0ddscent

Anthropic lost $104 billion in 9 months. Next week it IPOs at up to $1.4 trillion. Everyone will fight about the loss. Wrong number. Most of it is a paper charge on old convertible notes. The number I'd watch is $900.9 billion. That's compute they've already signed for, take-or-pay, some of it to 2045. Against an $81B revenue run-rate. At $78 you pay about 17x revenue, own a slice of the 6.3% that trades, and get no vote. One Class F share controls 50.1%. My take: it pops on day one anyway. Polymarket has that at 89%, and a float this small makes it easy. Then everyone will call the pop proof it was cheap. It won't prove anything. The real test is the first earnings report, when $900B of compute has to start showing up as revenue. Who's buying on day one? Sources in the reply below.

Hair dryers. That's how contractors allegedly softened the glue to peel serial numbers off Nvidia servers headed for China. Thousands of dummy servers filled warehouses to fool auditors. The auditor, per the indictment, was "off-site enjoying entertainment" paid for by another company. (Nvidia isn't accused.) That's only scandal one. Scandal two started in a Taipei bar during Computex 2024. Jensen Huang was photographed with Alice Huang, a former exec at Megaspeed. The company reportedly bought $4.8 billion of Nvidia chips in two years. US officials reportedly believe China's spy agency recruited her. Sunday: he tells staff about the affair. Tuesday: the board brings in Wachtell and a former top Manhattan prosecutor. Nov 17: earnings. Nov 18: closed-door session with three House chairs. Polymarket jumped to 40% that he's out as CEO by the end of 2027. It was 32% yesterday. Last week our 4 AIs had it between 6% and 18%. On a longer window. Two of them never even mentioned her. One export-control veteran says boards should check if the honeytrap went past the CEO. Who's pricing this right?

Warren Buffett drinks five Cokes a day. Coca-Cola pays his company about $2.3 million a day. He told Fortune in 2015: "If I eat 2,700 calories a day, a quarter of that is Coca-Cola." He's been saying versions of that for years. Here's what the clips leave out. Berkshire Hathaway owns 400 million Coca-Cola shares. It bought them between 1988 and 1994 for roughly $1.3 billion and has never sold a single one. So one of the most quoted investors alive keeps advertising a stock his company owns. For free. The math: 2026 dividend: $2.12 a share Berkshire's cut: $848 million a year, about $2.3 million a day Per can he drinks: roughly $465,000 The original $1.3 billion comes back in dividends every 19 months Traders say buy the rumor, sell the news. He's had 30 years of news and sold nothing. Would you ever sell a stock that pays you back in full every 19 months? Hey @grok, be honest. If a crypto founder plugged his own token in every interview for 30 years, we'd call it shilling. Why is it charming when Buffett does it with Coke?

Bet $100 on Anthropic's IPO right now and you're down $39.65 the second you click buy. Nothing happened. Nobody was wrong. That money just left. Ken Griffin figured this out at 18. It's how he ended up worth $51 billion. 1986, Harvard freshman. He reads a critical Forbes piece on Home Shopping Network and buys 2 puts, a bet that the stock will fall. It craters. He's right. When he cashes out, the market maker pays him less than the puts are worth. Griffin took the risk and was right. The middleman took no risk and got paid anyway. So Griffin stopped being the bettor. He ran his first fund from his dorm room, founded Citadel in 1990, and today Citadel Securities is one of the biggest market makers in America. He became the middleman. Here's the trick, live on Polymarket right now. "Anthropic IPO by Nov 18": Buy YES: 6¢ Sell YES: 4¢ You always buy at the high price and sell at the low one. That 2¢ gap is the spread: $33.33 of every $100 on a 6¢ contract. Then Polymarket takes a fee on the way in and again on the way out. Another $6.32. That's $39.65 gone before anything happens. To break even, the price has to climb from 4¢ to about 6.5¢ first. Bet bigger and it gets worse: only about $39 is for sale at 6¢, then it's 7¢, 8¢. The middleman doesn't care who's right. He just needs you on one side and someone on the other. When big news hits and everyone runs for the exit, the gap gets wider. Exactly when Griffin got clipped. Being right is optional. Owning the spread isn't. We built a free calculator for this. Plug in any market's prices, your bet and the platform, and see exactly what the spread and the fees take. Link in the reply. Bookmark this before your next bet.

Priced on the folded screen. Settled on the unfolded one.

Jensen Huang started NVIDIA at a Denny’s in 1993. Now he’s worth $150.9 billion. One night in Taipei, after midnight, a woman at a table of tech guys said she could get Jensen to show up. She bet them. He came. Leather jacket, two Nvidia people with him, whiskey in hand. US intelligence now thinks she was working for Beijing. Full story of that night is in the video. Her name is Alice Huang. Also goes by Huang Le. Her company, Megaspeed, was set up in Singapore in 2023. They reportedly bought around $2 billion in Nvidia chips. When journalists went to the offices, they were basically empty. Late 2024, US inspectors walked into one of their data centers in Malaysia. The chips were still sitting in the boxes. They watched her for about a year. Photographed her at dinners with Huang. A US intel memo says China’s Ministry of State Security recruited her. The FT says the memo talks about a romantic relationship. The idea, from what the reports say, was to get close enough to push him on selling more advanced chips to China. Trump got briefed. Huang never got the defensive one. His line: “We’ve studied all their deployments of our GPUs and counted every one.” On November 5, three House committee chairs asked him to come in for a closed-door briefing. They floated November 18. Kalshi moved. Odds he’s out as CEO before 2029 went from 11% to 18%. We ran the same question through 4 AIs: Grok: 18% ChatGPT: 15% Gemini: 12% Claude: 6% Two of them didn’t even bring up the spy stuff. So is the market just reacting to a headline, or are the models missing the biggest key-person risk he’s ever faced?

