Chetan Dube

I've spent the last 4 days testing UAE's K2 Think AI model. (a detailed review + demo from an Emirati) The UAE released this open source model on Sept 9th. And everyone's underestimating what this means. They call it "small but mighty" for a reason. OpenAI needs massive servers to run their models. Or costs a fortune if you try running it yourself. K2Think is just 32B parameters + runs on decent hardware. - It matches GPT-4/5 on AIME math benchmarks - It beats most models on reasoning - It does well with science and analytical tasks Every business I talk to has the same fears. - Legal won't let them upload contracts to ChatGPT - Finance can't share data with Claude - HR refuses to put employee information anywhere near AI The demand for AI is massive. But they can't use it properly because of security. K2 Think solves this completely. 1. Download the 62GB model from HuggingFace 2. Run it completely offline 3. Build your custom ecosystem around it "But Adham, open source models existed forever..." True. But they weren't this good. I stress tested this model for 4 days straight. - I created websites with it - I wrote emails, reports, and articles - I connected it to tools like accessing the internet (MCPs) (in both English and Arabic) If it stays this capable with complex tasks, I'm pushing it into every UAE business I work with. (secret: you can use AI to help build it, no tech skills needed) Another secret for consultants: Adding this to your delivery is a massive selling point. It's from the UAE. It solves security fears. And enterprises will pay premium for this. I recorded everything: Walkthrough + demo + documentation + setup. I did it in Cursor (you can do it anywhere). You can watch it here: https://linktw.in/ABGWXH The businesses that move on this now will win in the UAE market. Big leaps forward in the AI space from MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) & G42 & Institute of Foundation Models

189

34 Comments

#Agentic systems represent a fundamental shift: #software is no longer just a passive executor of instructions. It becomes a semi-autonomous actor —capable of interpreting, deciding, and operating in open-ended contexts. This evolution is redefining the very notion of intelligent automation and laying the groundwork for a new generation of digital systems: more flexible, adaptive, and collaborative by design - Wavestone. #ai #artificialintelligence #aiagents #agenticai #generativeai #automation #digitaldisruption #digitaltransformation Paidi O Reilly Prof. Dr. Ingrid Vasiliu-Feltes Martin Moeller Aaron Lax Richard Turrin JOY CASE Antonio Grasso Nicolas Babin Imtiaz Adam Phillip J Mostert 🇿🇦 Irene Lyakovetsky🎧🎙 Dinis Guarda Ian Jones Dr. Martha Boeckenfeld Giuliano Liguori Orlando Francisco F. Reis Hope Frank Ima Miri Mike Flache Dr. Marcell Vollmer Birgul COTELLI, Ph. D. Sharad Agarwal Olivier Kenji Mathurin Franco Ronconi Dr. Sindhu Bhaskar Patrick Maroney Amna Usman Chaudhry Helen Yu Neville Gaunt 💡⚡️ Lionel Costes Anthony Rochand Nafis Alam Eveline Ruehlin Avrohom Gottheil Dr. Debashis Dutta Imtiaz Hussain Dr. Monika Sonu Dr. Khulood Almani🇸🇦 د.خلود المانع Alberto Espinosa Machado

110

15 Comments

𝗔𝗻𝗻𝗼𝘂𝗻𝗰𝗶𝗻𝗴 𝘁𝗵𝗲 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲 𝟮𝟬𝟮𝟱! 🚀 Ready Tensor presents our most exciting competition yet with a $10,000 prize pool to celebrate innovations in autonomous AI agents and LLM workflows! 𝗣𝗿𝗼𝗷𝗲𝗰𝘁 𝗖𝗮𝘁𝗲𝗴𝗼𝗿𝗶𝗲𝘀: From foundational frameworks to practical applications: ✔️ Agent architectures and frameworks ✔️ LLM workflow systems ✔️ Task planning and decomposition ✔️ Multi-agent communication systems ✔️ Personal AI assistants ✔️ Code generation automation ✔️ Business process automation These are just examples - we welcome all innovative approaches in agentic AI! 𝗞𝗲𝘆 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: 🔹Submissions Open: Feb 10, 2025 🔹Deadline: March 25, 2025 🔹Who Can Join? AI devs, researchers, students & teams worldwide! 🔹Prizes for various project types including research papers, applied solutions, case studies, technical deep dives, tutorials, open-source packages/tools, and more 🔹Multiple recognition categories with cash awards 🔗 𝗟𝗲𝗮𝗿𝗻 𝗠𝗼𝗿𝗲 𝗮𝗻𝗱 𝗝𝗼𝗶𝗻: https://lnkd.in/gRJPQMjU 𝗣𝗮𝗿𝘁𝗶𝗰𝗶𝗽𝗮𝘁𝗶𝗼𝗻 𝗶𝘀 𝗙𝗿𝗲𝗲 𝗮𝗻𝗱 𝗘𝗮𝘀𝘆! 1. Register for a free Ready Tensor account (see link above) 2. To initiate your publication, click 'Create New' > 'New Publication' from the left menu 3. Follow the intuitive workflow to document your work in markdown format 4. Enter the completed publication into the competition during the submission period Typical submissions only take ~1 hour! 💡 No need to build something new - showcase your existing work in agentic AI! Just make sure it's your original project. 𝗔𝗯𝗼𝘂𝘁 𝗥𝗲𝗮𝗱𝘆 𝗧𝗲𝗻𝘀𝗼𝗿: We are a global AI platform empowering innovators to showcase their AI/ML projects and solutions. Join us in celebrating the next wave of AI technology! #AgenticAI #AIChallenge2025 #ArtificialIntelligence #AIAgents #MachineLearning #ReadyTensor #Innovation

