The price dropped 70% in the 12 days it took people to review it Every teardown filmed two weeks ago is now wrong about the only number that mattered. The verdicts agreed. Good architecture, priced out of reach. One reviewer said he would have a different analysis if it cost less. It does now. Look at what the screen shows instead of a price. Salesforce, 52 accounts pulled. 3 lookalike segments. 4 profiles skipped because someone already contacted them. Then 36 drafts sitting in a queue, each with a recipient, an opener and a badge. Nothing sends until a person looks. That is the part worth paying for. The agent did the pull, the match and the writing, then stopped at the one step where being wrong is expensive. Chapter 1 of 17 is titled I Was Wrong About Grok Bot. The price fell 70%. The verdicts fell with it.

MY JAW IS ON THE FLOOR, I MADE A GROK BOT AGENT WRITE ITS TRADING BOOK AS SHEET MUSIC AND CONDUCT IT FOR 16 DAYS $3,600 became $187,400. +5,106%. 12 voices playing. 9 resolved. 1,200 rehearsal notes. Sixteen bars, one bar to the day, nobody holding the baton. Each staff line is a position and the note length is how long it holds. Dynamics is not a metaphor here, it is capital at risk, and the marker sits at fff by the end. Sections carry the risk. Strings $134,367. Woodwind $66,709. Brass $22,685. Percussion down 12.1% and still sounding. The rehearsal notes are where it gets strange. Diminuendo, trimmed 3 parts back to a whisper. Rest, no plot above 16% of the book. Cut the line, minus $3,607, the phrase went wrong. Six resolved as written and paid $42,750. Three got cut and cost $9,608 together. Quarter Kelly, $780 written into the part. It sized every entry before the first note played. Deepest rest across the whole run was 7.4%. An orchestra does not get loud by everyone playing. It gets loud by knowing who rests. 12 voices. 9 resolved. $187,400 in the book.

8 use cases, 26 minutes, and the number that actually matters is 4 Chapter 4 is a CEO bot. Chapter 5 is revenue. Both run on the same primitives, and once you see them the other six stop being surprises. Every agent gets its own cloud computer. Browser, file system, terminal. It stays on when the laptop closes, and a session you sign into once belongs to the machine, not to a chat window. The workspace is shaped like Slack. Agents get names, titles, roles. Tag one and it answers. Tell one to hand work to another and the message lands in a thread you can read, while you watch the second agent's screen and take the mouse at a login wall. Routines fire on a schedule or a trigger. 9 AM on weekdays becomes a condition, not a reminder. Then the part demos skip. Record yourself doing something, press stop, and the agent watches the recording back and writes a reusable skill from it. 4 primitives: a persistent machine, a directory of named workers, a scheduler, and learning by demonstration. Every use case in that video is a configuration of those four. So is every one that isn't. Which changes the question. Not what it does, but what 8 of them cost to keep running. $200 per month is the floor, and only on a qualifying plan. Closed source, Grok models only. 8 use cases is a list. 4 primitives is a product.

AI agent reviewer, 12-minute Grok Bot teardown: "If it was 20 bucks a month, I'd have a different analysis" Everyone is arguing about which agent is smartest. Almost no one is arguing about what it costs to find out. 12:29, 9 chapters, full setup recorded start to finish. $200 per month is the cheapest way in, models locked to Grok. 7 days free before the bill. 100 attempts to connect X, and it never connected once. 10x the price that would have changed his verdict. Full teardown below. Chapter 9 is where he says what he'd use instead of

i handed one AI conductor a baton and 12 agents, went to sleep, and the orchestra turned $1,000 into $40,552 in 24 hours i played 0 notes. the orchestra played 770. only 41% were clean. the conductor cut the rest before they turned into noise. the desk is built like an orchestra on purpose. 12 players, one baton, and not a single human in the room. STRINGS hold the trend and never rush it. BRASS come in on momentum bursts, loud and short. PERCUSSION handle execution and fills. WINDS take the risk breath and trim size after a loss. CHOIR reads the crowd and scores it. one conductor routes all of them. it decides who plays, when, and how loud, then logs every cue so nothing is improvised twice. here's the score while my phone was face down: 21:19 BRASS hits $OPUS88, 2.3 sol at the open 21:19 WINDS trim $NOCTURNE, size down to 0.4 sol 21:19 BEAT $FUGUE filled +1.30 sol 21:19 BRASS hits $TREMOLO, 1.5 sol at the open the honest line is on screen and it isn't the 40x. eleven quiet notes, one crescendo carried the whole night. a tail event, not a strategy, in its own words. loudest note 214x. notes muted 11. missed cues 3 all day. a human plays every instrument at once, off-key, on adrenaline. this was one baton keeping 12 hands in time while i slept. the orchestra never rests. that's the only reason it was still playing when the crescendo hit.

