Tuana Çelik

Great blog post from Cornelius van Heerden at Google - https://lnkd.in/euPj5248. Spot on. Complex #queries, #analytics and #reasoning are enabled by business #semantics. Not simply #metadata catalogs and metric trees, which are lite semantics at best, but rich domain models (#ontologies). Organizations need horizontal #abstractions over their business silos, IT layers, cloud hosts and ecosystem partners. Done right, ontologies can enable global visibility, end-to-end #automation and virtually centralized policy-management. Data by itself is literally "meaningless". Ontologies provide domain #knowledge and real-time operational #context, this is critical to optimizing and controlling #GenerativeAI. It's ironic, but turns out #DeepLearning ("new #AI") and #StructuredKnowledge ("classic AI") are better together! The key is to ease and speed the design, use and maintenance of domain ontologies. That's what we do at EnterpriseWeb.

40

3 commentaren

Some LLMs are better at visual OCR 🎨 (e.g. Gemini 2.5 Pro) and some are better at structured extraction (e.g. GPT-5-mini). 🛠️ Our new update in LlamaCloud lets you mix and match models for structured extraction!  Check out the video below - we use Gemini 2.5-enabled parsing, but combine it with 5-mini to extract out a hundred+ toy listings from a large toy catalog. Check it out in LlamaCloud today: https://lnkd.in/g9Wpqn7w

187

7 commentaren

🤖How and when to build multi-agent systems Two really good blog posts (from Anthropic & Cognition) came out last week on building (or NOT building) multi-agent systems ❌Cognition: Don’t Build Multi-Agents 📃Anthropic: How we built our multi-agent research system Despite seeming different on the surface, they actually have a lot of common guidelines and advice. I tried to summarize some of that (and relate it to LangChain efforts) Blog: https://lnkd.in/g2uuc5YZ

384

10 commentaren

We cut the hallucinations of an LLM in half using a new technique: RLFR (Reinforcement Learning from Feature Rewards). RLFR reduces hallucinations in Gemma 12B by 58% (when run with our probing harness), for ~90x cheaper per intervention than the LLM-as-judge alternative. The technique uses lightweight probes on a model's internal representations as reward signals for reinforcement learning. This is all while avoiding: - off-target effects (no degraded performance on standard benchmarks) - degrading/reward-hacking the probes (the probes still work as monitors at test time!) This is just our first public demonstration of our work building towards "intentional design": the ability to intelligently guide gradient descent, enabling a new paradigm of far more precise, robust, and effective training. Kudos to Aaditya Prasad, Connor Watts, Jack Merullo, Dhruvil Gala, Owen Lewis, Tom McGrath, and Ekdeep Singh for this groundbreaking new work!

211

9 commentaren

Skyvern’s open source project just got faster. 🚀 We’ve swapped out Poetry for uv: the blazing-fast new Python package manager. That means your builds are now: • Much quicker: installs and resolves dependencies in a fraction of the time. • Simpler: no more waiting around for environments to set up. • Future-proof: leveraging modern tooling to keep the developer experience smooth. What this means for you 1. Clone and build Skyvern like before, but just 3x faster 2. Contribute confidently knowing your local setup won’t slow you down. 3. Enjoy a smoother dev loop when testing, fixing, or extending Skyvern. Try it out now: pip install skyvern

20

You see new models every week. But do you really know which one is right for production — GPT-4.1 or Gemini 2.5 Pro? Choosing a new LLM for your AI application can mean improved performance and cost savings — or it can lead to unexpected failures. Here’s a short demo of how we select the best model based on clear evaluation metrics. We test each model with 10 sample cases, using a simple mental healthcare AI as an example. Watch the video to see how we make better choices for safe and effective AI products.

42

2 commentaren

Loooove this post by Veerle Eeftink - van Leemput on Shiny and LLMs! 🤖 https://lnkd.in/dshZtXMd I've also been developing training materials on LLMs in Shiny recently, and it's become apparent very quickly that thinking about user experience is even more important when integrating LLMs into apps 👨‍💻 Some of the thing to consider: ⚡️ Handle requests asynchronously so multiple users don't block each other ⚡️ Provide visual feedback when components are loading ⚡️ Catch errors and display user-friendly messages #rstats #llms #shiny

29

1 commentaar

The AI/software space is in speedrun mode. OpenAI and others keep dropping new models. The barrier to entry? Lower than ever. Last December, I told Mees Konijnendijk 🌳: "We’re going to see a flood of new AI agencies in 2025. Some might become competition." Well—here we are. Every week, I see another AI agency or software company pop up on LinkedIn. And honestly? 𝗧𝗵𝗮𝘁’𝘀 𝗮 𝗴𝗼𝗼𝗱 𝘁𝗵𝗶𝗻𝗴. The market is moving absurdly fast, and 𝘁𝗵𝗲𝗿𝗲 𝗮𝗿𝗲 𝗴𝗮𝗽𝘀 𝗲𝘃𝗲𝗿𝘆𝘄𝗵𝗲𝗿𝗲: → Some companies need custom GPTs → Some need AI education & training → Some need AI workflow automation → Some need full-on custom AI products 𝗕𝗨𝗧. Here’s where it gets interesting: The ability to start an AI agency is 𝗲𝗮𝘀𝘆. The ability to deliver working AI solutions? 𝗡𝗼𝘁 𝘀𝗼 𝗺𝘂𝗰𝗵. Companies aren’t looking for hype. 𝗧𝗵𝗲𝘆’𝗿𝗲 𝗹𝗼𝗼𝗸𝗶𝗻𝗴 𝗳𝗼𝗿 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘄𝗼𝗿𝗸𝘀. The real winners won’t be the ones with the biggest LinkedIn presence. They’ll be the ones with working, proven, and successful AI products. By the end of 2025, we’ll see just how many AI agencies survive this phase. Who do you think will last?

35

“We met on a dating app.” - still felt a bit awkward a few years back. “I got this job without AI.” - might sound just as strange in 2030. Benedict Evans just published his AI report, diving deep into data on the platform shift from internet → mobile → AI. When dating moved to mobile, it wasn’t just a new channel. It was a paradigm shift in discovery: * You no longer met people only through friends, school, or geography * Algorithms started learning your preferences faster than your social circle * The surface area of who you could meet exploded That’s why the curve in this chart is so steep: in less than a generation, meeting online went from edge case to default. We’re now at the same kind of inflection point for careers—but this time the shift isn’t mobile. It’s AI. Today, most careers are still shaped by: * Who you happened to work with * Which brands are on your CV * How much time you are willing to spend scrolling through job postings That’s the “pre-Tinder” world of work. An AI-native talent market looks very different: 1. From static CVs to a deep understanding of what makes you unique and what you are truly looking for in a role. 2. From countless job applications to continuous matching. Instead of you chasing job posts, AI will bring the jobs to the talents. 3. Models can see that you shipped 3 products in Python last quarter, not just that you list 'Python' on your resume. In other words: what mobile did for “who you date”, AI will do for who you hire and who you work for. At Clera, we’re building an AI headhunter for this new reality. Not another resume filter, but an engine that understands talent, context, and potential—and proactively introduces the two sides. If Tinder could reshape who we marry, imagine what AI will do to connecting talent with the best suited opportunities, not just the next best one. So many people are stuck in roles that are OK'ish but never take action. There is still too much friction in the hiring market and we've set out to change that :) Can't wait for this graph to go up! #AI #FutureOfWork #Hiring #Talent #Recruiting #CareerDiscovery #Clera

358

5 commentaren