61

𝗧𝗵𝗲 𝗹𝗮𝘀𝘁 𝗵𝗼𝗹𝗱𝗼𝘂𝘁 𝗵𝗮𝘀 𝗳𝗮𝗹𝗹𝗲𝗻: 𝗢𝗽𝗲𝗻𝗔𝗜 𝗴𝗶𝘃𝗲𝘀 𝘂𝗽 𝗮𝗻𝗱 𝗮𝗱𝗼𝗽𝘁𝘀 𝗠𝗖𝗣 🙌 When Anthropic introduced the Model Context Protocol (MCP) in late 2024, it promised an important component for interoperability - a standard way for LLMs to securely access tools, files, APIs, and databases. Most leading providers aligned early. Except one: OpenAI 😉 For nearly a year, OpenAI pushed a more controlled, internal approach - "Function Calling" and "Tool Calling" - prioritizing proprietary tool use over open standards. But that resistance is now over. With the new Developer Mode in ChatGPT, OpenAI has officially added full support for MCP tools, allowing developers to: - Connect custom MCP-compliant backends - Execute both read and write actions - Chain tools into complex, stateful automations 𝗣𝗦: 𝗠𝗖𝗣 𝗶𝘀 𝗼𝗻𝗹𝘆 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝘄𝗲 𝗻𝗲𝗲𝗱. Other construction work is happening - like A2A for agent communication by Google, or our own 𝗧𝗿𝘂𝘀𝘁𝗲𝗱 𝗔𝗴𝗲𝗻𝘁 𝗛𝘂𝗱𝗱𝗹𝗲™ protocol we developed at Accenture, to enable collaboration between agents across different clouds and enterprise environments! ♻️☁️ But MCP solves a critical piece: 𝘁𝗵𝗲 “𝗡×𝗠” 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 𝗽𝗿𝗼𝗯𝗹𝗲𝗺. Before, every new model and every new tool meant a custom connector. Now, we have a protocol to reduce that friction - and that’s great! Peter Seeberg | Thorsten Maier | Dr. Dirk Alexander Molitor | Igor Aranovsky | Dr. Thomas Holzmann

40

6 Comments

OpenAI’s GPT-5 required more than 50,000 NVIDIA H100 GPUs to train — over double the compute of GPT-4. The Compute Revolution: Why GPT-5 Proves the World Needs Neurolov.ai To put that in perspective: • 50,000 H100s = $2 billion+ in hardware costs. • Power usage = the electricity consumption of a medium-sized city. • Compute requirements are now 4,000× greater than what GPT-3 needed just a few years ago. • Training runs consume hundreds of megawatt-hours, stretching across weeks of nonstop GPU operation. This isn’t just about one model. It’s a wake-up call. AI is scaling at a pace that centralized infrastructure simply cannot sustain. The Centralization Problem Right now, only a few tech giants have the resources to secure this level of compute. But this creates bottlenecks and inequality: • Startups and researchers can’t afford entry. • Innovation is restricted to those who control infrastructure. • Costs spiral while demand skyrockets. If we continue down this path, AI’s future will be centralized, exclusive, and unsustainable. The Neurolov Breakthrough This is why we built Neurolov. We are creating the world’s first browser-based decentralized compute network: • Anyone can contribute their device’s GPU power through a browser. • The network aggregates idle power into a global GPU marketplace. • Developers, startups, and enterprises can access compute at a fraction of today’s cost. No billion-dollar data centers. No locked gates. Just open, scalable compute for everyone. Why This Changes the Game The leap from GPT-3 → GPT-4 → GPT-5 proves one thing: demand will keep multiplying. • GPT-6 or GPT-7 could require hundreds of thousands of GPUs. • Centralized clouds won’t keep up — financially or environmentally. • Neurolov’s decentralized model thrives as demand grows: more contributors = more compute. This is not just a solution for cost. It’s a paradigm shift in ownership and access. The New Law of Compute At Neurolov, we believe in a simple principle: Compute should be as decentralized and open as the internet itself. Neurolov ensures that the next wave of AI breakthroughs won’t be locked in corporate vaults — they’ll be powered by a global, browser-based network owned by the people. The world doesn’t just need more GPUs. It needs a new way of thinking about compute. That’s the revolution Neurolov is leading. Closing Thought The age of centralized AI infrastructure is ending. The future belongs to decentralized, browser-powered compute networks like Neurolov. Because when compute is open, innovation knows no bounds.