i watched a rocket lose an engine one minute into flight then i watched six AI agents vote 6 to 0 to keep going, and make orbit anyway no human touched the console. the top bar says it: 6 stations, 1 vehicle, no humans in the room liftoff was clean. 4 of 4 engines, velocity climbing, margin 72 percent at T+00:44 the prop station catches engine 3 chamber pressure falling. it commands the engine off. now they are down to 3 and short on thrust here is the part that gets me. the abort call goes out and every station votes no. guidance immediately extends the burn 14 seconds to recover the missing velocity MECO comes at 7.4 km/s. final orbit 412 by 398, target was 400 engines 3 of 4 abort votes 0 of 6 margin fell to 34 percent and held it wasn't the launch that was impressive. it was six machines agreeing, under load, that a broken rocket was still good enough to finish the job

HE GUY RUNNING 8,000,000 COLD EMAILS A MONTH TESTED GROK BOT FOR ONE HOUR AND KILLED TWO TOOLS He said it on camera, in the middle of a tutorial, with no build-up. "we're not going to use Hermes and OpenClaw" Don't bookmark this if it crosses your feed, it's only 17 minutes. Some context on who is saying it. His agency was Clay's first marketing contractor back when Clay had 4 employees. It has since run outbound for 300+ customers and become Clay's largest user by enrichment volume. Watch the volume line move: 2023 → 1,200,000 emails a month 2025 → 4,000,000 now → 8,000,000+ That is the desk he was willing to rip a tool out of after 60 minutes. The whole setup is in the video. He shows the roles, the handoffs, the sending stack. Nothing gated. Steal it. Everyone else is still choosing. He already chose and moved on.

CEO of OpenAI, on day one of the company: "I have no idea what I'm doing." 12 people in Greg Brockman's apartment, January 2016. Someone went to find a whiteboard. It arrived, and the room went quiet again. Everyone studies the launch. Almost nobody studies the 4 and a half years OpenAI ran with no product and nobody to tell them the research was working. So they invented customers. Internal leaderboards. Demos staged for one visitor the researchers wanted to impress. Fake deadlines, which failed. Watch the part where Altman names the 2 products he killed last year. Both worked. Both lost their compute to something else. The stumbling produced GPT. ChatGPT has been public under 4 years and now runs at roughly 1,000,000,000 users a week. 4 and a half years with nothing to show. One decade to a billion.

He ran research at the lab everyone was chasing, then walked out with 6 colleagues and built the competitor. Dario Amodei left OpenAI at the end of 2020. His sister Daniela went with him. Watch how he frames the decision. Not a fight over a product, over what should be built next and who decides that. Anthropic started in 2021 with no model, no customers, and a group of people who had already shipped the state of the art elsewhere. The company is now valued at $183,000,000,000, with 80% of revenue coming from businesses. He spends interviews warning about the technology while racing to build it faster. Seven people left a room. The room they built is now the thing to beat.

A founder sat on a Silicon Valley stage in 2013 running the largest social network in Russia, and within a year he had lost it and left the country. Pavel Durov built VK with over 100,000,000 users and was still in his twenties when he took that seat. Telegram was months old at that point and existed as a side project. Watch how he talks about the two products. One was the asset everyone wanted. The other was the one he could carry out with him. In 2014 he refused a state demand for the data of Ukrainian protest organisers, sold his VK stake, and flew out. He has run Telegram from outside Russia ever since, funding it out of pocket for years before it took revenue. Telegram passed 1,000,000,000 monthly users last year. He gave up the company with the users. He kept the one with the encryption keys.

