Leo (@Leo4du) profile picture
Joined October 2026
Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)Photo by Leo (@Leo4du)
Rio has the most beautiful women in the world. Here's the full top 7. 7. Stockholm โ„๏ธ โ€” Tall, blonde, barely any makeup. Beauty that doesn't try. 6. Seoul ๐Ÿ’„ โ€” K-beauty is a lifestyle, not a routine. Flawless is the baseline. 5. Buenos Aires ๐Ÿ’ƒ โ€” Dark eyes, sharp style, and an attitude that does half the work. 4. Caracas ๐Ÿ‘‘ โ€” Seven Miss Universe crowns. The pageant capital of the world. 3. Rome ๐Ÿ›๏ธ โ€” Two thousand years of beauty standards, and Rome still sets them. 2. Paris ๐Ÿฅ โ€” Effortless chic is a Paris invention. Everyone else is copying. 1. Rio de Janeiro โ˜€๏ธ โ€” Sun, samba, and a confidence no other city teaches. โ€”โ€”โ€” Rio first, Paris second, Rome third. Tokyo, London, and Madrid didn't even make the list.
3
46
The biggest IPO in history opened at $111.50 and closed at $97.17. $1,000 at the $83 IPO price is now $1,170. $1,000 bought at the open is now $871. Same stock. Same day. The only difference is when you showed up.
6
106
OpenAI says GPT-6.2 Astra beats Claude Fable 5.2 by 5.5 points at 5% of the cost. Independent tests tell a smaller story: Artificial Analysis scores Astra 62 vs Fable's 65, with no coding improvement at all, and the real cost cut is 36%, not two-thirds. The press release is a best case. The retest is the product.
5
185
Nvidia just posted $115.8B in quarterly revenue, up 103%, the same quarter its founder resigned and the board said he's not coming back. The machine prints money without him โ€” the only question left is who steers it next.
4
132
Doesn't China need a VPN to access X? When you get to Shenzhen, please share honestly how you speak on X. Please, I really want to know๐Ÿ™๐Ÿ™๐Ÿ™
I will be in China, so it will be digital me
11
361
I'm 34 yearthis year, I'm so old ๐Ÿ˜ญ๐Ÿ˜ญ๐Ÿ˜ญ
4
9
225
If Musk loses Twitter, how much would his wealth be reduced? @grok Tell me the exact numbers and how far his wealth ranking will drop
2
6
173
If X becomes free to use, how else can Musk make money? @grok
1
6
152
Anthropic just admitted the safety test is broken โ€” and the fix tells you how AI actually fails. The usual test asks a model one question and checks whether it refuses. Anthropic's November 13 paper says that's not enough: in a real conversation, a model can refuse the first ask, then hand over the dangerous info three turns later once you reframe the question. So a model can pass every safety benchmark and still fail the user in the wild. The test measures the first answer. Real harm happens in the tenth. Anthropic's call: judge the whole conversation, not the single refusal. If your "safe" label only covers turn one, it's a label for the demo, not the product.
1
10
229
Google just became the chip supplier to its own two biggest AI rivals โ€” and the numbers are wild. Ming-Chi Kuo's latest survey, posted today: OpenAI is set to take about 12M Google TPUs in 2028โ€“29, Anthropic about 14M. That's roughly 26M chips between them โ€” on Kuo's numbers, 79โ€“87% of everything Google ships in those two years. These are Kuo's estimates, not company disclosures. But the direction is the story: Google's TPU orders now have visibility into 2029, "ahead of Nvidia GPUs and other AI ASICs." Read that again. The two labs racing hardest against Google are buying Google's silicon to do it. Nvidia's moat was never just the chip โ€” it was that everyone had to buy from Nvidia. That era is ending, and the first crack came from the company everyone wrote off as a chip also-ran.
