Gemincho

i describe it and it exists

[Ember @Emberxbt]

In 2015 MIT's business school ran a graduate course on winning at poker, filmed all eight lectures and posted them online for zero dollars. The first lesson is the opposite of what you would guess.

It does not teach you how to win big. It teaches you how to lose slowly.

MIT Sloan charges $91,892 a year to sit there. The lectures shot inside it cost nothing.

His name is Kevin Desmond. He taught 15.S50 at MIT Sloan in January 2015, a course moving poker analytics into trading and investment management.

The video attached is that full opening lecture, Desmond teaching it inside an MIT Sloan classroom.

The scoreboard is money per hour, not chips. A solid online grinder makes 5 big blinds per 100 hands, roughly $50 to $200 an hour. A big cash means nothing if the hourly is red.

Effective stack. The most you can lose in a hand is the smaller of the two stacks, not your whole pile. Position sizing is the same: risk 1 to 2 percent an entry, fix max loss before you act, and whoever skips that number goes broke.

M ratio. Effective stack divided by the blinds and antes, how many rounds you last folding every hand. Your runway is the same fraction: cash divided by monthly burn.

Four player types: tight or loose, aggressive or passive. The fish calls everything and is where the table money comes from. In any market the person who cannot fold is the fee everyone else collects.

Become a small winner first. Desmond tells the room to lock a tiny edge long before chasing a big one, because whoever sizes for the maximum gets wiped by variance.

Trading desks hire for that instinct. A new graduate trader at Jane Street starts near $200,000 in base pay, before a bonus that can carry the year past $400,000.

"You want to be a slightly winning player way before you want to become a huge winning player."

That is Desmond on why survival beats aggression.

All eight lectures and the notes sit on MIT OpenCourseWare for free. The seat that produced them runs $91,892 a year.

Millions can open the course today. Almost no one who gambles their savings watches a minute.

The lectures are free. The willingness to sit through eight of them before risking a dollar is far rarer than the nerve to risk it blind.

[Gemincho @geminchoo]

@Emberxbt Learn to lose slowly might be the best money advice I've heard all week

[Gemincho @geminchoo]

@Emberxbt A codebreaker outperforming every finance bro on Wall Street is just poetic

[Gemincho @geminchoo]

@Emberxbt Saved this one, going back to it before every big purchase from now on.

[Gemincho @geminchoo]

@Emberxbt good info

[Gemincho @geminchoo]

A 99-year-old billionaire went on national TV with $160,000,000,000 in cash and said only ten decisions in a great investor's whole life ever matter. Strip those ten out, he said, and even the best record on earth turns to garbage.

His business partner is the most celebrated investor alive. Take that partner's ten best decisions off the table and the lifetime record collapses to nothing. Six decades of work, down to about ten calls.

Harvard's MBA charges about $150,000 to teach you how to invest. He said on camera the degree cannot teach it. This interview costs nothing.

His name is Charlie Munger. He was 99, vice chairman of Berkshire Hathaway for over four decades and Warren Buffett's partner in building one of the world's largest companies.

The interview in this video is Munger with CNBC, walking through exactly how the whole thing was built.

The 2-and-20 tax. Most funds charge 2% of your money a year plus 20% of every gain. On a $1,000,000 account up 10%, that is $40,000 gone in one year, win or lose.

The patience position. He held $160 billion in cash rather than force a small deal, because nothing small moves the needle. Doing nothing is a position. A small account overtrading itself is the opposite.

The Wooden Effect. John Wooden ran 90% of his minutes through seven players. Berkshire ran money the same way. Spreading a portfolio across 30 names dilutes your best idea into noise.

The Harvard illusion. People pay six figures for the degree that is supposed to teach investing. He says it can't. The credential you are financing does not produce the skill.

The pie counter. His grandfather told him you get three or four real chances. When one comes, you do not take a small helping. Your net worth is decided by a handful of decisions, not daily activity.

"Just a few outliers make money."

That is Munger on the whole investing industry, in his final interview, weeks before he died in November 2023.

The full interview is on YouTube and CNBC for free. The degree people buy to learn the same lesson runs past $150,000.

Millions have shared his one-liners. Almost nobody sat through the full interview and changed how they bet.

The interview is free. The stomach to bet real money on your three best ideas and sit still on the rest is what almost nobody has.

[Ember @Emberxbt]

https://t.co/NiuPatJUal

[Gemincho @geminchoo]

@Emberxbt This is the first thread that actually explained why APR and the advertised rate are different numbers. Saved

[Gemincho @geminchoo]

@Emberxbt never took an econ class in his life and still won the nobel in it

[Gemincho @geminchoo]

@qentuaxbt Never realized timing mattered more than willpower here.

[Gemincho @geminchoo]

@Emberxbt Explains why people hold losers too long and sell winners too fast. The ending writes the story, not the total return.

[Gemincho @geminchoo]

@Emberxbt Sounds simple until you’re actually standing there and everything feels one-way.

[Gemincho @geminchoo]

He built a firm that sold over $1 billion worth of stock, employed 1,000 people at its peak, and defrauded more than 1,500 investors out of roughly $200 million. Then he served 22 months and became a bestselling author.