The CEO of Goldman Sachs was sworn in before the Senate while protesters in prison stripes stood behind his chair. Lloyd Blankfein spent hours explaining why a bank that sells a security has no duty to tell the buyer what it thinks of it. The senators kept returning to one word. Market maker. He took the other side of client trades and called that the job. Watch the room when the exchange sharpens. Carl Levin reads the internal emails aloud, and the language in them is not the language of the testimony. Goldman took $10,000,000,000 in TARP money in 2008 and repaid it in 2009. It received about $12,900,000,000 through the AIG rescue, paid at full value on contracts the market had already written down. Three months after this hearing the firm settled with the SEC for $550,000,000. The bank paid the fine and posted a profitable year. Nobody at the table lost a job.

The richest man in the world spent 20 hours under oath arguing about what the word "concerned" means. August 1998. A federal antitrust case that could break Microsoft in two, and the camera runs the whole time. He rocks in the chair. He pauses for 10 seconds before easy questions. He answers his own emails back to himself as if he had never seen them. Watch the moment the government lawyer stops asking and starts reading. The tape gets played in open court and the judge watches it with the jury box empty. Judge Jackson ordered the company split in two in June 2000. The appeals court threw out the split, and kept the finding that Microsoft held a monopoly. Every fact in the case was already on paper. He lost the room on posture.

Elon Musk sat on a conference sofa in November 2008 with both his companies weeks from running out of money and the credit markets frozen shut. Lehman had collapsed 8 weeks earlier. No investor in the room was writing cheques for an electric car startup. Tesla had laid off staff in October and was down to a balance that would not clear another payroll cycle. Watch how he handles the funding question. He is being asked to defend a business that had already burned through most of what he made at PayPal. The Model S existed as a plan and a price nobody believed, and the Roadster was shipping in small numbers at a loss per car. Six weeks after this conversation the Tesla round closed on Christmas Eve. He was on a sofa arguing for a company he was days from losing.

Richard Roll, throughout his career, had a habit of dissecting financial models, and then he interviewed the person who had created three of the most lasting ones. Few people watch that interview. Stephen Ross, wearing a bow tie, in May 2009, eight months after Lehman. The Arbitrage Pricing Theory was introduced in 1976, and he substituted a single market factor with several factors while giving up all the assumptions that everybody else had accepted. Look at the agency theory paper from 1973; it established the framework within which executive compensation has since been discussed. The binomial option pricing model originated in the same mind and is the version that all trading desks teach first. His risk-neutral work had just been put to the test by the crash in public. There were three basic papers during that career, but the Nobel Prize was awarded elsewhere.

Larry Page and Sergey Brin sat for an interview in 2000 with a product used by millions and no working way to charge for any of it. Almost no one watches the tape. Two Stanford graduate students, a company founded in September 1998, a search engine growing faster than the plan for paying for it. They had tried to sell the whole thing to Excite for about $1,000,000 and been turned down. Watch how they answer the money question. Neither reaches for a business model, and neither sounds worried about it. The rule they set was the one everyone else broke. Nobody buys a place at the top of the results. Their competitors sold the ranking. All of them are gone. That interview sat unwatched while the company built on it passed $2,000,000,000,000.

The four words were written on the board and it was that layer not the machine that Apple bought for $429 million. Almost nobody watches him giving the explanation. Steve Jobs was at NeXT years after Apple had expelled him and years before they brought him back. The board states Custom App 3x and the entire presentation depends on that single figure, something he never alters. See the little time he spends on the hardware. Although he has built one of the most beautiful machines of the decade, he still directs the conversation towards the software that runs on it. Under it he wrote IPC. He turns the plumbing detail into a reason for purchasing. No slides. No agenda. Just a marker and a board. Eighteen minutes. One take. In 1993 he closed the factory and retained the layer which he had drawn on the board.

Sergey Brin and Larry Page wrote in 1998 that advertising would corrupt search. Google took about $295,000,000,000 from advertising last year. The early footage shows 2 graduate students standing next to hardware they assembled themselves. At Stanford their first storage rack was a case built from LEGO bricks holding 10 drives at 4 gigabytes each, because a proper enclosure cost more than they had. They incorporated in September 1998 and worked out of a rented garage in Menlo Park for $1,700 a month. Andy Bechtolsheim had already handed them a check for $100,000 made out to a company that did not legally exist yet. Their paper that year carried an appendix on the business model. They wrote that "advertising funded search engines will be inherently biased towards the advertisers," and that the bias would be almost impossible to detect from outside. They gave an example. A major search engine would not return a large airline's own homepage when you typed the airline's name, because the airline had paid for that keyword. Their conclusion was that a better search engine needs fewer ads, and that this kind of thing belongs in the academic world where nobody has to sell anything. Sequoia and Kleiner Perkins put in $25,000,000 in June 1999. AdWords launched in October 2000 with 350 advertisers. By the 2004 prospectus the position had flipped. Advertising was now the principal source of revenue, and the ads were described as relevant and useful rather than intrusive. Alphabet finished 2025 at $402,800,000,000 in revenue. Search alone brought $224,530,000,000. They predicted the bias in 1998. They collected $295,000,000,000 from it in 2025.

