SuccessDecoded @SuccessDecodedX

Ninety-two minutes into a podcast, OpenAI's former chief scientist defends belief over data. Ilya Sutskever sat down with Dwarkesh Patel last November, and most people clipped the headline: the age of scaling is over and research is back. The part worth keeping sits in the last chapter, "Research taste," which starts at 1:32:42 in the episode's chapters. He says "The top-down belief is the thing that sustains you when the experiments contradict you." Then he asks the question anyone who builds things hits sooner or later. Is it a bug, or is it the wrong direction? Data alone can't tell you, so he leans on a belief about how things have to be, built on simplicity and on the brain as the model. What follows is my reading, not his. A top-down belief only earns the right to overrule your data if you can say in advance which result would kill it. Without that line, taste turns into stubbornness with a nice name. He never names that line, and for him skipping it is cheap. In the same interview he mentions his startup, Safe Superintelligence (SSI), had raised $3B, so a wrong belief costs him a year. Bet your own savings on one and it costs you the company. Look at what the belief is doing at SSI now. It was valued at $32B in April 2025. As of this July it had around 50 people and no product or revenue, and it had just signed a deal with Nvidia for roughly ten times its compute. That works out to roughly $640M of valuation per person, and my reading is that almost all of it is priced on top-down belief. Ninety-two minutes of everything else, about three minutes of taste. What's the belief you'd keep debugging for, and which result would make you walk away?

@drvineetgoyal agreed when it's a choice. The trouble is when silence is just missing machinery

Why does Snap's first product manager leave his best people alone? Peter Sellis sat down with Lenny's Podcast for ninety-seven minutes nine days ago, and the line I saw quoted from it was the one about designing teams like terrorist organizations. Sharp line. It tells you how he sorts people, and says nothing about what he does with them after. That line comes from minute six. The sentence that matters is three quarters of the way through. The chapter called ride your best people hard is where Sellis lays out the method plainly. People who keep making good decisions get handed more and more decisions, until they start making bad ones. A few minutes later he gives the reason underneath it. He doesn't have a good way of helping people improve, so he lets them fail on their own. 15:16 why the median product manager is probably bad 46:25 ride your best people hard, and the admission sitting inside it 50:13 why collaboration is often overrated Those are the episode's own chapter marks. The line I quote below comes from the transcript, so read it as close to word for word. Put the two halves together and the philosophy turns over. The autonomy his best people get is real, and it arrived by accident. He's routing around a hole in his own toolkit, and from the outside that hole looks exactly like autonomy. He knows how it reads. What you have to control for, he says, is that "not paying attention to somebody doesn't mean they're failing." This next part is my reading, and I'll mark it as mine. Your manager's silence isn't feedback in either direction. It's information about their system, and what it says about you is nothing at all. Whether it means trust or neglect comes down to whether that person ever built a way to coach anyone, and most never did. Sellis was Snap's first product manager and spent about seven years there, finishing as VP of business product, then ran product at Discord. That background matters for one reason, and this is my reading again. He was hiring from a pool where nearly everyone was already strong, which is the condition his method depends on. Which is also the price of the principle for you. If you're not already the person making consistently good decisions, his approach hands you nothing, and nothing in the episode suggests he sees that as a problem. Most people walk out of a review waiting to hear what they did wrong. The uncomfortable version is that plenty of managers never built the machinery to tell them.

@dlLambo Yep ))

