Introducing Wood Wide AI: A New Table-Native AI

New code-along series: From AI hype to building AI value. Too many AI initiatives stall at the idea phase. Let's change that. Join DataCamp's Build Week: AI, Data, Copilots (Nov 17–24) to transform experimentation into execution. Through hands-on code-alongs, learning tracks, and real-world projects—learn how to build AI that delivers real (and measurable) value. What you'll build:  -> A deepfake avatar: Combine video, audio, and image-generation AI -> AI that works: How to avoid the 95% of AI projects that fail -> A GPT-like LLM from scratch: Train and evaluate with PyTorch -> ROI-ready ML evaluations: Estimate the business value of a model Whether you're a technical leader or individual contributor, walk away with tools, strategies, and confidence to deliver AI projects that launch. Save your seat and start building: https://ow.ly/8xTq50XiUCJ Jonathan Cornelissen (DataCamp) Lexi Reese (Lanai) Yusuf Saber (Optima) Rhys Phillips (DataCamp) Eric Siegel (Gooder AI) Jim Liddle (Nasuni) Rajeev Butani (MediaMint) Jonathan Welch (Albert Invent) Jacob Buckman (Manifest AI)

DeepSeek’s New OCR Model Might Redefine How We Handle Long-Context AI!!!!! As a Data Science professional, I’m always fascinated by innovations that push the boundaries of how AI processes information — and the recent release of DeepSeek-OCR is one such breakthrough that deserves attention. This model doesn’t just improve Optical Character Recognition; it reimagines how text can be compressed, represented, and understood by large language models. The Core Idea Instead of treating text as a sequence of tokens, DeepSeek-OCR treats it as a visual medium. It converts large bodies of text into high-resolution visual representations — and then decodes them. This approach reduces token usage by nearly 10×, while maintaining around 97% accuracy. That’s a significant leap in how efficiently models can handle long documents, research papers, or even entire books. (Reference: DeepSeek-OCR Paper on arXiv) Why It Matters To me, the importance of this development goes far beyond OCR: - Efficiency: Visual compression drastically cuts down compute cost and memory load. - Scalability: Models can now process much longer contexts without running into token limits. - Enterprise Potential: This could transform document-heavy industries like finance, legal, and healthcare — where large reports and archives need to be analyzed in real time. The Bigger Picture This move signals a shift toward multimodal reasoning, where AI no longer “reads” or “sees” separately — it does both, in harmony. If visual tokenization continues to evolve, it could redefine the future of context-aware AI, enabling systems to process massive amounts of information more efficiently than ever before. As someone passionate about AI and Data Science, I see this as more than an incremental improvement — it’s a new direction in model design, where vision and language finally merge to expand what AI can truly comprehend. #DeepSeek #AIInnovation #ArtificialIntelligence #MachineLearning #MultimodalAI #DeepLearning #OCR #DataScience #GenerativeAI #TechTrends #AILeadership #FutureOfAI

🧠 The Tiny Recursive Model: Less is More: Recursive Reasoning with Tiny Networks We're witnessing something remarkable in AI: a 7-million parameter model outperforming billion-parameter LLMs on complex reasoning tasks. Meet the Tiny Recursive Model (TRM) — and it's changing how we think about AI efficiency. What is TRM? TRM is a recursive reasoning approach that uses a single tiny network with just 2 layers and 7M parameters. Instead of relying on massive scale, it recursively improves its answers through iterative refinement. How Does It Work? Think of it like solving a puzzle: 1️⃣Start with an embedded question and initial answer 2️⃣Recursively update the internal state based on the question, current answer, and latent understanding 3️⃣Refine the answer iteratively for up to K improvement steps 4️⃣Each iteration brings it closer to the correct solution The Results Speak for Themselves 1️⃣45% accuracy on ARC-AGI-1 and 8% on ARC-AGI-2 2️⃣87.4% accuracy on Sudoku-Extreme 3️⃣85.3% on Maze-Hard tasks 4️⃣Outperforms Deepseek R1, o3-mini, and Gemini 2.5 Pro on these benchmarks. All with less than 0.01% of the parameters of traditional LLMs. 👉Why This Matters 🔹Efficiency Over Scale: We don't always need bigger models. Sometimes smarter architectures win. 🔹Sustainability: Smaller models mean lower computational costs, reduced energy consumption, and more accessible AI. 🔹Specialized Reasoning: For logical puzzle-solving and structured reasoning tasks, recursive refinement proves more effective than brute-force scale. 👉The Bigger Picture TRM challenges the "bigger is better" narrative dominating AI development. It shows us that: 🔹Architecture innovation matters as much as scale 🔹Recursive reasoning can match or exceed transformer performance on specific tasks 🔹The future might be hybrid: specialized small models for specific reasoning + LLMs for general knowledge The race isn't always to the biggest model. Sometimes, the smartest approach wins. #AI #MachineLearning #DeepLearning #ArtificialIntelligence #Innovation #TechTrends #Efficiency

