many separate lineages, that eyes alone evolved independently dozens of separate times across the animal kingdom, arriving at wildly different designs, an octopus eye, an insect's compound eye, a human eye, each one a different solution to the exact same problem of catching light. If one master designer were behind all of it, you'd expect one master design, reused. Instead you get dozens of different, independently improvised answers, which is exactly what blind trial and error produces and exactly what a single planner wouldn't.

The watch on the ground still needed a watchmaker. The eye in your skull didn't need one at all. It needed time, competition, and an almost unimaginable number of small, unplanned mistakes that happened to work.

[clayne @0xclayn]

https://t.co/j85C6aHKUs

[Codi @0xcod1]

@0xcodix yoo good post man

[Codi @0xcod1]

Your genome can lose a piece and nothing happens. Most of the time, biology doesn't even notice.

Delete a gene in yeast and roughly 80% of the time, nothing visibly changes. Same organism, same conditions, missing instructions, business as usual. Biology has spent decades building the assumption that DNA is a program and the body is just running it. That assumption keeps quietly failing the moment anyone actually runs the experiment.

What's really happening is that living systems are absurdly good at absorbing damage. Knock out one gene and a backup mechanism picks up the slack. Knock out two, and sometimes a third has to fail before anything at the level of the whole organism finally shows a crack. The genome isn't a dictator handing down orders. It behaves more like one voice in a very large, very redundant conversation.

That changes what a gene even is. One definition ties it to a stretch of DNA. An older one ties it to a trait you can actually observe. Those two definitions don't line up, and which one someone uses quietly decides whether they even believe DNA is fully in charge.

𝘈 𝘤𝘦𝘭𝘭 𝘥𝘰𝘦𝘴𝘯'𝘵 𝘳𝘶𝘯 𝘰𝘯 𝘢 𝘴𝘤𝘳𝘪𝘱𝘵. 𝘐𝘵 𝘳𝘶𝘯𝘴 𝘰𝘯 𝘢 𝘯𝘦𝘨𝘰𝘵𝘪𝘢𝘵𝘪𝘰𝘯 𝘣𝘦𝘵𝘸𝘦𝘦𝘯 𝘵𝘩𝘰𝘶𝘴𝘢𝘯𝘥𝘴 𝘰𝘧 𝘱𝘢𝘳𝘵𝘴, 𝘢𝘯𝘥 𝘢𝘭𝘮𝘰𝘴𝘵 𝘯𝘰𝘯𝘦 𝘰𝘧 𝘵𝘩𝘦𝘮 𝘤𝘢𝘯 𝘣𝘦 𝘪𝘴𝘰𝘭𝘢𝘵𝘦𝘥 𝘢𝘴 𝘵𝘩𝘦 𝘴𝘪𝘯𝘨𝘭𝘦 𝘤𝘢𝘶𝘴𝘦 𝘰𝘧 𝘢𝘯𝘺𝘵𝘩𝘪𝘯𝘨.

[clayne @0xclayn]

https://t.co/j85C6aHKUs

[Codi @0xcod1]

@0xcodix omg its crazy man

[Codi @0xcod1]

A startup started interviewing its own customers on camera.

Real customers. Real answers. Real case studies.

The person asking the questions isn't real.

They needed a consistent interviewer, someone customers would recognize across episodes, without pulling an actual employee off their real job for two hours a week.

So they generated one instead.

Claude writes the interview questions from the customer's actual account data, so nothing feels generic.

A voice model asks them out loud. A render tool keeps her face and tone the same, episode after episode.

The customer being interviewed is real, on a real call, answering real questions.

The person nodding along on the other side of the screen was rendered after the call was already over.

Nobody being interviewed has asked who she actually is.

They just answer the question and move on.

The series has become the startup's best-performing content type by a wide margin.

Not because the questions are smarter.

Because it finally happens every single week, instead of whenever someone on the team had two free hours.

[Codi @0xcod1]

@0xclayn Me too

[Codi @0xcod1]

A startup started interviewing its own customers on camera.

Real customers. Real answers. Real case studies.

The person asking the questions isn't real.

They needed a consistent interviewer, someone customers would recognize across episodes, without pulling an actual employee off their real job for two hours a week.

So they generated one instead.

Claude writes the interview questions from the customer's actual account data, so nothing feels generic.

A voice model asks them out loud. A render tool keeps her face and tone the same, episode after episode.

The customer being interviewed is real, on a real call, answering real questions.

The person nodding along on the other side of the screen was rendered after the call was already over.

Nobody being interviewed has asked who she actually is.

They just answer the question and move on.

The series has become the startup's best-performing content type by a wide margin.

Not because the questions are smarter.

Because it finally happens every single week, instead of whenever someone on the team had two free hours.

[clayne @0xclayn]

https://t.co/4GmVaB61Yh