@lctrnotes plain sight is where it usually is. that's the whole job

OpenAI priced its own agent swarms at minute four, and nobody clipped it. Noam Brown from OpenAI sat down with Dwarkesh Patel on September 17. The moment that traveled was a swarm of ten thousand agents chewing on an unsolved math problem. Fair enough. It's a good clip. Four minutes in, before any of that, he's going through the scaling numbers from OpenAI's own release and putting a price on running agents in parallel. His framing, close to verbatim: you're paying 2x more to get an answer twice as quickly. Sit with that, because it's an exchange rate, not a discount. Double the spend buys you half the wait. Push further and it gets worse, because the curve flattens as agents pile up. He also names the kind of work that refuses to split at all, and his example is writing a novel. Worth opening the episode at these points: 03:58 the scaling numbers and what speed costs 15:28 how he thinks companies made of agents would organize 1:08:34 why he thinks chain of thought is getting worse, and what that does to everything above My reading now, and he didn't say this. Parallelism is a deadline instrument. Nobody fans out to spend less, they fan out to finish sooner, and the calendar is the thing being purchased. So the question was never how many agents to run. It was whether the job has parts that don't need to talk to each other, and whether the deadline is worth what the split costs. There's a second reason to fan out, worth separating: run four attempts and keep the best one. That buys a better shot at a good answer rather than a faster one. Different purchase, different bill, and most people running agents today couldn't tell you which of the two they're paying for. He can price speed because he has OpenAI's compute behind him. Asking which purchase you're making costs nothing. It works the same on three contractors, on a team you're about to hire into, on the sixteen tabs open in front of you. Somebody is going to charge for that judgment eventually, because right now everyone makes it by feel and learns the answer from the invoice. Think about the last job you split across three people to move faster. Did it have three parts that could run without each other, or did you pay three times to watch two people wait?

@mmoklaaa a stop-doing list only works when you name who gets disappointed by it

NASA's administrator stood in front of investors and didn't ask for money. All-In Summit, Los Angeles, mid-September. Jared Isaacman had a room built for asking, and he opened by taking the ask off the table: NASA has 25 billion dollars, and that is enough to build some pretty incredible hardware. What everyone carried home was the Moon. Artemis III next summer, Artemis IV in 2028, a nuclear reactor called SR-1 Freedom meant to push a ship toward Mars, China landing by 2030. Loud, quotable, and it's the part that got written up everywhere. The line that explains him came earlier and he dropped it as a diagnosis, not as a method. NASA's resources, in his words, "were spread everywhere, trying to make everyone happy." Districts, contractors, anyone with a phone number. So his fix isn't more. It's subtraction first. NASA walks out of launch and communications, because those have customers and margins now and somebody else will do them anyway. Whatever survives that cut goes into the missions with no obvious business case. Run it as a sequence and it reverses how most people plan. The usual order is: here is my time, what do I add. His order is: what am I doing that the market already does better and cheaper without me, cut it, then look at what's left. That's my reading, not his framing. He never says subtract first. He describes it as the agency's new job description and moves on. And the honest half. Subtraction is free. Anyone with ten hours a week spread across five projects to avoid disappointing five people can do that tonight. The second half isn't free. He can chase the impossible because he has $25 billion and a presidential appointment, and he can annoy congressional districts because his job doesn't depend on them liking him. Yours probably does. The version that survives without his leverage is smaller: not what nobody can do, but what nobody else has a reason to bother doing in your specific corner. Two moments worth going back to in the recording. 12:40 — a dark aircraft silhouette slides onto the screen behind him during the X-plane section, and he never mentions it. Not then, not later. 28:39 — Chamath Palihapitiya, hosting, finally asks about it, and the answer is "You sure that's a Blackbird?" The one thing he refused to explain got more headlines in 24 hours than the nuclear program he explained in detail. Which of the five things you're keeping alive is somebody else already doing better, cheaper, and without needing you?

@chenzhangreal right, and that's exactly the window. year two you're a keyword match

