Nobody trains their own model anymore. A Stanford professor showed why in one slide:
05:37 - why one long chat eats most of a GPU's memory, and why that caps everything you build
36:00 - 128 experts in memory, 8 fire per token. How frontier models got huge and cheap at once
54:00 - he writes a pattern on the board, the model finishes it. Nobody trained it to do that
1:02:01 - the money slide. One off-the-shelf model, and your only edge is the task description you write for it
That last one is the whole business model of AI agents, said out loud in a lecture hall.
The hard part was never the model. It is who splits it into roles that actually finish work - which is exactly what Grok Bot is.
Save this - the article below is how to build that agent team in an evening ↓
A GROK BOT BOUGHT MY CAR FOR ME. IT WATCHED THE MARKET FOR 9 DAYS AND TALKED THE DEALER DOWN $4,200.
Nobody loses money on a car because they picked the wrong model. They lose it because they got tired.
You look at listings for two weekends, you cannot tell a fair price from a hopeful one, and by the third dealer you just want it to be over. The seller knows that. It is the whole business.
So you stop haggling and put a crew on it:
→ CHIEF - takes your budget, your must-haves and your walk-away number, and briefs the other seven
→ SCOUT - watched 1,412 listings for 9 days, not two Saturdays
→ VALUER - built the actual price curve for that model and mileage. $19,400 is fair. Everything above the line is someone hoping
→ INSPECT - pulled the VIN, the history and the service records on 9 cars. Found the timing belt due at 70,000 km
→ HAGGLE - messaged the dealer and did not get tired. 11 messages, every one backed by a comparable or a cost
→ FINANCE - pulled two loan quotes and cover so the dealer's finance desk had nothing to sell
→ PAPERS - had the transfer ready before Saturday
→ BRIEF - one page. What it costs, what it is worth, what is still unknown
Here's why this works on Grok Bot and not on a listings site.
Every Grok Bot agent gets its own computer and browser, so SCOUT reads the real listings and HAGGLE writes to the real seller. They share memory, so INSPECT finding the timing belt is what HAGGLE uses to take $700 off - nothing gets pasted between tools. And it runs in the cloud for nine days straight, which is exactly what a person cannot do.
And you write zero code. You describe each agent in plain English, tell it where to look, give it one job.
The one honest catch: you have to give it a real walk-away number. Say $18,000 and mean it, or it will negotiate against itself.
Here's the part nobody says out loud. Negotiating is not a skill. It is a stamina contest.
The side that is willing to walk away and wait wins. A bot is always willing to wait.
Start with two: VALUER and HAGGLE - one tells you what it is worth, the other refuses to pay more.
Prices on the dashboard are an example purchase.
Everyone else is still deciding in a dealership on a Saturday with a coffee going cold.
Bookmark this & read more about Grok Bot in the article below ↓
You have a second job you never applied for.
Inbox, subscriptions, appointments, documents, follow-ups.
Six hours a week, every week, forever.
I gave all of it to six Grok Bots in four weeks.
This article is the entire setup, including what breaks. https://t.co/cq2kDudWIg
I APPLIED TO 64 JOBS IN A WEEK AND WROTE ZERO COVER LETTERS. 8 GROK BOT AGENTS SCANNED 2,140 POSTINGS, PUT 7 INTERVIEWS ON MY CALENDAR AND SURFACED AN OFFER $18K ABOVE THE POSTED BAND.
Job hunting does not fail on talent. It fails on volume.
You need to see 2,000 postings to find 60 worth your time. Nobody does that after work, so people apply to nine roles, hear nothing, and decide the market is broken.
So you stop scrolling boards and put a crew on the hunt:
→ CHIEF - takes your level, your stack, your money and your red lines, and briefs the other seven
→ SCANNER - reads 2,140 postings across 9 boards. Every night, not once on Sunday
→ MATCH - dropped 1,954 that did not actually fit your seniority or your band. That is the part that saves your evenings
→ TAILOR - rewrote the CV per role. 64 versions, none of them a copy paste
→ SENDER - applied to all 64 and logged which version went where
→ CHASE - followed up on day 4 with the 41 that went silent. 3 of the 19 replies only came because of that
→ PREP - booked the interviews and wrote a brief per company. Their stack, their people, what they will ask
→ BRIEF - one page. Two live offers, one decision to make
Here's why this works on Grok Bot and not in a job board.