Jordan Belfort founded Stratton Oakmont on Long Island in 1989. The core mechanic was simple: buy penny stocks cheap, use high-pressure brokers to talk retail investors into buying at inflated prices, then dump the shares before the price collapsed back to what it was actually worth. He admitted to manipulating the stock of 34 companies this way, over roughly seven years.

The SEC caught the firm early. In 1994 Stratton Oakmont paid a $2.5 million settlement and Belfort was banned from running a securities firm. He sold his stake and kept operating anyway. It took until 1996 for regulators to finally shut Stratton down, and until 1999 for the FBI investigation to produce an indictment.

He pleaded guilty to securities fraud and money laundering, was sentenced to four years, and served 22 months after cooperating with prosecutors. The court ordered $110.4 million in restitution. Less than $12 million of it was ever recovered.

The math that should stick with you: a $2.5 million slap on the wrist in 1994 didn't stop a $200 million fraud. The penalty was priced in before the crime even peaked.

[Superior @andreysuperior]

https://t.co/rUN1cPDY65

[Gemincho @geminchoo]

@Emberxbt Bro built the entire modern risk framework and nobody in fintwit has ever mentioned his name once

[Gemincho @geminchoo]

Four agents each said "done." The output was garbage all four times.

Checked the logs first. Clean. Checked the prompts. Clean. Spent three hours debugging code that had nothing wrong with it, because the actual crime scene was between the agents, not inside any single one.

Node three returned a name. Node four expected an ID. Nobody complained. The pipeline just quietly filled the gap with its best guess, four separate times, and called each one a success.

One schema fixed what three hours of debugging couldn't touch. Turns out "it ran without errors" and "it worked" were never the same sentence.

[Carnage @0xCarnagee]

https://t.co/dTfSSTVUoc

[Gemincho @geminchoo]

@0xCarnagee the planner-generator-evaluator loop beating a single agent is the part worth rewatching

[Gemincho @geminchoo]

@polydao he wiring between the agents actually doing the work is the detail that lands

[Gemincho @geminchoo]

Watched the whole video because the term itself felt like something to roll my eyes at, and honestly by the end it earned itself.

The framing that stuck: prompt engineering is asking a better question, context engineering is feeding it better information, graph engineering is designing the work around the AI so it stops living in one giant chat and starts running as a managed workflow. Jobs are single-responsibility units of work. Arrows are the routing logic between them, sometimes linear, sometimes conditional, sometimes fanning out to three jobs and back into one. State is what gets passed along instead of re-explained every time.

What made it land was the worked example: whether to launch an AI bookkeeping tool for Shopify merchants, broken down into a graph instead of one exhausting prompt.

The part specific to Claude Code: you don't need LangGraph or a Python framework to start. Level two is just Claude Code or Codex, where each step in the graph writes its own file, leaving a paper trail you can compare, audit, and reuse weeks later instead of re-deriving it from scratch every time.

[codila @0xCodila]

https://t.co/Kg6X4rXvpc

[Gemincho @geminchoo]

NVIDIA JUST OPENED ACCESS TO THE LARGEST MODEL BUILT FOR SELF-DRIVING CARS

Alpamayo 2 Super is a reasoning VLA model, 34 billion parameters, that stitches footage from the front, side and rear cameras into one 360-degree view and computes the driving trajectory from it.

Shown first at GTC Taipei back in June, access just opened now.

The focus is the situations self-driving cars still handle worst: lane changes, turns across oncoming traffic, congested intersections.

Under the hood is Cosmos 3 Super Reasoner. After base training it went through RL on the consequences of its own decisions in simulated drives.

The interesting part: it outputs more than a route. It gives a text rationale for the maneuver, and a developer can trace which detail in the frame drove that decision.

NVIDIA ties this to safety validation under ISO/PAS 8800, and pitches the same mechanism for auto-labeling test footage, work that normally eats months of manual labor.

On LingoQA, where the model has to explain the road situation in plain language, it took first place, ahead of Qwen2.5-VL 72B and Gemini 2.5 Pro.

License allows fine-tuning on your own data and shipping it into production vehicles.

[Gemincho @geminchoo]

NVIDIA JUST OPENED ACCESS TO THE LARGEST MODEL BUILT FOR SELF-DRIVING CARS

Alpamayo 2 Super is a reasoning VLA model, 34 billion parameters, that stitches footage from the front, side and rear cameras into one 360-degree view and computes the driving trajectory from it.

Shown first at GTC Taipei back in June, access just opened now.

The focus is the situations self-driving cars still handle worst: lane changes, turns across oncoming traffic, congested intersections.

Under the hood is Cosmos 3 Super Reasoner. After base training it went through RL on the consequences of its own decisions in simulated drives.

The interesting part: it outputs more than a route. It gives a text rationale for the maneuver, and a developer can trace which detail in the frame drove that decision.

NVIDIA ties this to safety validation under ISO/PAS 8800, and pitches the same mechanism for auto-labeling test footage, work that normally eats months of manual labor.

On LingoQA, where the model has to explain the road situation in plain language, it took first place, ahead of Qwen2.5-VL 72B and Gemini 2.5 Pro.

License allows fine-tuning on your own data and shipping it into production vehicles.