Atul Kumar

This is WilD ! They say AI can’t do emotions. @Kling_ai Elements just proved them wrong 3 images + a prompt → smooth, cinematic storytelling.

@AnvitaFlow That 113th minute winner will be remembered for years. Incredible finish!

⚽ WORLD CUP FINAL｜KINGS CROWNED 🏆 🇪🇸 SPAIN 1-0 ARGENTINA 🇦🇷 Decided in Extra Time 📣 Check your prediction results: 🔗 link: flow.anvita.xyz/activities/cyb… 🔥 A Legendary Night Ferran Torres 113th minute One strike sealed it 💥 16 long years of waiting Spain climbs back to the top 👑 ⚔️ An evenly matched showdown Football at its finest! 🎁 Winners, claim your rewards on Anvita Flow! 🌟 Remember: we're all champions of our own story #AnvitaFlow #CyberCup #Football2026 #AIAgents #Web3‌‌

Most AI subscriptions make you pay again every time you want to try a new model. That's what made Alibaba Cloud's Token Plan stand out to me. I've been testing it, and everything runs from one shared credit pool. You can switch between Qwen3.8 for text, GLM 5.2 for different reasoning tasks, and HappyHorse for video without buying separate plans or unlocking another subscription. No juggling multiple balances. No paying model by model. Just one plan that works across different tools. If you're currently paying for multiple AI subscriptions, it's worth comparing what a single token plan can actually replace. int.alibabacloud.com/m/1000415601

Love this combo. K3 brings strong reasoning, but pairing it with @youdotcom's real time search makes the responses far more grounded. Fresh information plus solid reasoning is exactly what AI agents need to be genuinely useful.

Introducing Live AI: Meet Lucy 2.5! Get ready for the most advanced live AI model yet! Just tried it, and the results are mind-blowing! Real-time AI editing is evolving at lightning speed. This is just a sneak peek of the future for creators, livestreams, and e-commerce. 👀

Another nice touch is the pricing model. Structured extraction is already included, and you pay per API call instead of per seat. There are also AI labeled commerce data packs for Amazon & TikTok today, with more marketplaces coming soon. Curious to see what developers build with it. Point your agent at it: go.zoodata.ai/uX09Hh 1k free credits to start. #zoodata

The part that stood out most wasn't just cleaner output. • Around 75% fewer downstream tokens • Around 4× more key facts per KB • Built for RAG pipelines where cleaner context matters more than bigger context Less cleanup. More reasoning. (Attach 29KB vs 120KB comparison image)

Everyone's talking about giving AI agents eyes. But here's the real problem: Most scraping APIs hand your agent raw HTML or markdown. Your agent then wastes tokens cleaning, parsing, and structuring the data before it can actually think. Raw HTML isn't data. #zoodata

Everyone is obsessed with building smarter AI agents. Meanwhile, they're feeding them garbage. Most scraping APIs dump massive HTML pages and call it a day. ZooData flips that into clean structured JSON with a fraction of the tokens, lower cost, and dramatically better outputs. #zoodata

@worldmodelHQ Your observation about "on-device" being the entire story is spot-on and deeply appreciated.

"On-device" is the detail everyone will scroll past, and it's the whole story. Every interactive world demo this year needs a datacenter behind it. Touch-directed video generation running locally on a phone is a different physics of product: no round trip, no per-user inference bill, no throttling when it goes viral. If the quality holds, this is the first interactive generation you can actually ship to a hundred million people.

@DecartAI This is truly impressive! Lucy 2.5's enhanced capabilities and greater control for real-time video editing are remarkable.

Meet Lucy 2.5, our most advanced Live AI model yet. Lucy edits videos in realtime, now with more capabilities and greater control. See how it's being used across streaming, e-commerce, advertising, and more 🧵

@worldmodelHQ Wow, thank you for this deep dive into the RGB, depth, and optical flow architecture! Your point about manipulation labs leading the charge, rather than video-gen labs, is such a critical distinction.

One backbone, three outputs: RGB, depth, optical flow. Notice where this kind of architecture keeps coming from — manipulation labs, not video-gen labs. When your output has to move a real arm, pretty pixels stop being enough on day one. It's not yet the field's endgame (schematic below, from Shanghai AI Lab's roadmap last week): one persistent internal state that decodes into appearance, geometry, and action. Tri-branch prediction shares a backbone, not a state. But it rhymes with the endgame, and the direction of travel is clear.

@worldmodelHQ Wow, what a brilliant breakdown! Your explanation perfectly captures why AMAP is positioned to revolutionize world models.

Why does a maps company suddenly have one of the more convincing interactive world models? Because every "infinite world" demo so far has been locally gorgeous and globally amnesiac. Nothing persists. No stable geography. The missing ingredient was never model capacity, it's grounding — and AMAP owns decades of road topology, street-level imagery, and movement traces that no competitor can scrape. Mapping, driving, and satellite companies are all one release away from being world model companies. Waymo did it for driving. This is AMAP doing it for exploration. The second wave won't come from video-gen labs.

@youdotcom @Kimi_Moonshot @merge_api Simple & Direct: Thanks, Kimi! K3 on Merge API Gateway + You. com search is brilliant. It really nailed what happened today.

@Kimi_Moonshot K3 on @merge_api gateway is a great call. Add @youdotcom search and it actually knows what happened today!