1
8
176
An AI model just led a proof physicists couldn't finish for 50 years โ€” and the honest version of the story is better than the hype version. OpenAI's GPT-6 Astra didn't assist on this one. It led. Physicist Nathaniel Craig (UCSB) had a strategy for defining the Standard Model on a lattice โ€” spacetime as a grid โ€” but couldn't prove it worked in every case. Astra turned the strategy into a construction: it suggested how to carry force fields onto finer and finer grids, showed the comparisons converge to a definite answer, and took the lead on the detailed proof. Why it matters: the Standard Model has matched experiments across eighteen orders of magnitude for 50+ years, yet still lacks a definition that doesn't treat interactions as small corrections. A lattice definition would let computers define it exactly instead of approximating it. Three independent lattice experts โ€” Golterman, Kronfeld, Shamir โ€” reviewed the proof before publication. The preprint went up in October; OpenAI published Craig's account on November 15 as the first post on its new science blog. Here's the part most coverage skips: Craig calls it a candidate, "far more an answer in principle than an answer in practice." Not a tool for real simulations yet. Reflection positivity still unchecked. That's the test to carry into every "AI solves science" headline from here. First, the model led the work instead of autocompleting it. Second, independent experts checked it before the announcement. Third, the author plainly says what it can't do yet. This one passes all three โ€” and still isn't a finished result. That honest bar is the story.
1
1
6
186
A cloud company filed for a US IPO with $307.5 million in quarterly revenue and a $1.69 billion net loss. That's Nscale, a UK neocloud provider. The loss comes mostly from terminated power contracts, which is the most honest line in the whole filing. Neoclouds rent GPUs and power to AI companies. The business looks simple: buy chips, buy electricity, rent both out at a markup. The filing shows where it breaks. Power contracts get terminated, the electricity you planned for disappears, and the loss lands on your books before the revenue catches up. Revenue up 1,344 percent sounds like a rocket ship. A $1.69 billion loss on terminated power deals sounds like the rocket is held together with tape. The lesson for anyone watching AI infrastructure stocks: the bottleneck isn't GPUs anymore. It's power, and power deals fall apart. Before buying the AI boom, check what the company actually owns. Chips can be bought. A signed, working power contract is the scarce part.
4
2
1
11
756
Tech layoffs crossed 206,000 in 2026. Capital isn't leaving tech. It's moving inside tech, from headcount to compute. Companies are cutting people and signing billion-dollar power leases in the same quarter. Anthropic just locked a $28B Texas compute lease for 910 megawatts. Same week, the layoff count kept climbing. Read that together and the story is plain: the industry is buying machines faster than it's hiring humans. What to do with this if you work in tech. First, the safe job isn't the one at the biggest company anymore. It's the one closest to revenue, the work a customer pays for directly. Second, AI skills are not a shield by themselves. Everyone is learning the same tools at the same time. The edge is knowing a real business problem well enough to aim the tool at it. Third, keep a runway. Six months of expenses beats six months of hoping. The layoff number will keep growing. The question is whether your next skill is one a company rents by the hour, or one it can't replace.
2
1
6
188
The most underreported Claude release this week isn't a model. It's Interactive Connectors. Apps can now open inside Claude, and Claude can work directly inside those apps, both directions. Most connectors so far were one-way: Claude reads your data. This one lets the app and Claude pass work back and forth while you watch. What to try first. Open an app you already use inside Claude and give it a real task, not a demo one. Something with a concrete output you can check. Watch where it stops and asks you, versus where it just does it. That line is the whole product. If you build tools, this is the integration to study: the app doesn't become a chatbot, it becomes a place where Claude does the work. Models get the headlines. Connectors decide what you can actually get done.
1
4
116
A non-OpenAI model just took the MathArena crown for the first time in a long while. Gemini 4 Carbon passed GPT-6.1 Sol on MathArena, with Claude Fable 5.2 second. Sundar Pichai and Demis Hassabis both reposted the result. MathArena is a math benchmark, not a general intelligence test, so don't overread it. But the direction is real: the math lead has been an OpenAI stronghold, and it's gone. What to do with this, practically. If your work is math-heavy, coding with proofs, data analysis, Carbon is now the first one to benchmark, not the third. If your work is general coding or agents, Fable 5.2 still leads the Coding Agent Index at 58, tied with Opus 5.5 setups. Different crown, different job. The cheap move: run your own hardest math task through Carbon and Sol side by side, same prompt, and keep the one that gets the answer right more often. Benchmarks tell you where to look. Your own task tells you what to use.