Every Grok Bot agent gets its own computer and browser, so SCANNER signs into the boards the way you would and SENDER fills the real application forms. They share memory, so MATCH's reasoning is what TAILOR writes against and PREP briefs from - nothing gets pasted between tools. And it all runs in the cloud, so the scanning happens overnight with your laptop shut.
And you write zero code. You describe each agent in plain English, tell it where to look, give it one job.
The one honest catch: it needs your real red lines. Say no to relocation and no to on-call, or it will book you both.
Here's the part nobody says out loud. The best candidate rarely gets the job.
The candidate who was in front of the right team in the right week gets the job. That is a volume problem wearing a talent costume.
Start with two: MATCH and CHASE - one stops you wasting nights on roles you would hate, the other gets replies from people who simply forgot.
Roles, fits and numbers on the dashboard are an example week.
Everyone else is still writing one cover letter a night and calling it a search.
Bookmark this & read more about Grok Bot in the article below ↓
Hearth home floor
MP4
MY SUPPORT INBOX HAD 212 TICKETS AND A 4 HOUR REPLY TIME. 7 GROK BOT AGENTS TOOK IT TO 41 SECONDS, ISSUED 9 REFUNDS THROUGH STRIPE, AND NOT ONE DOLLAR MOVED WITHOUT MY TAP.
Support does not break because the answers are hard. It breaks because they are the same answers, all day, forever.
Where is my order. Charged twice. Wrong size. Never arrived. You already know every reply. You just cannot type them fast enough at 2am.
So you stop typing and put a crew on the floor:
→ TRIAGE - reads every ticket before you do. Sorts into needs a human, needs a reply, needs nothing. Folded 14 duplicates into 4 threads
→ DRAFT - writes the reply in your tone and queues it. You send or you edit, you never start from blank
→ POLICY - matches the ticket against your actual refund rules and scores the fit. 96% on a duplicate charge is not a judgement call
→ REFUND - talks to Stripe directly. Issues the money in 1.4 seconds once approved
→ ESCALATE - spots the genuinely angry thread and hands it to a person instead of trying to smooth it over
→ KNOWLEDGE - turns resolved tickets into new answers, so the same question is cheaper next week
→ BRIEF - one page at close. Money returned, what got escalated, one thing that needs you
Here's why this works on Grok Bot and not in a helpdesk tool.
Every Grok Bot agent runs on a shared cloud computer, so TRIAGE's decision lands straight in DRAFT's queue and POLICY's check - nothing gets exported between tools. It connects to Stripe by API and to your support mail directly. And you set approval rules, so nothing sends, charges or refunds until you tap.
And you write zero code. You describe each agent in plain English, tell it where to look, give it one job.
The one honest catch: your refund policy has to be written down somewhere. Vague policy in, vague decisions out.
Here's the part nobody says out loud. Nobody churns because a refund was slow.
They churn because nobody answered for four hours and they had to ask twice. Speed is the product.
Start with two: TRIAGE and DRAFT - one decides what matters, the other has the reply waiting before you open the tab.
Ticket volume and refund totals on the dashboard are an example shift.
Everyone else is still hiring one more person for the night shift.
Bookmark this & read more about Grok Bot in the article below ↓
You have a second job you never applied for.
Inbox, subscriptions, appointments, documents, follow-ups.
Six hours a week, every week, forever.
I gave all of it to six Grok Bots in four weeks.
This article is the entire setup, including what breaks. https://t.co/cq2kDudWIg
I HAVEN'T OPENED A SINGLE TAB AND MY 11-DAY TOKYO TRIP IS BOOKED. 7 GROK BOT AGENTS SEARCHED 1,840 OPTIONS, CAUGHT A $212 PRICE DROP AND BUILT THE WHOLE ROUTE WHILE I SLEPT.