22

𝗬𝗮𝗻𝗻 𝗟𝗲𝗖𝘂𝗻: 𝗟𝗟𝗠𝘀 𝗔𝗿𝗲 𝗡𝗲𝗮𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝗘𝗻𝗱, 𝗯𝘂𝘁 𝗕𝗲𝘁𝘁𝗲𝗿 𝗔𝗜 𝗜𝘀 𝗖𝗼𝗺𝗶𝗻𝗴 (𝘸𝘦𝘦𝘬𝘦𝘯𝘥 𝘭𝘰𝘯𝘨 𝘳𝘦𝘢𝘥) While the AI world celebrates LLMs as the pinnacle of intelligence, Yann LeCun quietly challenges the consensus: 𝙇𝙇𝙈𝙨 𝙝𝙖𝙫𝙚 𝙥𝙚𝙖𝙠𝙚𝙙. 𝙏𝙝𝙚 𝙣𝙚𝙭𝙩 𝙗𝙞𝙜 𝙡𝙚𝙖𝙥 𝙞𝙣 𝘼𝙄 𝙬𝙤𝙣'𝙩 𝙗𝙚 𝙡𝙖𝙣𝙜𝙪𝙖𝙜𝙚-𝙗𝙖𝙨𝙚𝙙. In 2025, that's practically heresy. But is he wrong? Language is powerful, yet fundamentally limited: → It compresses reality rather than fully capturing it → It describes the world instead of truly representing it → It excels at pattern prediction but falls short on genuine reasoning LeCun argues that transformative AI needs to understand: → Gravity → Object permanence → Thermodynamics → Physical and chemical laws In other words: not just how we talk about the world, but how the 𝙬𝙤𝙧𝙡𝙙 𝙖𝙘𝙩𝙪𝙖𝙡𝙡𝙮 𝙤𝙥𝙚𝙧𝙖𝙩𝙚𝙨. His perspective builds on Kahneman's Systems 1 and 2, with a crucial distinction: → Current LLMs = intuitive heuristics (System 1) → True intelligence = deliberate reasoning (System 2) → And we haven't bridged that gap yet 𝗦𝗼 𝘄𝗵𝗮𝘁'𝘀 𝗼𝗻 𝘁𝗵𝗲 𝗵𝗼𝗿𝗶𝘇𝗼𝗻? Multi-modal models that transcend language as their primary thinking mechanism. Systems that will "know" the world the way your hand instinctively knows to pull away from a hot surface. Less text. More embodiment. Less prediction. More understanding. ChatGPT might impress you today, but the real revolution is still coming

25

15 Comments

China 🇨🇳 has become the first country in the world to launch a fully AI-powered hospital. The AI hospital was developed by Tsinghua University. It is staffed by 42 AI doctors and 4 AI nurses who treat simulated patients using real medical data in a Virtual Environment. The AI hospital is capable of treating up to 10,000 patients in just a few days, something that would take human doctors nearly two years. AI doctors and nurses will shape the future of healthcare. In this AI era, ANYTHING IS POSSIBLE. Check out what we are building at Gemsbok Group - https://lnkd.in/efBhp8sR

