Thesky

Web3 in a nutshell: ○ Use a good product ○ Become a regular ◉ Ask for the airdrop the second you like it

Gm My day starts with 2 pair of eyes 👀

Me soon 👀

X officially has its own AI Labubu now. Give it a few months and this thing will be on every SF engineer’s MacBook, backpack, and water bottle lol

Gm Alpha is Everywhere 👀

OpenAI says it solved a 90-year-old Millennium Prize problem in 88 hours. 10,000 AI agents. A Lean formalization. Navier–Stokes singularity. Then the mathematicians spoke up. Here’s the part to focus on: The claim @OpenAI’s internal system was set on the Navier–Stokes existence and smoothness problem one of the Clay Millennium Prize problems. About $1M if you solve it the old-fashioned way. They say the agents found a finite-time singularity proof and formalized it in Lean in roughly 88 hours. That alone would be one of the wildest AI-math moments ever published. The other team Tristan Buckmaster (NYU) and @__alpoge__ (at @AnthropicAI) had spent months on a related path forced Euler heavily using Claude and OpenAI’s Codex, logging drafts into the tools as they worked. They say they broke through around August 15. Then the rumor mill moved. The collision @OpenAI says it started September 1 after hearing a rumor connected to their progress, finished around September 6, then reached out to offer a concurrent release and recognize priority. Buckmaster says OpenAI only ramped after learning they were close, on a route that looked familiar and that he asked, twice, whether Codex sessions from their unpublished drafts had trained or informed the model. He says he never got a clean answer on training. @SebastienBubeck (OpenAI) has said they did not see the pair’s work until it was public, and that the proofs differ. The sentence nobody can unsee @OpenAI denies accessing their private work directly. It also wrote that it “cannot rule out” de-identified product data helping improve the models. Different proofs. Different exact statements. Same week. Same tooling surface. Same credit war. Millennium Prize math just got a product-data problem.

Gm

Pokémon Red created an open-world metaverse using less RAM than your calculator. - 8KB. - No Unity. - No Unreal. - Mostly assembly. And somehow Game Freak shipped an open world with 151 monsters. Here’s the part everyone forgot: - The machine Original Game Boy. Sharp CPU at ~4MHz. About 8KB of work RAM. Four shades of gray. A cartridge you couldn’t patch after launch. No engine. No asset pipeline. No “we’ll optimize later.” If it didn’t fit, it didn’t ship. - What they packed anyway 151 Pokémon. Maps. Music. Sprites. NPCs. Stats, moves, type IDs, health bars. Experience curves. Evolution pointers. Save files that multiply when you trade. Two Game Boys syncing battles in real time over a serial link cable. One iPhone photo is bigger than the whole game. - The joke inside the constraint Early on, the team thought even one map wouldn’t fit in 1MB on an 8KB machine. Then Satoru Iwata rewrote the compression so hard they fit two entire new regions into Gold and Silver. - Why it still hits Modern games burn terabytes, cloud servers, and millions of lines of code. Pokémon used engineering violence. Gotta catch them all wasn’t just a slogan. It was an engineering brief.

How do i make a funny tweet about AI going down without AI ?

Fei-fei spent 20 years teaching machines to see Yesterday her lab shipped one that can move this is @drfeifei we're talking about. ImageNet, 2006. she sat on millions of labeled photos so a computer could finally tell a dog from a toaster. 2012, alexnet trains on her homework and the entire field loses its mind. she didn't even ship the model. she shipped the eyes. then it's stanford, google, HAI, the whole godmother-of-AI tour. every talk is the same obsession: pixels aren't intelligence. space is. 2024 she actually walks out of the university and starts @theworldlabs with her people. justin johnson. mildenhall, the nerf guy. lassner. not a random founder stack. marble drops last november. you throw it a photo and you can walk around the room. that was the product. yesterday they ship atlas. you don't prompt the camera in english. you hand it the path. a minute of 1440p. then it rebuilds the scene in 3D. Marble was the demo Atlas is the actual real deal.

ASML doesn’t make the most important part of an EUV machine. NVIDIA doesn’t either. A German microscope workshop from 1846 does. And you cannot buy it. Here’s the forgotten @ZEISS_Group lore: - 1846 Carl Zeiss spent 20 years shaping glass lenses by hand in a German university town. It was artisan work. One man. One workshop. Microscopes. Then in 1866 he hires a young physicist named Ernst Abbe. Abbe writes the math for how light forms an image inside a microscope. d = λ/2NA. That equation is still the reason chips can get smaller. Once you can write the physics, you can make specialized lenses at scale. Zeiss stops being a craft shop and becomes an optics company. - The foundation that still owns it Zeiss dies in 1888. Abbe takes over. He does something almost no founder ever does. He puts the entire company into a foundation so it cannot be looted, and writes a constitution: pensions, sick pay, paid vacation, an 8-hour day. Before German labor law exists. That foundation is still the only shareholder of the parent. One owner. For 130+ years. - The dark chapter they don’t hide In the 1930s Zeiss made precision optics for Nazi weapon systems, using forced labor. The company doesn’t bury this. In 1945 it splits in two, one Zeiss on each side of the Iron Curtain. Same name. They spend the next 40 years suing each other over the trademark. After reunification they re-merge. Same glass. Same math. Same foundation sitting on top. - Why ASML would be toast without them Around then a small Dutch company, ASML, is figuring out EUV lithography. Lasers. Extreme ultraviolet light. A hall of mirrors etching circuits onto silicon. The mirrors and lenses have to be so precise they barely exist as a physical object. Zeiss makes the only ones that work. ASML has a total monopoly on EUV. They have said themselves they would be toast without Zeiss. Every advanced AI chip on earth goes through a Zeiss optic before it goes through anything else. - The 24.9% ceiling In 2016 ASML tried to buy in. €1 billion. They got 24.9% of one subsidiary. That’s the ceiling. The parent is still owned by Abbe’s foundation. Not listed. Not for sale. Not to NVIDIA, not to TSMC, not to a sovereign wealth fund. That semiconductor arm is now worth about €5B, growing ~23% a year, and is 42% of Zeiss. One of the most critical companies in all of chipmaking. It started as a guy grinding glass. Before NVIDIA. Before TSMC. Before ASML’s monopoly. There was a microscope. And @ZEISS_Group still can’t be bought.