Everyone wants the trip. Nobody wants the six hours of tabs it takes to book one.
Same flight at three prices. Reviews you cannot trust. A route that crosses the city twice a day and a hotel that looked closer on the map.
So you stop opening tabs and hire a crew that does the sitting-and-refreshing for you:
→ CHIEF - you say where you want to go and how you like to travel. It briefs the other six
→ SCOUT - ranked 14 destinations against your real criteria, not a top-10 list
→ FARE - watched 5 routes every 4 minutes for a week. Caught LIS to NRT falling $212 overnight
→ INSIDER - pulled 62 local spots with opening hours checked, not the ones with the most reviews
→ ROUTER - built 11 days at 4.2km average walking, so you are not crossing the city twice
→ CONCIERGE - held the seats, waited for my tap, then booked and filed every confirmation in one place
→ SENTRY - watches visas, weather, strikes and closures. Hakone line shut on Tuesday, so day 4 got rebuilt before I knew
Here's why this works on Grok Bot and not in a booking site.
Every Grok Bot agent gets its own computer and browser, so it signs into the real airline and hotel sites the way you would. They share memory, so FARE's price drop lands directly in ROUTER's plan and CONCIERGE books it - nothing gets pasted between tools. And it all runs in the cloud, so the refreshing keeps going with the laptop shut.
And you write zero code. You describe each agent in plain English, tell it where to look, give it one job.
The one honest catch: it needs your real preferences first. Say you hate early flights and long museum days, or it will book you both.
Here's the part nobody says out loud. A travel agent was never selling knowledge.
They were selling the six hours nobody wants to spend. That is the part that just got automated.
Start with two: FARE and ROUTER - one buys the ticket at the right moment, the other makes sure the days are actually walkable.
Prices and route on the dashboard are an example trip.
Everyone has a trip they have been meaning to book for months. This is what finally books it.
Bookmark this & read more about Grok Bot in the article below ↓
WELL F**K, THE NIGHT DECK HANDED OFF TO A DAY DECK I DID NOT BUILD
cobalt twelve - harbor deck, deck 9, runs the day burn. same floor logic, new crew, picks up exactly where the night shift dropped it
9 agents → 3 rows → one duty officer → human approval
every handoff is a bot physically walking the work across the floor. #cbt-9052 opens at sev 9.0, evidence lockers fill, the case forge stacks confirmed cases, mesh deck sweeps for what got missed
freeze hosts, pull line 40%, roll creds - the dangerous half sits in the budget panel and waits for a human. everything else the deck just does
two decks now, one grid. i keep asking for a chat and getting a building
HOLY SH*T, GROK BOT GAVE ME A SECOND DECK AND IT DEFENDS THE F**KING MESH
halcyon nine - mesh desk, deck 7, keeps the grid honest. same night shift idea, mirrored floor, completely different job
9 agents → 4 zones → one shift commander → human approval
sifter, tracer and latch hold the wire row. mirage, crypt and picket run the cold row. ember, relay and glass sit in archive. the shift commander owns field ops in the middle and hands work down by zone
this deck fights instead of just watching. tarpit open, honeypot bay filling, isolating host, hash matched, beacon spotted, mesh resealed - every one of those is a bot physically walking to the bay and doing it
#orc-2291 runs at sev 9.4 while the contain counter ticks 02, 05, 08 and the seal number climbs past 32. evidence lockers fill on the right, the case foundry stacks confirmed cases, mesh deck radar keeps sweeping in the corner
response budget holds the dangerous buttons - sweep hosts, isolate 60%, rotate keys, block egress - and none of them fire until i say so
and when the shift rolls over it prints shift handover across the floor like a real desk changing hands
xai built a chat product and i keep accidentally turning it into staffed buildings
Before the agency and the hundred million, Gary Vaynerchuk was a guy in a plaid shirt yelling at a room about a business nobody respected: selling wine on the internet.
https://t.co/94tDiklL56 launched in 1997, when almost no one was buying alcohol online. He scaled his father's store roughly twentyfold, from $3 million to $60 million, by doing what the industry considered undignified - reviewing wine daily on camera, in normal language, without the vocabulary that kept ordinary people out.