OpenAI cofounder just gave away a test that replaces your resume. Andrej Karpathy, Sequoia's AI Ascent in San Francisco, April 20 of this year. Twenty-nine minutes with Stephanie Zhan, and the sentence clipped everywhere was the one already printed in the talk's title: vibe coding raised the floor, agentic engineering raises the ceiling. Accurate, memorable, and useless to anyone trying to get hired. The part worth having comes at 18:19, when he's asked how he would interview an engineer now. "Give me a really big project and see someone implement that big project." Then the specifics: build a Twitter-style site whose users are AI agents, deploy it, and he will point agents at it and try to knock it over. They shouldn't be able to. He is describing a filter. The inverse is the useful reading, and he never makes it, because he's the one holding the job rather than looking for one. Inverted, it reads like this. Your credential is one thing you built, then attacked on purpose with the same class of tool that built it. It held. That's the whole claim, and it is not a degree or a whiteboard hour. The weak joint sits inside his own answer, and it's where this gets interesting instead of inspirational. A model hunting for holes in code a model wrote is not an independent auditor. It works anyway because the attacking agent carries none of the choices the building agent made and has no stake in defending them, so it probes what a design assumes rather than what its author intended. That's a fresh pair of eyes. It is not security expertise, and confusing the two will cost somebody a production incident. Whether the principle survives depends on who opens the artifact. Karpathy is filtering for elite AI roles, where a human actually clicks the link. At a company screening applications on keywords, nothing you built gets seen, and this does nothing for you there. It works sideways: a referral, a founder who reads what you posted, a team small enough to have no filter between you and an engineer. That's the honest scope, and he had no reason to state it. There's a service in the same paragraph for anyone who'd rather sell than apply, since non-technical founders are shipping agent-built products right now with no idea where they leak. Run the attack on their app, return the list of places it falls over, charge a fixed fee. Interviews didn't get harder. They moved. The evaluation now happens on some weekend months before anybody reads your name, and most people are still preparing for the room.

@0x_Vexly wholesome undersells it. a guy who built his whole identity on one skill just ranked a duller one above it, on record

Buffett just told you why he trusts his own successor. You read the wrong sentence. Last week Buffett stepped down as Berkshire's chairman after sixty years in the seat. His son Howard now holds the chairman title. Greg Abel, the man actually running the company day to day, is who Buffett was talking about. The real succession decision happened eight months earlier, in a CNBC interview nobody replayed. Every outlet ran the same line from it. He'd rather have Greg handling his money than any top CEO in the country. Great soundbite. Not the point. The line right before it is the actual reason. "Greg's operated more than I have when you get right down to it." Abel had been CEO for exactly one day. Read it again. The most famous stock picker alive just ranked a different skill above his own signature one. Sixty years of Buffett headlines are about capital allocation. Abel got the job for something duller. Seven years running every non-insurance business Berkshire owns, making calls nobody puts in a documentary. Control doesn't go to whoever mirrors your best trick. It goes to whoever logged the most hours making real decisions, in the part of the business nobody puts in a profile piece. You don't need a trillion-dollar company to use that filter. Next time you're picking a deputy, skip whoever's best at your signature move and count decision hours instead. That filter is worth packaging as a five-minute hiring checklist for any founder about to make their first ops hire. Would you actually trust the operator over the star, or only when Buffett says it's fine.

Fifteen years at Apple, and Tim Cook's whole strategy was one 1997 line. Cook just stepped down after running the company for a decade and a half. In his farewell interview with Esquire this spring, he put the whole run in nine words. "We say no to a thousand things to get to that one thing." You can't spread your energy like a peanut-butter spread, he said. That's not new wisdom. Steve Jobs told a room of developers the same thing in 1997, saying focus means turning down the hundred other good ideas in the room. Cook didn't write the law. He spent fifteen years being the one person patient enough to enforce it after Jobs was gone. Enforcing it still cost him. Apple burned ten years and more than ten billion dollars before it finally killed the Apple Car, the project his own filter should have caught early. The rule only pays off applied before the ten billion, not quoted after. Cook needed a decade next to Jobs before that sentence turned into instinct. What are you skipping to go straight for the line.