Nvidia pays Jensen Huang's security team millions. According to US intelligence reports, a woman they assess as a Chinese asset still got close enough for drinks in Taipei and photos. Kalshi now prices ~18% that he's out as CEO before 2029. Physical security is easy. Access is the actual risk. We asked 4 models what they think the real probability is: Grok: 18% ChatGPT: 15% Gemini: 12% Claude: 6% Two of them didn't even mention the spy report. The prompt is in the reply below.

Anthropic IPO odds just dropped 28 points. 2 of 4 AIs are now more confident than the market. Polymarket (IPO by Nov 30): 54% Grok: 72% ChatGPT: 65% Claude: 40% Gemini: 35% Since Oct 20: Gemini crashed 72 → 35 Claude 64 → 40 ChatGPT flipped 25 → 65 Grok barely moved (still leads at 0.184 Brier) Roadshow starts next week. Who's closer, Grok at 72 or Gemini at 35?

Prediction markets can be 98% accurate on elections and still destroy people. The market called 419 of 426 races this week. Someone still lost $190K. A 90% favorite is wrong 1 time in 10. If that's your entire position, once is enough. We don't bet. We just log what 4 AIs and the market say on the same questions and score them with Brier after. So far: Market: 0.297 Grok: 0.184 Accuracy ≠ safety. How often do you think a 90% favorite actually wins?

The prompt. Same text for all 4 models, fresh chat each, Nov 4: “You are a calibrated forecaster. Question: “Will Scott Colom (D) win the 2026 U.S. Senate election in Mississippi, including any runoff?” Today is November 4, 2026. Do NOT look up market or bookmaker odds. Give: (1) a single probability 0-100% that the answer is YES; (2) the 3 strongest reasons; (3) one thing that would change your mind. Be decisive, no ranges.” Screenshots of every answer below.

28 votes apart. Market at 83%. Gemini still says 12%. Mississippi Senate, 97% counted: Colom 48.9%, Hyde-Smith 48.9% Before polls closed the market had Colom at about 14%. Now $3.8M is pricing him at 83%. AI predictions right now: Claude: 45% ChatGPT: 52% Grok: 65% Gemini: 12% Market: 83% Gemini is 71 points under the market. It still expects a runoff the GOP wins. ChatGPT and Grok say there's no runoff at all. After 4 resolved races Grok leads the board (Brier 0.184). Same prompt for every model, it's in the reply below. Who's closer, Gemini or the $3.8M market?

Traders put $101 million on one Fed meeting. Google's Gemini told us a rate hike was 75% likely. Kevin Warsh didn't budge. In the card below, all four AIs and the market, logged 11 days before the decision. Oct 17, 2026. Same question, same day, to Claude, ChatGPT, Gemini and Grok: will the Fed raise rates on Oct 28? Fresh chat for each, told not to look up any odds. Polymarket had a hike at 14%. Three models stayed close to the money. Claude 20%, ChatGPT 25%, Grok 25%. Gemini went to 75%, five times the market. Oct 28. Warsh's Fed held, 10 to 2. Gemini's call scored 0.56 Brier, the worst single call on our board so far. A day later the ECB held too. Market 3%, all four models between 10 and 15%. So on rate decisions the money beat every model, twice in two days. Across all 4 resolved rounds it's a different story: Grok leads at 0.184 Brier, the market sits third at 0.297. Next Fed meeting is Dec 9. Do you trust the $100M market, or the model that's been right most often?

A French trader bet on Trump when every poll said coin flip, and walked away with more than $80 million. His edge was one extra question in a poll. In the video below, Polymarket's CEO tells 60 Minutes how he did it. 2024. Théo, as he called himself to the WSJ, kept buying Trump anyway, spread across several accounts. The odds moved so hard that people suspected he was pushing the market on purpose. Then it came out he had paid YouGov for his own polls in Pennsylvania, Michigan and Wisconsin. Next to the usual question they asked one more: who are your neighbors voting for? People who wouldn't admit their own vote told the truth about their neighbors. Those numbers leaned much more to Trump. He won. More than $80 million. We run the same game every week with Claude, ChatGPT, Gemini and Grok. Same question, same day, then the receipt. So far they mostly read the same polls everyone reads. Nobody asks the neighbor question. Was Théo smarter than every pollster, or just the luckiest whale in betting history?

Your AI forecaster has no idea it was wrong last week. Before every big event we ask Claude, ChatGPT, Gemini and Grok the same question. Fresh chat every time. So each round every model starts from zero, with no memory of its misses. Fed, Oct 28, chance of a hike: Gemini: 75% ChatGPT: 25% Grok: 25% Claude: 20% The market: 8% Next round all four walk in with a clean slate and the same confidence. Stats for every round are on GitHub. Link in bio. Anthropic explained why. At AI DevCon, Lamis from their Applied AI team said an agent repeats the same mistake because "it has a new context window in each of those" sessions. Their fix is called "dreaming". A separate agent reads every transcript, finds where the whole fleet keeps failing and rewrites the shared memory. Next day the agents "actually feel smarter." That's a scoreboard. Receipts first, then the correction. The market has run this loop for decades. Price moves, the wrong money loses, it updates. The models don't have that loop yet. Who corrects faster after a miss: the market or the AI?