109

21 Comments

Mo, Larry, and Curly, aka the three stooges, made fun of passing the same money between the three of them back in 1925, appropriately 100 years ago. Now OpenAI, Nvidia, AMD, and Oracle are doing something similar, yet the media has been slow to catch on. Much of the media seems to think circular investments means something economically important when they just represents #hype and an attempt to pull the wool over investors who so far have been willing to play along with the generative AI bubble. Some people will claim that this is how the economy works but those decisions are for the most part independent of each. We are paid money in wages, we buy things with that money, and we pay taxes. (Sorry taxes aren’t optional except for the richest) In the case of OpenAI, Nvidia, AMD, and Oracle, the actions are decided beforehand, and they are not independent of each other. Mo, Larry, and Curly were making fun of banks passing around the same money back in 1925 but for banks, there were interest payments and charges and for the most part the decisions were independent of each other. Things are different for the #tech bros. Some claim that the money being passed around by Nvidia, AMD, OpenAI, and Oracle are really a “trillion-dollar illusion” or “a house of cards” “where circular money, contingent warrants, and hype keep Silicon Valley’s boom standing on borrowed time.” As one article says: “Investors aren’t chasing power anymore. They’re chasing press releases. Dollars loop through vendor financing and prepaid contracts, showing up twice or more on different balance sheets. It’s round-tripping dressed as revolution.” “For now, the house of cards gleams. But its foundations are paper: promises instead of profits, milestones instead of cash flow. The stone building—the real power—lies elsewhere in today’s market, in assets that don’t need a press release to hold value. This is Silicon Valley’s House of Cards moment. Elegant on the surface, empty underneath.” The worst thing about these superficial deals is that they distract the tech bros from important work they need to do. They need to offer products that are so good that customers will pay them more than it costs them to offer the product. So far, OpenAI has not been able to do this; instead, it pays big money to the AI cloud providers, more than it takes in from subscribers, and it’s going to get worse. Videos created in Sora 2 cost about $5 each (according to Ed Zitron) and users can make 100 of them a month with a $20 a month service. This is partly why OpenAI admits they will have cumulative losses of $119 billion by 2029. Superficial media companies probably ignore those losses because OpenAI, Nvidia, AMD and Oracle are passing around hundreds of billions in their strategy borrowed from Mo, Larry, and Curly and thus the media thinks that big things are coming from generative #AI #technology #innovation #artificialintelligence

198

69 Comments

Tesla's Ex-Director of AI Andrej Karpathy dropped the bombshell about AI Agents! He said: AI agents like Claude or Codex are still far from acting as real “employees.” They lack memory, multimodality, and true learning. Fixing that will take about a decade, not because it’s impossible, but because deep intelligence just takes time to build. -- What this means for builders: You’ll win by shipping systems, not chasing sentience. 1) Build workflows, not ‘workers’ Think pipelines, not people. Chain tools and decisions that produce reliable outcomes. 2) Small agents > grand agents Design 3–5 step micro-agents that do one job 10x better: Ingest → ground facts → act → verify → deliver Add a review step (auto or human) before you call it done. 3) Thin wrappers can be powerful “Thin” doesn’t mean trivial. Look at Perplexity: a lean UX + retrieval + strong ranking loop = massive utility. Your wrapper = orchestration + guardrails + clarity, not just a chat box. 4) UX is the moat Cursor showed that great process + interface beats raw model magic: Tight context controls, multi-file edits, agentic actions, visible plans Reliable outputs people trust > clever demos they don’t adopt 5) Memory, but make it practical Until long-term learning is native, use working memory: Task files, run manifests, audit logs, and checklists the agent reads/writes Treat memory like a shared workspace, not a black box ✅A simple playbook to ship this week: A. Frame a single high-value workflow Example: “Weekly customer health summary” or “Invoice QA before payout.” B. Chain 3–5 steps Ground: fetch sources (CRM, tickets, docs) Synthesize: produce structured draft (JSON/Markdown) Validate: fact checks, schema checks, policy checks Repair: auto-fix common failures (missing owner/dates) Output: send to Slack/Email/Notion with next actions C. Add thin-but-sharp UX Clear inputs, visible plan, one-click approve, one-click revise Show your work: sources, diffs, and confidence notes D. Measure what matters First-try-right %, handle time, rework %, user saves/edits, incident rate Promote versions that beat baseline, retire the rest Common traps (skip these): “General agent” fantasies (do everything, does nothing well) Endless chats with no definition of done Hidden memory (no artifacts, no auditability) Evaluating with vibes, not metrics We don’t need artificial employees to create value today. We need well-designed workflows, small dependable agents, and great UX.

774

96 Comments

"Alibaba Unveils Advanced AI Model, Claims It Outperforms Rival DeepSeek-V3" was the headline that dominated the WSJ on January 29, 2025, sending shockwaves through the AI industry. The news follows a market frenzy sparked by DeepSeek, a Chinese AI startup that has quickly gained global attention. Launched just 20 months ago in Alibaba's hometown of Hangzhou, DeepSeek has made waves by announcing it can rival its American competitors at a fraction of the cost, causing a $589 billion drop in #NVIDIA's market value. It seems like we can expect even more exciting developments to come out of Hangzhou, China, in the near future. . #AI #DeepSeek #Alibaba #Tech Leadership

99

4 Comments