GROK + JEV + OPUS 5.5 checked a puzzle where one boring detail, "born on a Tuesday", changes the answer... a family has two kids. one of them is a boy born on a Tuesday. what's the chance both are boys? most people say 1/2. the day of the week can't matter, right? question → explain → check → simulate → animate Grok took the explanation: the Tuesday boy can be the first kid or the second. two boys give him two ways to exist, a boy and a girl only one Jev checked the math: each kid is one of 14 types, boy or girl × 7 days families with a Tuesday boy: 14 + 14 − 1 = 27 of those, both boys: 7 + 7 − 1 = 13 13/27 = 48.1% drop the Tuesday and it's 1/3 Opus 5.5 simulated 1,000,000 families 100 families: 61.5% 1,000: 47.5% 10,000: 45.8% 100,000: 48.1% 1,000,000: 48.4% the weirder part: the more specific the detail, the closer it gets to 1/2 "one is a boy": 33.3% "born on a Tuesday": 48.1% "born in July": 48.9% "born on July 17": 49.97% where this actually shows up: how you found out matters as much as what you found out. "I asked if there's a Tuesday boy and heard yes" gives 13/27. "a parent told me about one of their kids at random" gives 1/2 that gap is why people still argue about this one, 16 years after puzzle designer Gary Foshee posed it 27 families, 13 with two boys, 1,000,000 simulated ↓ save this

GROK + JEV + OPUS 5.5 checked a treatment that wins in every group and still loses overall, and the data is real... 1986, kidney stones, 700 patients. treatment A vs treatment B small stones: A 93%, B 87% large stones: A 73%, B 69% all patients: A 78%, B 83% A wins both groups. B wins the total question → explain → check → recompute → animate Grok took the explanation: doctors gave A the hard cases. 75% of A's patients had large stones, only 23% of B's Jev checked the math: an overall rate is a weighted average A = 0.25 × 93% + 0.75 × 73% = 78% B = 0.77 × 87% + 0.23 × 69% = 83% Opus 5.5 recomputed it with the same mix for both, half small and half large A: 83.1% B: 77.7% same patients, fair weights, A wins by 5.4 points where this actually shows up: Berkeley 1973: 44% of men admitted, 35% of women. split by department, women did as well or better in most. they applied to the hardest ones your dashboard: conversion up in every channel, down in total, because traffic moved to the weak channel sports: a player can have a higher average in each season and a lower one across both 700 patients, two groups, one total that points the wrong way ↓ save this

GROK + JEV + OPUS 5.5 checked why your followers seem to have more followers than you, and the math says they almost always do... it's not that you're unpopular. it's who shows up in your list question → explain → check → simulate → animate Grok took the explanation: popular people sit in a lot of friend lists. so when you look at your friends, you keep running into them Jev checked the math: friends' average = average + variance / average. any spread at all and it beats your average Opus 5.5 built a network of 10,000 people and counted every connection average person: 6.0 friends average friend of that person: 22.4 friends people with fewer friends than their friends: 90.2% the biggest node: 407 friends the top 1% of people hold 11.2% of all connections where this actually shows up: on Twitter, 98% of users have fewer followers than their followers do on average you feel behind because you compare yourself to the people everyone knows doctors use it in epidemics: ask random people to name a friend and test the friend. the friend is more connected, so you see the outbreak earlier 10,000 people, 90.2% "less popular" than their friends, math working as intended ↓ save this

GROK + JEV + OPUS 5.5 checked the math trick that wins bar bets, and 23 people is all it takes... a room with 23 people. what are the odds two of them share a birthday? most people guess under 10%. it's 50.7% question → explain → check → simulate → animate Grok took the explanation: you're not matching birthdays to yours. every pair counts, and 23 people make 253 pairs Jev checked the math: P = 1 − (365 × 364 × … × 343) / 365²³ = 50.7% Opus 5.5 built 10,000 rooms of 23 random birthdays and counted the matches 10 rooms: 6 had a match 100 rooms: 52 1,000 rooms: 499 10,000 rooms: 5,009 push the room size and it gets silly: 30 people: 70.6% 50 people: 97.0% 70 people: 99.9% where this actually shows up: 2014 World Cup, 32 squads of 23 players. 16 of them had two players with the same birthday a class of 30? better than 2 in 3 that two kids share a birthday the same math sets the size of the hashes behind your passwords. a collision takes only half the bits in tries, so they're 256 bits, not 128 10,000 rooms, 5,009 matches, 23 people ↓ save this