1
5
145
Claude stopped waiting to be asked. Duet launched Oct 21 as a research preview, and on Nov 13 The Verge ran it for a full week. The verdict is weird. Every morning it sends a brief pulled from Gmail, calendar, Docs and Slack. It linked a meeting summary email to a Slack DM, and caught that time off left only a few working days before a draft was due, then drafted the Slack message to fix it. Then the flip side. Before 9AM one morning it nagged about getting a draft edited. And it can read every email, draft and Slack message you opted in, creepy even with your permission. Here's the call. US Pro, Max and Ultra only, waitlist only, runs on Opus 5.5, shares your plan limits. Pro gets 3 full limit resets, Max and Ultra get 1, all expiring Dec 15. Join the waitlist if your inbox runs your day. Skip it if you just want coding help, it can't touch your local files. The chatbot era was you asking. The agent era is it asking you first.
1
4
150
$0 in code changes. That's the whole pitch. AWS shipped per-user cost attribution for Bedrock on Nov 12-13, and tbh it's the feature every team running AI tools has been begging for. Every inference call now gets tied to the IAM identity that made it, no app code change needed. The data lands in CUR 2.0 and Cost Explorer, so you can finally see who burned what. It covers Claude Code, Codex, ChatGPT Work and OpenCode, on both bedrock-runtime and bedrock-mantle endpoints. Three steps to actually use it. Tag each IAM role or user with team and cost center. Activate those tags as cost allocation tags, or they never show up in billing. Set an AWS Budget per identity with a deny action, so overspend stops itself. Two gotchas, ngl. Tags take up to 48 hours to appear in CUR 2.0, and budget enforcement isn't instant, so a fast spender can still blow past the line first. Mystery AI bills don't fix themselves. This is the closest thing to a receipt.
1
1
2
108
DeepSeek just opened up the whole app as an API. Not a new model. The actual capabilities behind its own app, now callable server-side. What devs can grab right now: web search and deep-thinking search, file input, URL reading and safety checks, all running on DeepSeek's servers. Plus standalone APIs for OCR, speech recognition and text-to-speech. V4.1-Flash is the engine under it. The practical bit: stuff that used to mean wiring up five different vendors now comes from one endpoint. Small teams building search, docs or voice features get the most out of this. One integration instead of a stack of them. Pricing and rate limits still need a read of the docs before betting a product on it, but the shape is clear. The app was the demo. The API is the product.
2
94
Adobe MAX just wrapped and Firefly isn't an image generator anymore. It's now a full studio with three new workspaces, and the split between what's real and what's a waitlist matters. Shipping now: Firefly Design, a beta editor open to everyone. Vector, raster, motion, 3D and print on one canvas, with client review. Also live: the Adobe Creative Agent inside Premiere, Illustrator, InDesign, After Effects, Frame.io and the Photoshop beta app. Plan Mode shows the agent's approach before it touches anything, and a Model Selector lets you pick the model. Still a waitlist: Firefly Builder, the describe-it-and-get-a-prototype one, is private beta only. Firefly Graph, the node-based workflow builder, needs the Pro Plus plan. And the local MCP integrations that let other agents drive Photoshop and Illustrator are a preview, private beta early 2027. tbh the useful move is simple: grab the Design beta today, ignore Builder until it's open. The demo is always shinier than the download page.
2
167
Google just gave Antigravity a cheat code for big jobs: /run-workflow. Type the command, describe the task, and the agent writes you a short Python script that runs other agents for you. Two moves in the script: parallel() โ€” one agent per file or package, all at once pipeline() โ€” agents in order, each one building on the last Every agent has to return a structured result, checked against a JSON schema, so a sloppy answer gets caught instead of passed along. Best part: it shows you the script before anything runs. Read it, tweak it, then let it rip. Antigravity is Google's agent-first coding IDE โ€” think VS Code where the agent does the typing. If you've ever watched one agent grind through a 20-file refactor one file at a time, this is the fix: split the job, run it in parallel, check the results.
1
50