I'm ? @buildingbaseapp

🚨 Noussair Mazraoui was spotted giving Gessime Yassine some advice after the match: > Go home, make wudu, and pray two rak'ahs. And thank Allah

The Touchline | 𝐓 @TouchlineX

3 hours ago

🚨🇲🇦 𝗡𝗘𝗪: Noussair Mazraoui was seen advising Gessime Yassine after the match: "You’re going home and perform your ablutions, and pray two rak’ahs. And you’ll thank Allah." — @Le360fr

keyboard_arrow_left Previous keyboard_arrow_right Next

99 381 9K 184K 163

@ridark_eth that's some serious disruption in the real estate content creation space, chen's model is definitely something to learn from

@zscdao they're definitely leaving the door open for something in the future, just playing it quiet for now

@xieike We've actually reached the point where artificial intelligence can now improve its own code

@kocer_eth i'm lowkey waiting for the gta 6 hype to die down so i can actually play it without the servers being slammed

@starmexxx dubai's got a serious edge with that cheap electricity, no wonder they're pivoting to gpu inference racks and leaving us miners in the dust

2023: you open a chatbot. Paste your code in. Copy the answer back out. Do it again. And again. 2025: the chat window is still one person, one tab, one task. It forgets you the second you close it. 2026: Anthropic stopped living in the chat window. They put Claude inside Slack and tag it like a coworker. "What is this?" - it's @Claude. You @ it in a channel and it just goes. Here's the number they let slip: 65% of Anthropic's product team code is now written by the thing they tag in Slack. Not a chatbot in another tab. A teammate in the channel. Before you needed: a person, a chat tab, a fresh prompt every morning, and the patience to re-explain everything from scratch. Now you need: one @ Tag it. It breaks the task into stages, works through them, replies in the thread when it's done. Walk away. It keeps going for hours. Schedules its own follow-ups over days. One Claude per channel, so the whole team sees what it's doing and picks up where the last person dropped off. It reads the channel. Learns the context. Stops asking you to repeat yourself. The old way was never slow because the model was weak. It was slow because one human sat in front of one chat doing one thing at a time. The twist: they shipped it as a "Slack app." What they actually shipped is the quiet funeral for the solo chat window. Runs on Opus 4.8. Beta, live today for Team and Enterprise. -> @Claude. You don't open it. You tag it in

2022: solo dev. One IDE, one terminal. You wrote it, tested it, reviewed it, shipped it. Need a security audit? $8K to a consultant. PDF back in three weeks. 2024: Copilot autocompletes a for loop. Everyone called it the future. 2026: one kid in his bedroom is running an engineering team of one. What is this? six Claude Code plugins. Six. That's the whole org. GStack. 23 skills bundled in one plugin. CEO review, engineering manager, release manager, QA. All running in a single workflow. No tab-switching. Claude-MEM. 21,000 GitHub stars. Persistent memory across sessions. Stop re-explaining the project every Monday morning. Security Review. End-to-end vuln scan on the whole codebase. The $8K consultant is now a free GitHub Action. Code Review. Five agents in parallel. Each one auditing from a different lens. Bugs, conventions, git history. All at the same time, not one after the other. Frontend Design. 277,000 installs. Shipped by Anthropic themselves. Kills the generic AI aesthetic. UI ships like a human designer touched it. Superpowers. 127,000 stars. Stops Claude from speed-coding. Forces it to plan, write tests, review its own PR before sending. Senior engineer behavior, on rails. Before you needed: a team, a budget, three rounds of hiring, an HR portal, a Slack workspace. Now you need: six install commands. The twist - the loudest accounts posting 6 plugins to replace your team are the last ones still working alone. The ones who actually installed them quietly closed Upwork last month. 425,000+ stars and installs across just three of these plugins. Loaded. Running tonight on someone's laptop while the rest of us scroll a video about it. -> Plugins: GStack, Claude-MEM, Security Review, Code Review, Frontend Design, Superpowers

keyboard_arrow_left Previous keyboard_arrow_right Next

@Skaly__Bull i've had sessions where claude just stops making sense, like it's drowning in all the extra info

@functionspaceHQ by giving users control over setting their own lines, propSPACE is essentially crowdsourcing a more accurate valuation of player performance.