GROK + JEV + OPUS 5.5 broke down the most argued-about math problem on the internet, and the numbers don't lie... 3 doors, 1 car, 2 goats. you pick one, the host opens a goat. stay or switch? question → explain → check → simulate → animate Grok took the explanation: your first pick wins 1 time in 3, and the host never touches the car Jev checked the math: P(stay) = 1/3, P(switch) = 2/3 Opus 5.5 ran 10,000 games and wrote the animation 10 games: switching won 7 100 games: 60 1,000 games: 663 10,000 games: 6,636 then 100 doors. you pick 1, the host opens 98 goats. nobody keeps their door at 1 in 100 where this actually shows up: Deal or No Deal looks the same, but you open the cases yourself. random reveal, no edge, it's 50/50 someone who knows the answer cuts your options? the one they skipped is the one to look at most people still keep their first door after hearing the math. same reason traders hold a bad position 3 agents, 10,000 games, 6,636 wins for switching ↓ save this

🚨 GPT-6.1 ASTRA IS SCARY GOOD AT GAUSSIAN SPLATS Asked it to rebuild one of the new ChatGPT agent dots using Blender MCP plus Lichtfeld Studio, this is what came back Every strand of fur holds up from every angle, not just the hero shot, that's usually where splats fall apart Looks razor sharp in the SuperSplat web editor too, no blur, no floaters, no mushy edges Put this next to what Gemini 4 Argon's been doing lately and the gap is obvious, Argon leans toward big environments and scale, this is the opposite skill, nailing one small object down to the individual hair Different kind of flex, same question either way, how long until this is just what "making a 3D asset" means Disclaimer: this footage may be staged, edited, or AI-generated, its origin is unconfirme

A 16 year old in China told his mom he needed a monitor at the fruit stall for an inventory spreadsheet. Mom said fine. $220 ultrawide, secondhand. He never opened a spreadsheet. He ran chip layouts for IoT startups between customers buying mangoes. He opened Claude Opus 5.5 and typed ai inference accelerator for keyword spotting, 180nm, 2x2mm tile, int8 mac array 32 wide 32 deep, floorplan with openroad, write the verilog, flag every drc risk, no handwaving. Opus wrote 1,800 lines of Verilog and the openroad config. The design routed cleanly on the second pass. By Sunday he had three accelerators and a tiny image sensor driver sitting in his portfolio. A Berlin IoT startup building a low power smart garden sensor DMed him on Discord. They needed a custom always on keyword spotting block. A freelance ASIC designer had quoted them $45,000 and 4 months. He sent them a working GDSII in 10 days. Paid $1,800. They taped out at the next Tiny Tapeout window. The block came back working on the first silicon. Now he does chip layouts for 4 IoT startups and one indie VR headset company. Low power vision tiles, tiny speech blocks, a wake word accelerator, a sensor fusion front end. Opus writes the Verilog and the floorplan scripts. He signs off the design. Never opens Cadence. 7 to 12 days per block. Month one $4,200. Month three $12,000. Month six $19,000. Spend on tools that month $148. His mom walked past the monitor between customers. Saw a yellow and blue city map covered in grid lines. Asked if the inventory system was ready yet. He said almost. She said the mangoes are ripening fast. A freelance ASIC designer in Berlin bills $45,000 for a wake word block and takes 4 months. He bills $1,800 and takes 10 days. His physics teacher asked when he was going to turn in the circuit lab writeup. He said this weekend. She said good, chip design is a serious career if he's interested.