Steve Jobs sold about 50,000 NeXT computers in 8 years. Two of the people who bought one built the web and Doom. October 12, 1988, Davies Symphony Hall, San Francisco. 9:30 in the morning to noon. Over 3,000 invitations went out to educators, developers and reporters, and not one to an Apple employee. He had been pushed out of Apple 3 years earlier. He opened with 5 words. "It's great to be back." The machine was a 1-foot magnesium cube. 25 MHz processor, 256 MB optical disk, no floppy drive, $6,500 to universities. He called it 5 years ahead of its time. They had rented a concert hall for the acoustics. He pulled the complete works of Shakespeare off the disk in seconds, then had the Cube play a duet with one of the symphony's violinists. The room stood up. Then it shipped more than a year late. Businessland took it to market near $10,000 and moved a few hundred units in 1989. About 10,000 existed by late 1990. NeXT killed the hardware business in 1993. One of those 10,000 sat at CERN. Tim Berners-Lee wrote the first web browser and ran the first web server on it in 1990. The machine still carries a handwritten label telling people not to power it down. Another sat at id Software, where John Carmack built Doom on it. In December 1996 Apple paid $429,000,000 and 1,500,000 shares, not for the cube but for the operating system inside it. That code became Mac OS X, and Mac OS X became the iPhone. The hardware sold 50,000 units. The software ended up in your pocket.

Mark Zuckerberg dumped a plastic bucket over his head in 2014. Bill Gates answered with blueprints, a blowtorch and a pulley rig, and the campaign raised $115,000,000. August 13, a backyard, one camera. Zuckerberg, challenged by Chris Christie, lifts the bucket by hand and names three people: Bill Gates, Sheryl Sandberg, Reed Hastings. Gates watches the clip on a laptop. "I'm glad to give to ALS. It's a great cause." Then the line that changes the video. "I want to do it better than it's been done." He goes to a drafting table. Blueprints, drawn in Comic Sans. A paper prototype. A welding mask and a torch, a steel frame with a bucket hung above a chair and a rope running down to his hand. August 15. He sits, pulls the rope, takes the whole thing. Then he names Elon Musk, Ryan Seacrest and Chris Anderson of TED, 24 hours on the clock. The internet filed it under rich men being silly. That summer produced 17,000,000 videos and $220,000,000 worldwide. The ALS Association alone took $115,000,000 in 8 weeks, against $2,800,000 in the same 8 weeks the year before. $77,000,000 of it went to research. $1,000,000 went to a gene-sequencing effort called Project MinE, 80 researchers across 11 countries. July 2016, Nature Genetics. A gene called NEK1, found in roughly 3% of ALS patients, now ranked among the most common ever tied to the disease. The stunt got the laughs. The $1,000,000 got the gene.

Two men built a $800,000,000,000 empire together over 64 years and never signed a contract with each other. Omaha, 1959. A mutual friend seats Warren Buffett next to Charlie Munger at lunch. Buffett is 29, Munger 35. They talk, and something clicks that never comes apart. Munger was a Harvard-trained lawyer who could have papered every handshake. He didn't. He formally joined Berkshire in 1978 with no employment agreement, no partnership deed spelling out who gets what. The whole thing ran on trust. Munger reshaped Buffett's entire approach, pushing him from buying cheap junk to paying up for great businesses. Buffett called him the architect; he was just the general contractor building the vision. The proof is in the friction, or the lack of it. "In 62 years, we've never gotten mad at each other," Buffett said, "and it's not that we agree on everything." Munger died in 2023, weeks short of 100. Buffett's line: "I've lived a better life because of Charlie." Lawyers draft contracts to plan for betrayal. These two never needed the paper.