Casper

$1.5 billion. 8 years. Over 1,000 people. And not one pixel of this was generated by AI. Take-Two's CEO called an AI-made GTA "laughable" and kept generative AI out of the entire game. Every building, every street, every wave was placed by a human. GTA 5 grossed roughly $10 billion, the most profitable entertainment product in history. GTA 6 will likely beat it, with zero AI in the product. In a world where anyone can generate average in seconds, the most valuable thing on earth is still what humans made by hand.

Warren Buffett, asked the secret behind his success: "You have to be just plain lucky. My great skill has been avoiding bad luck, but that isn't a skill. That's luck." The most successful investor in history, the man the whole world calls a genius, says the honest answer is luck. Right country, right decade, and every drunk driver that could have ended it and never did. That's the entire point of Fooled by Randomness. The people who mistake luck for skill blow up. The one who compounds for 60 years is the one humble enough to admit how much of it was luck. The edge was never certainty. It's the doubt.

Nobel laureate Robert Merton on how to actually fund a retirement: 4:41 – why it's a challenge, not a crisis 11:04 – the only 3 real ways to fund retirement 14:00 – why "stocks are safe in the long run" is a wish, not a policy 23:18 – measure your retirement in income, not a pot of money Two hours from a Nobel Prize winner that reframes how you think about money. Watch it, then read the breakdown in the article below.

@Rossst_03 Very smart guy

@caspr_exe wow, so interesting! saved, man!

Eric Rosenfeld, co-founder of Long-Term Capital Management: "We should have run this at a lower risk." That is the understatement of the century, and it is the whole lesson in one sentence. LTCM was supposed to be unbeatable. Two Nobel laureates, Myron Scholes and Robert Merton, the men who literally won the prize for pricing risk, sat on the board. Their models said the trades were almost riskless. For three years the fund printed around 40% a year, and the smartest money on earth begged to get in. Then the improbable landed. Russia defaulted, the correlations that were supposed to cancel out all moved the same way at once, and the same math that made them look like gods turned on them. The fund lost $553 million in a single day. Roughly $4.6 billion evaporated in weeks. It was so big and so leveraged that the Federal Reserve had to organize a bailout to stop it from dragging the whole system down. None of them were stupid. That was the point. Being the smartest people in the room is exactly what let them mistake a few calm years for a permanent edge. Their model was fit to a market that had never been stressed, and when the stress came, the edge turned out to be luck all along. The market does not punish the dumb. It punishes the most certain. The edge was never the model, the returns, or the Nobel prize. It was the doubt they never had.

Marcos Lopez de Prado, Cornell professor, at London Business School on how machines quietly make money off human fear and greed: 2:33 – reading investor overconfidence straight from news sentiment 7:46 – 20,000 tweets a day, and how their imbalance predicts Tesla's next move 10:03 – the 2010 flash crash, explained as pure human panic 12:28 – what actually stopped it: cold machines bidding with no fear 14:54 – why the whole game is bet sizing, not being right 18:50 – meta-labeling: a second AI that decides how much to bet The market doesn't reward the smartest human. It quietly moves money from the fearful to the machines that don't panic. Watch it, then read how machines exploit human bias in the article below.

@Rossst_03 most people mistake luck for skill and never even stop to think about it.

@caesar_aii fast enough that the workflow stops feeling rented. that's the whole pitch

THIS MAC STUDIO M4 MAX TURNED A LOCAL AI PC TEST AGAINST THE M3 ULTRA INTO A BUYING PROBLEM A guy sets up two Apple boxes in LM Studio and runs the same model on both. Mac Studio M3 Ultra on one side. Mac Studio M4 Max on the other. Same local chat. 48 GPU offload. A small “hi” warm-up before the real run. This is where the benchmark should become obvious. The bigger machine should walk away, justify the price, and make the smaller one feel like the wrong desk computer for local AI. Instead, the test gets annoying in a useful way. One run sits around 42-45 tokens/sec. Another jumps around 49-53 tokens/sec. The M3 Ultra is faster, but not in the clean “forget the M4 Max exists” way people expect when they shop by chip tier. That is the local AI PC decision in one screen. If your day is small and medium models, private prompts, client code, research notes, localhost agents, and dumb experiments you do not want metered by an API, the M4 Max-class box starts to look like the quiet default. Not because it wins every chart. Because it is fast enough that the workflow stops feeling rented. The M3 Ultra still has the better argument when memory headroom matters: bigger models, bigger context, more parallel work, fewer moments where the machine starts acting like unified memory is the whole product. But the video kills the lazy version of the purchase. You are not buying a local AI PC only for peak tokens/sec. You are buying the point where privacy, rate limits, cloud bills, noise, thermals, desk space, and waiting all become one machine decision. M4 Max is the efficient private AI desktop. M3 Ultra is the heavier local model workstation. The uncomfortable part is that for a lot of daily local LLM work, those are closer than the spec sheet wants them to be.

@caesar_aii absolutely right

@caspr_exe drilling those pure data structures early is the real cheat code for building long term

Want to know how people earn $500k a year writing code? It starts exactly here. While everyone argues AI killed coding, this kid is silently grinding LeetCode in Python. 334 problems solved by hand, no AI, no talking. Data structures and algorithms, the exact thing he's drilling, is the gate every $500k engineer at Jane Street, Citadel and Google had to pass. He's just starting a decade early. The fundamentals aren't dead. Someone's quietly compounding them while the rest of us debate.

@0xRokko amazing animation!!

This is every named chess opening ever studied, ~3,700 of them, rendered as one force-directed graph. A map of where human theory has actually been. Your AI's memory should look exactly like this. Not a giant transcript of every session, but a graph of compressed lessons, each one linked, each one recalled only when it's relevant. That's why a graph beats a log: a log gets noisier as it grows, a graph gets smarter, because every new node connects into the web and makes the rest more useful. The model was never the moat. The graph of what it has learned is.

Tejas Kumar, AI engineer at IBM, just explained AI harnesses from scratch, live-coding one in 18 minutes: 1:48 – why a harness beats prompting a black-box model 4:08 – the 5 parts of every harness: tools, model, context, guardrails, verify 6:01 – live demo: making a bad 2023 model do a real task 7:03 – the agent fails, then lies that it succeeded 13:07 – the verify step that catches the lie 15:39 – deterministic login, handled by the harness not the agent 19:10 – why 2026 is the year of harnesses 18 minutes that replace 20 paid agent courses. He never touched the prompt once. Only the harness. The outcome flipped. Watch it today, then read how to build your first AI loop in the article below.

@sairahul1 nice article

@caspr_exe oh nice. thanks for this new video