OpenAI's new Dots keep working after you close the laptop. But most people set them up wrong. My 3 locks, mistake log and 30-day plan are in the article 👇

A 15 year old in Belgium told his dad he needed a used Vision Pro for a school AR project. Dad found one with a cracked light seal. $780. He never submitted the school project. He used the headset to build spatial mini-apps that print real objects out of volumetric cartridges. He opened Claude Opus 5.5 and typed visionOS mini-app, floating 1998 gameboy running pokemon red, when the player hits a print button the active pokemon mesh ships to the user's bambu lab printer over the network, print progress as an ar card next to the printer, swift ui, under 40mb. Opus wrote 2,400 lines of Swift. The build compiled first try. The first thing he printed was a Bulbasaur. By the end of the week he had 6 mini-apps. A volumetric Game Boy. A card scanner that pins the card's story in the air above the sleeve. An AR shelf planner for figurine collectors. A tabletop battle map that fires haptics on real dice rolls. A Czech 3D printed minis shop that sells on Etsy DMed him on Discord. Their return rate was 12 percent because customers couldn't picture a 28mm model on their terrain before buying. He built them a scan to AR preview app in 7 days. Paid $1,400. Their return rate dropped to 3 percent in a month. They asked for a second app for armies. Now he builds visionOS mini-apps for hobbyist 3D print shops, collectible stores, and three indie TCG makers. Opus writes the Swift and the Reality Composer scenes. He wires in the hardware. Never opens Xcode. 4 to 6 days per app. Month one $4,100. Month three $11,500. Month six $18,000. Spend on tools that month $184. His dad walked in during a test. Saw a Game Boy floating over the kitchen table and a Hatsune Miku in the corner that wasn't actually there. Asked when the school project was due. He said next week. Dad asked why there was a Pokemon on the table. He said that's part of it. An Apple partner studio in London bills $45,000 for a visionOS retail demo and takes 10 weeks. He bills $1,400 and takes 6 days. His teacher asked for the AR project update. He said he was polishing the final build. She said good, most students are still on the proposal.

A 16 year old in Switzerland told his mom he needed a used laptop for a drafting class CAD unit. Mom bought a gaming laptop with a crack in the lid. $220. He never installed CAD. He used the laptop to build interactive 3D houses that architects pitch to their clients. He opened Claude Opus 5.5 and typed two story modern house, 240 square meters, pool, double height living room, folded plate steel stair, limestone rainscreen, service core, label every system, let the user toggle plumbing electrical hvac drainage, add an explode slider, three.js, mobile 60fps. Opus wrote 1,200 lines. He deployed to Cloudflare Pages. The explode slider worked on the first click. By the end of the weekend he had 4 houses, 2 duplex stacks, and a townhouse row. A 3 person architecture firm in Zurich DMed him on Instagram. They were pitching a $4.2 million villa the following Tuesday. They had been quoted $18,500 and 4 weeks by a local viz studio. He sent them a working interactive link in 4 days. Charged $1,800. The client signed the next Monday and told the firm the explode toggle was what did it. Now he supplies interactive pitches to 6 small firms and 2 custom home builders. Villas, cabins, three warehouse conversions, one chapel. Opus writes the Three.js and the UI. He tunes the materials. Never opens Rhino. 3 days per house. Month one $4,200. Month three $11,000. Month six $19,000. Spend on tools that month $164. His dad walked into his room during a preview. Saw the house exploded with every pipe and wire floating. Asked if it was for Minecraft. He said school project. Dad said the detail is really good. A Zurich arch viz studio bills $18,500 per interactive pitch and takes 4 weeks. He bills $1,800 and takes 4 days. His drafting teacher asked when he was starting the CAD unit. He said next week. She said good, most students put it off.

A 14 year old in Finland told his dad he needed a new smart TV for family movie nights. Dad bought a used LG for the living room. $340. The family never watched a movie on it. It has played a koi pond loop 22 hours a day for 5 months. He opened Claude Sonnet 5.5 and typed koi pond top down, gerstner waves, cherry blossoms drifting, four koi in a boid school, dawn key light, loop cleanly at 60 seconds, three.js, one index.html. Sonnet wrote 800 lines. He dropped it in a Cloudflare Pages URL. The TV browser played it fullscreen. It loops without a seam. By the end of the week he had 22 different loops. Zen garden. Rain on a window. Aquarium. Firepit. Sakura in wind. Northern lights over a fjord. A boutique ryokan chain based in Amsterdam DMed him on Gumroad. 6 hotels, one custom lobby loop per property, each branded with the chain's sakura mark. Paid $4,200. Delivered in 4 days. Now he sells the same package to boutique hotels, spa lobbies, sushi restaurants, and one funeral home. Sonnet writes the shader. He tunes color. Never opens Blender. 6 hours per loop. He also sells 43 stock loops on Gumroad. $12 each. Dentists, meditation apps, two Twitch streamers, one guy running them on a wall of 9 TVs in his living room. Month one $4,600. Month three $10,000. Month six $17,000. Spend on tools that month $128. His dad walked into the living room during a preview. Asked when they were going to watch a real movie. He said soon. Dad asked why the fish were still there. He said it's not broken. A signage studio in London bills a boutique hotel chain $30,000 for a custom lobby content library and takes 4 months. He billed $4,200 and took 4 days. His mom asked why the electricity bill went up. He said it's the TV. She told him to turn it off at night. He said the fish were paying rent now.