🤖 Boost AI Model Accuracy by 45% with High-Quality Datasets 📈 The success of any AI system depends on one key factor — data quality. Our latest research highlights how the right #DatasetsForAI model training can drastically improve performance, accuracy, and scalability. 💡 Here’s what we found: 📊 45% accuracy boost by refining data selection and removing bias. 🧠 Smarter model behavior — better generalization and reduced overfitting. ⚙️ Scalable data pipelines that fuel faster retraining cycles and real-time updates. 🌍 Domain-specific datasets designed for AI in retail, finance, healthcare, and more. “Well-structured, diverse, and bias-free datasets are the foundation of truly intelligent AI.” From computer vision to NLP, #CuratedDatasets are powering the next generation of intelligent automation — and redefining how data drives decisions. 🚀 👉 Read More: https://lnkd.in/dWiqucWD 🔖 Hashtags #ProductDataScrape #AIModelTraining #MachineLearning #ArtificialIntelligence #BigData #DataScience #ModelAccuracy #AIInsights #TrainingData #DataDriven #Analytics #AIML #Automation #SmartData #DeepLearning #TechInnovation #ModelOptimization #NeuralNetworks #AIResearch #BusinessGrowth #AITrends #DataQuality #AIDatasets #DigitalTransformation #TechSolutions

We recently had an interesting (and slightly painful) lesson in using AI to train AI. We were testing a dual-model setup for transaction categorisation. Using a large model to train a smaller model or ML model. Clean and simple; Two AI models would classify each transaction. If both models agreed, we’d take that consensus as high-confidence data to feed into our machine learning training set. If they disagreed, the transaction would go to our human tagging team for review. In theory, it was an elegant way to scale high-quality labelled data. In practice, our AI models found consensus but only on one or two categories. Those “agreements” dominated the dataset, and our ML model learned that almost everything appeared to be “groceries_and_meals.” 🍎🥦🍔 The accuracy of our categorisation model deteriorated significantly, and all of a sudden, it became irrelevant. It’s a good reminder that AI pipelines amplify whatever assumptions (or imbalances) we build into them. Crap in means crap out We have reverted to our previous best steady state model and are going to look into ways we can debug our distillation process. AI isn’t magic, it’s math, data, and constant iteration.

🔥💥The Wikidata Embedding Project: Semantic structured data source for entities and facts as a vector database for LLMs👇 With the Wikipedia Embedding Project, LLM providers such as OpenAI and Google now have the opportunity to access this semantic database. 📚 Wikidata is a free and structured knowledge graph that organizes information in a standardized way, unlike Wikipedia. 🔗 Its knowledge graph structure connects items through statements, creating a web of data for both humans and machines. 🎯 The project's main goal is to create an open-source vector database from Wikidata to bridge its structured data with generative AI. ✅ Connecting AI to Wikidata helps reduce misinformation by grounding AI responses in verifiable and up-to-date sources. 🗺️ A vector database is built by converting Wikidata items into numerical vectors, which maps their contextual meaning. 🔍 Semantic search allows users to find contextually relevant information using natural language queries. 🔧 A standardized toolkit, the Model Context Protocol (MCP), enables AI models to directly interact with and retrieve data from Wikidata. 💡 Key uses for this technology include automated query generation, question-answering systems, and fact-checking. It is unclear whether they are already accessing it. If so, Wikidata is a valuable source for optimization approaches for GEO. We have been offering Wikipedia services for 10 years and now also offer Wikidata services. Wikidata is particularly worth looking at for company and person entities that are not relevant enough for their own Wikipedia entries. https://lnkd.in/eB8jmSK4

AI systems keep evolving but what if the real secret lies in how they understand and learn patterns? In this fascinating piece by Samuel Dagne on Mindplex Magazine, we explore the architecture powering today’s most advanced models and what it means for the future of intelligence. A must-read for anyone curious about the building blocks behind modern AI. https://lnkd.in/ekSQhJyu #AI #mindplex #mindplexmagazine

AI systems keep evolving but what if the real secret lies in how they understand and learn patterns? In this fascinating piece by Samuel Dagne on Mindplex Magazine, we explore the architecture powering today’s most advanced models and what it means for the future of intelligence. A must-read for anyone curious about the building blocks behind modern AI. https://lnkd.in/ekSQhJyu #AI #mindplex #mindplexmagazine

In this feature from AI Time Journal, Binny Gill, CEO of Kognitos, shares his vision for “English as code”—a world where anyone who can communicate can automate. His approach puts human intent and control back at the center of AI, making automation more natural, trusted, and accessible to all. 🔗 Check it out: https://lnkd.in/eH_PZvNM