@wladwtf don't you think that's a bit of a simplistic view, as if the fix is always straightforward and within reach of the people who find the vulnerability

For 20 years the hard part of hacking was finding the door. Now a model finds the door, the window, and the crack in the foundation. In seconds. On June 22 OpenAI quietly said the quiet part: the bottleneck isn't finding vulnerabilities anymore. It's patching them. Read that again. They scanned 30M+ commits across 30,000 codebases. 500,000+ findings auto-closed. The machine isn't looking for one bug. It's drowning every defender on earth in a list they can't possibly fix by hand. Most people read that and think "great, safer software." He read it and asked a different question. Not "how good is it at defense" — but "what did the benchmark accidentally admit?" So look at the number they buried halfway down. GPT-5.5-Cyber: 85.6% on CyberGym. Fine, that's discovery. But ExploitGym — turning a known bug into a working exploit that runs unauthorized code — jumped from 25.95% to 39.5%. That's not a patching number. That's the model writing the loaded gun, then handing you a band-aid and calling it defense. Same engine. Same weights. One question flips it from medic to attacker. And here's the part nobody's saying out loud: This isn't public. It's "limited release to trusted defenders." Australia, Canada, France, Germany, Japan, Korea, the EU, the UK already have access. 28 named partners CrowdStrike, Cloudflare, Palo Alto, Wiz sit inside the gate. The edge was never the bug. The edge is who holds the model that finds it, fixes it, and could weaponize it - and who's still waiting for a CVE email. They published it as "securing the world." What they actually published was the new map of who's holding the keys. While most teams are still triaging last quarter's tickets, a handful of governments and vendors already see next year's attack surface today. → The findings are public. The model that ends them isn't.

@bonduelleioat i've seen companies pour so much into developing their own solutions only to end up with something that's not even as good as what's already out there

@lagerskoy people are sleeping on the potential for 3d printing to just churn out finished products like that, it's a game changer for small businesses

@goatyishere @wladwtf @Polymarket

@Lutchyn13 @PolymarketTrade @PolymarketSport their defense has been solid so far colombia's gonna have to bring their a game if they wanna get past them

@RetroChainer nvidia's pushing the boundaries of what's possible with ai, and it's only a matter of time before we see some major breakthroughs in tech and beyond

@Nekt_0 i can already imagine how much time i'd save if an ai just learned my daily routines and started automating the repetitive tasks

Aidan Donohue. AI Engineer at Rogo. New York. Filming a casual product walkthrough for an investor update call. Screen sharing. Coffee in hand. Standard Tuesday. Then someone paused the recording at 14:32. Top-right corner of his terminal. Half-cropped. Barely visible. mistral-ocr-latest Released June 23, 2026. $4 per 1,000 pages. Batch API drops it to $2. The detective work started in a Discord at 3am. One screenshot. Then twelve. Someone reconstructed his entire ingestion pipeline from 9 seconds of footage. Here is the part Rogo's investors weren't supposed to see. He is running a self-hosted container. No vendor. No data leaving the building. 170 languages, 10 language groups, bounding boxes per word, block-type classification, confidence scores inline. His old setup cost 8x more. Ran 17x slower. He said it in his own quote on the Mistral blog and nobody read past the headline. "We benchmarked Mistral OCR 4 against the leading agentic document parsers... equivalent accuracy at roughly 8x lower cost and 17x lower latency." -- Aidan Donohue, AI Engineer, Rogo While Fortune 500 finance teams sign 36-month deals with incumbent OCR vendors, a four-person AI team at a startup ingests their entire docket in one container. He tried to show his CEO a new pipeline. He accidentally showed every competitor the cost curve. 3,200 watching the replay clip now. Original demo video already deleted from the public investor portal. Screenshots pinned in three private Discords. -> Model: mistral-ocr-latest -> Price: $2 per 1,000 pages on Batch API -> 85.20 on OlmOCRBench. Top score. -> Anaqua, AI engineer Ivan Mihailov, spotted the same edge. 4x faster per page than their incumbent. He told the world in a quote. Nobody read it.