Your AI is 99% accurate. That last 1% is about to get expensive. That remaining 1% looks harmless until you repeat the task. Suppose each request has a 99% chance of being correct, and errors happen independently. The probability of getting at least one wrong answer becomes: -10 requests: 9.6% -50 requests: 39.5% -100 requests: 63.4% -1000 requests: 99.996% The calculation is simple: P(at least one error) = 1 - 0.99ⁿ At 1000 requests, you’d expect around 10 errors on average. The model hasn’t become less accurate. You’ve given that small error rate more opportunities to show up. Now imagine those requests are steps in an automated workflow. The AI reads an invoice, extracts the amount, matches a customer and updates a record. A wrong amount could pass through the remaining steps without anything crashing. The system finishes. The result is wrong. If a task requires 100 correct steps, and any single mistake ruins the outcome with no recovery, its success rate under these assumptions is only 36.6%. That’s a simplified model, not a measured failure rate for every AI agent. Real errors can be related, steps vary in difficulty, and systems can catch and correct mistakes. Those differences are exactly why the surrounding software matters. Anthropic’s guidance on agents explicitly highlights the risk of compounding errors and the need for testing and safeguards. Validate critical outputs. Check totals with code. Make changes reversible. Review consequential actions before execution. 99% accurate tells me how often it gets things right. I also want to know what happens when it doesn’t.

A 15 year old in Hungary told his mom he needed a mechanical keyboard for school typing exams. Mom bought it. $80 secondhand. He never took a typing test. He used the keyboard to write animated 3D hero sections for boutique brands his mom has never heard of. He opened Claude Opus 5.5 and typed roman chariot race, three lanes, four horses at gallop, dust particles, sunlit colosseum, three.js, keep it under 60kb, mobile 60fps. Opus wrote 400 lines of Three.js. He dropped it in a Vercel deploy. It worked first render. By the end of the weekend he had a portfolio of 6 chariots, 4 spinning sneakers, 2 titanium bike frames, and a rotating perfume bottle. A boutique cycling brand in Copenhagen DMed him on his portfolio site. Their agency had quoted them $8,000 for a 3D hero and 6 weeks. He sent them a live link in 3 days. Charged $1,350. The brand's conversion rate on that page doubled the first week. They asked for two more. Now he sells animated 3D hero sections to DTC brands and small agencies. Cyclewear, watchmakers, indie fragrance, one Kickstarter for a folding kayak. Opus writes the Three.js. He tunes the camera. Never touches Blender. 2 days per hero. Month one $4,300. Month three $11,000. Month six $18,000. Spend on tools that month $147. His mom walked in during a preview. Saw the chariots running in a loop. Asked what game he was making. He said school project. Mom said stop wasting your evenings on games. A senior Three.js developer at a Copenhagen agency bills $8,000 per hero and takes 6 weeks. He bills $1,350 and takes 3 days. His typing teacher asked when he was going to take the practice test. He said next week. She said good, most kids don't take typing seriously anymore.

AMD is paying $8.2B for World Labs. World Labs builds models that understand and generate 3D worlds. After the deal closes, Fei-Fei Li becomes AMD's chief scientist. That's a lot of money for AMD to spend this far above the chip. It puts the people designing new AI workloads inside the same company designing the hardware they'll run on. Nvidia has spent years building up the stack. AMD just bought itself a much bigger piece of one.