Then he did the same thing to social media before it had budgets attached.
His pitch here is simple and it predates his fame: care more about the customer than about looking professional. Reply to everyone. Show up daily for years before it pays.
He was an immigrant kid with a D average and a card table at a flea market.
He didn't find a shortcut. He found a channel nobody wanted, and refused to leave it.
Jim Rogers retired at 37 with enough money to never work again, and he spent the next two years riding a motorcycle around the world.
Net worth today: around $300 million. He earned the core of it in a decade. With George Soros he ran the Quantum Fund from 1970 to 1980 and returned about 4,200% while the S&P managed roughly 47%. He was the research half of that partnership - the one who read the annual reports of Portuguese utilities and Botswanan mines.
Then he walked away at the top, taught at Columbia, crossed six continents by bike, and later drove around the world by car.
He moved his family to Singapore in 2007 so his daughters would grow up speaking Mandarin. He said the 19th century belonged to Britain, the 20th to America, and the 21st to Asia.
He does not diversify to feel safe. He waits for money lying in a corner and then picks it up.
The rare skill isn't finding the trade. It's leaving when you've already won.
Tim Ferriss turned down a job offer, wrote a book 27 publishers rejected, and built a fortune estimated around $100 million.
Before that he was running a supplement company, working around 14 hours a day, sleeping badly, and close to burnout. He took what was meant to be a short trip abroad and discovered the company ran fine without him. That became The 4-Hour Workweek.
He is not primarily an author. His money came from early positions in Uber, Facebook, Shopify, Twitter, Alibaba and Duolingo - many taken as advisor stakes when nobody was competing for them.
Then he stopped angel investing at the peak, saying the psychological cost of tracking dozens of companies was ruining his ability to think.
His whole method is subtraction. Cut the inputs, cut the commitments, then examine what's left.
Most people optimize how much they can carry. He optimized how little.
Jeff Bezos walked away from a guaranteed Wall Street bonus at 30 and his boss told him it was a stupid idea for someone who already had a good job.
He is now worth roughly $230 billion.
The man at this podium is a former hedge fund vice president at D. E. Shaw. He noticed the web was growing 2,300% a year, made a list of twenty products that could be sold online, and picked books because there were more titles than any physical store could ever stock.
His parents put in about $245,000 - most of their savings. He told them there was a 70% chance they would lose all of it.
He drove to Seattle, wrote the plan in the car, and packed orders on his knees on a concrete floor until someone suggested knee pads. He asked why they didn't just buy tables.
Amazon lost money for years while critics called it a bookstore with a stock price.
He built a $2 trillion company by being willing to look wrong for a decade.
Sam Altman built the graph that explains why most startups die, and it has nothing to do with size.
The chart shows number of users against intensity of liking. Two peaks. On the left, a lot of people who mildly like your product. On the right, a small number who love it.
The left group kills companies. The right group builds them.
He speaks from the founder's side of it - he started Loopt at 19, ran it for seven years, and sold it for $43 million, which he describes as a mediocre outcome, not a win. Then he ran Y Combinator, which funded Airbnb, Stripe, DoorDash and Coinbase, before taking over OpenAI.
His point is that it is far easier to expand something a few people love than to deepen something many people tolerate.
Building for everyone is how you build for no one.
Seth Klarman is worth about $1.5 billion, and his out-of-print book sells for thousands of dollars a copy.
He runs Baupost Group - roughly $27 billion - and has averaged double-digit returns for four decades, largely by refusing to be fully invested. He has held 30-40% of the fund in cash for years at a stretch, returning money to clients when he could not find anything cheap.
His entire framework is downside first. What can this lose. Who is the forced seller. What breaks if I'm wrong.
He built the fund on distressed debt and things nobody else would touch, buying when the seller had no choice.
Most managers are paid to be invested. He is one of the few willing to be paid for waiting.
Steve Jobs was worth $3 billion when he told 23,000 people he had been fired from his own company at 30.
Apple had grown from a garage into a $2 billion business with 4,000 employees. The board removed him. It was public, it was humiliating, and he considered leaving Silicon Valley entirely.