Wells Fargo's CEO blamed 5,300 low-level employees. What about his own stock sales? Ten years ago today, Senator Elizabeth Warren asked John Stumpf that exact question, and he never really answered it. Wells Fargo had been firing roughly three employees a day, five straight years running, chasing an internal target called "Eight is Great" that pushed every customer toward eight accounts, almost three times what surveys found the average bank customer actually held. Nobody wrote "open a fake account" into the policy manual. The quota did that work on its own. That's the detail people skip while quoting the apology. Two days after the hearing, analysts pegged Stumpf's total exit package, three decades of stock, options and bonuses, at close to $200 million, and his defense to the Senate still came down to 5,300 "bad apples" on the sales floor. A week after this hearing, the board clawed back $41 million from him anyway. The lesson was never about compliance training. One number turned into the entire scorecard, and people hit it however the manual forgot to forbid. Any founder handing a team a single KPI this quarter is running the exact same experiment, just smaller and without cameras in the room.

@KostyaAI salesforce hasn't published that number, only cost and headcount. tells you which metric they wanted you to see

Benioff's own 2025 line just resurfaced, and it wrecks his math. September 2025, The Logan Bartlett Show podcast. He said it plainly. "I need less heads." Support team went from 9,000 to about 5,000. Half of tickets now go to an AI agent instead of a human, versus zero a year before that. Costs down 17%. This week, different stage, same guy. Telling reporters Salesforce has more employees than it's ever had, over 83,000 total. Both statements are true. Nobody put them next to each other until now. Total headcount hides which department got hollowed out. You don't lie with a number that big. You just answer a question nobody actually asked. In that same interview, he told the AI industry to regulate itself before the lawsuits arrive, borrowing a Hawaiian word, kuleana, for personal responsibility. Sounds like ethics. He said all that three weeks after Salesforce reported its lowest attrition in years, longest contracts ever, and Agentforce revenue past $1.5 billion, up 240%. Trust isn't the message here. Trust is the line item. You already know which sentence a CEO reaches for when a number needs defending. He picked his on camera, twice, twelve months apart.

this is the least talked about layoff story of the year

Verizon's CEO fired 13,000 people and predicts a far bigger wave. Dan Schulman spent nine years running PayPal before Verizon called him in October last year. He later told Harvard Business Review the company "had a lot to be proud of, but was resistant to change." Within months he cut 13,000 jobs, about 13% of the workforce, in one stretch instead of four quarters of layoff rumors bleeding morale dry. Almost every recap stopped at that number. He'd rather everyone know the worst on day one than spend a year bracing for it. Schulman told the Journal in April that blunt beats "happy talk." He backed it with a $20 million retraining fund, about $1,500 per laid-off worker, calling it "an obligation of every CEO," not charity. He also warned automation could push unemployment to 20-30% within two to five years. Coming from a man who had just cut his own staff by 13%, that sounds less like a forecast and more like a confession. Nine months in, Verizon stock is up 29% this year, the only telecom beating the S&P 500. Guidance raised twice. Nobody rewarded the cost cutting itself. They rewarded a CEO who told you the ugly number before the earnings call could force it out of him.

One traffic jam built Richard Branson's next CEO. In 2005, Branson double-booked his own interview. Instead of cancelling on the 31-year-old lawyer sitting across from him, he brought the guy along for the ride to his next meeting. Traffic turned that ride into two and a half hours. Josh Bayliss spent it asking Branson questions instead of answering his. He got the job as Virgin's general counsel that same year. Six years later, he ran the entire Virgin Group as CEO. Everyone reads this as a luck story. It isn't. Branson didn't hire Bayliss for a sharp answer inside a scheduled 45-minute slot. Most candidates never get an unscheduled 2.5 hours alone with the person who can actually hire them. Bayliss got it by accident and used it to listen, not perform. Principle's actually simple. Unscripted time beats a rehearsed answer, every time it's available. AI can already write a perfect interview answer in ten seconds. That makes real, unscripted access to a decision-maker newly rare, because no script can fake it. Recruiters who build that into the hiring process are selling something no AI-prepped candidate can out-prepare. When's the last time you had two unscripted hours with someone who could actually change your career?

six years from general counsel to ceo of the whole group is still a wild jump even with the traffic jam story