What followed: NeXT, and Pixar - a computer graphics division he bought for $5 million and funded with $50 million more. Pixar's IPO in 1995 made him a billionaire, and it happened before Apple was ever worth anything to him personally again. Apple then bought NeXT, and he was back.
At Apple he drew a salary of $1 a year. His wealth came almost entirely from that Pixar detour.
Getting fired was the most expensive thing that ever happened to him. It also built everything he came back with.
The founder of SpaceX wore a stick-on name tag at his own event.
The footage shows him in the early days, walking through a room of engineers, no entourage, no stage. His actual job then was not vision - it was chief engineer. He read propulsion textbooks. He asked suppliers why a valve cost $50,000 and then had it built in-house for a fraction.
That is the whole cost structure of SpaceX in one habit: refuse the industry price, decompose the object into raw materials, rebuild it.
He put the desk on the factory floor and slept there when launches were close.
Most executives buy expertise. He acquired it, then priced everything from first principles.
You cannot negotiate a supplier down if you don't know what the thing is made of.
Warren Buffett filed his first tax return at 13. He deducted his bicycle.
He was delivering 500 newspapers a morning. Before that he was reselling Coca-Cola by the bottle, hauling golf balls out of ponds, and running used pinball machines in barbershops with a friend - the barbershop got half, and neither of them ever fixed a machine they could avoid fixing.
His first stock came at 11. Three shares of Cities Service Preferred at around $38. It dropped, he panicked, he sold near $40 for a small profit. The stock later went to roughly $200.
He said that taught him more than any win would have.
By 26 he had about $174,000 and calculated he could retire on it. He didn't. He bought the house in Omaha for $31,500 and never moved.
The compounding wasn't just the money. It was that he started at 11 and never stopped.
Charlie Munger spent his life studying why smart people do stupid things. He put the answer in one sentence: "Never, ever, think about something else when you should be thinking about the power of incentives."
His favorite proof is FedEx. The whole system depends on every plane meeting in one place overnight and swapping packages. For years the night shift ran late. Management tried everything. Nothing worked.
Then someone stopped paying the workers by the hour and started paying them by the shift - finish early, go home. The problem disappeared.
The workers weren't lazy. They were being paid to be slow.
Munger's rule follows from that: when you find an absurd outcome, don't look for stupid people. Look at who profits from the absurdity. He also said the man who sold fishing lures painted purple ones - not because fish like purple, but because fishermen do.
Whenever behavior makes no sense, you are looking at the wrong incentive.
Ray Dalio keeps an email a subordinate sent him after a client meeting.
"Ray, you deserve a D-minus for your performance today."
He didn't fire the guy. He built a $160 billion company around making that email normal.
The reason traces back to 1982. Dalio publicly predicted a depression. Mexico defaulted on cue - and the market bottomed and rallied for 18 years. He lost everything, laid off every employee, and borrowed $4,000 from his father.
The lesson wasn't "be more careful." It was a permanent change in the question he asked. Not "am I right" but "how do I know I'm right?"
So Bridgewater runs on radical transparency. Meetings recorded. Disagreement mandatory. Every opinion weighted by the holder's track record, not their rank - a junior analyst with a good record outvotes a partner with a bad one.
Most firms pay for confidence. Dalio pays for doubt.
The most dangerous person in any company is the one nobody is allowed to grade.
In 1982 Ray Dalio went on TV and predicted a depression. Mexico defaulted right on schedule.
Then the stock market bottomed and went almost straight up for 18 years.
He lost everything. He laid off every employee and borrowed $4,000 from his father to pay his bills.
That failure produced one question that rebuilt him: instead of thinking "I'm right," he started asking "how do I know I'm right?"
The answer became Bridgewater's operating system. Radical truthfulness and radical transparency - every meeting recorded, every disagreement surfaced, every opinion weighted by the person's track record, not their title. An idea meritocracy where the best idea wins even if it comes from the most junior person in the room.
That system turned a one-man shop into the largest hedge fund in the world.
His edge was never being right. It was knowing exactly how easily he could be wrong.