Cohorte

Hello world 🖖 We’re living exciting times, isn’t it? We can work from everywhere, learn anything, collaborate with anybody, and create almost anything on the web! There are limitless possibilities and every day many people are changing their lives and adapting their work to their lifestyle - not the reverse. This is particularly true for tech and data workers. You need to access data, and a machine terminal and that’s it 🚀 At Cohorte, we believe in a future where people & businesses collaborate and build together with high agility and minimal friction. A new order where intermediaries and rigid organizations give way to communities of independent and passionate talents to self-organize and deliver their best work every day while finding meaning and fulfillment. In 2022, companies need access to on-demand expertise from the best talents worldwide to build fast, shift quickly, and adjust to production demands. Technology is becoming an essential component of any business, and having the right mix of skills is critical for success. That's why we want to create a platform where skilled and passionate people with different technology skills can organize themselves to collaborate and complete complex projects. We believe that between freelance consulting catalogs and traditional consulting firms, there's a place for an alternative order, where people can be independent, enjoy flexibility, join a structured network and expand their business. In such an order, there is no room for subordination, but only for supportive and mentoring relationships, where people grow together, and experienced talents help younger ones to succeed. On the one hand, independent workers enjoy flexibility and freedom without sacrificing learning from peers, the scalability of their work, and the scope of their missions. On the other hand, businesses benefit from the agility and premium services from highly committed and cohesive professionals. ⚡️Do you believe in our project? Click the “Follow” button to support us. ⚡️If you are a tech/data freelancer, join our network, get access to our opportunities and discover our capabilities to scale and expand your business. ⚡️If you are a business, follow us, and boost your projects now by onboarding the best tech and data collectives with the right combination of skills. Let’s shape the new normal 🙌 #technology #data #consulting

AI alignment just got a significant upgrade. The latest innovation in AI training, Trust Region Direct Preference Optimization (TR-DPO), is pushing boundaries. Alexey Gorbatovski and his team at Tinkoff reveal how this method outperforms traditional Direct Preference Optimization by introducing dynamic reference model updates during training. Here’s why it matters: - 19% Performance Boost: TR-DPO enhances model alignment by periodically updating the reference model, resulting in a substantial 19% improvement. - Enhanced Human-Centric Metrics: Models trained with TR-DPO show marked improvements in key areas—coherence, correctness, detail, helpfulness, and harmlessness. - Versatility: The method proves effective across diverse datasets and model sizes, indicating its broad applicability in AI development. Why TR-DPO could be a game-changer: - Seamless Integration: Enhances existing models without the need for extensive retraining. - Targeted Precision: Fine-tunes model behaviors to align closely with specific objectives. - Training Efficiency: Reduces the dependency on manual adjustments, streamlining the process. For a deeper exploration of TR-DPO and its potential to revolutionize your AI systems, read the full research paper. https://lnkd.in/dHi3p2-r ______________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Alexey Gorbatovski, Boris Shaposhnikov, Alexey Malakhov, Nikita Surnachev, Yaroslav Aksenov, Ian Maksimov, Nikita Balagansky, Daniil Gavrilov

The line between technology and humanity just got blurrier. China's unveiling of lifelike female robots at the Beijing World Robot Conference is a major leap—and a bit unsettling. Imagine walking into a room and not knowing if you're talking to a person or a machine. These robots aren't actors playing a role; they're actual humanoid robots, engineered to mimic human appearance and behavior. With 27 types of humanoid robots on display, the blend of innovation and eeriness is hard to ignore. It’s fascinating to see how far robotics has come, but it also raises questions about the future. Will we prefer our robots to look distinctly robotic or human-like? For me, the uncanny valley is real—I'd rather my future robot look like a robot, not a near-perfect replica of a person. It keeps the lines clear between us and the machines. What about you? _____________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Rinor Restelica

Elevate Your AI Projects with Azure AI Search and LlamaIndex. Here’s how this integration redefines retrieval-augmented generation (RAG): - Pre-retrieval Techniques: Start strong with query transformations like rewriting and sub-question generation, laying the groundwork for precise data retrieval. - Advanced Retrieval: Harness Azure AI Search for hybrid search and semantic ranking, guaranteeing the most relevant outcomes. - Effortless Integration: Scale seamlessly with platforms that boost both performance and scalability. Why it matters: - Real-time Data Access: Your AI stays updated with the latest data. - Enhanced Control: Full control over data processing, from input to output. - Scalability: Designed for complex queries at scale, effortlessly. Want to implement these features in your projects? Full guide here:https://lnkd.in/gnSM3Vam ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Khye Wei, fsunavala-msft

Understanding the 𝗺𝗼𝗱𝗮𝗹𝗶𝘁𝘆 𝗴𝗮𝗽 is crucial for effective 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹.  Here's what’s really happening: When using CLIP models to embed text and images for retrieval tasks, you expect a unified embedding space. However, the PokémonCards dataset experiment reveals otherwise. Despite training CLIP models to align images and text, their embeddings don’t merge into a single coherent space. Instead, they split into two distinct subspaces, creating a clear boundary—the modality gap—between text and image data. This gap arises from: 1. Initialization bias (Cone Effect):   - The initial embeddings are biased towards forming cones, pulling text and image data into separate clusters. 2. Temperature reductions during training:   - Lower temperatures in contrastive learning intensify this bias, making it hard to overcome. 3. Contrastive learning's unintended reinforcement:   - Contrastive methods, meant to align image-text pairs, ironically accentuate the gap, maintaining the divide rather than closing it. This separation is why you can’t simply use CLIP models to retrieve and sort text and image data simultaneously by score. _____________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Han Xiao

AI is learning to answer real-world questions—but it’s not there yet. Meta AI's new benchmark, OpenEQA, challenges AI to understand and respond to questions about real environments, pushing the boundaries of Embodied Question Answering (EQA). What’s EQA? EQA demands AI to process visual data and answer open-ended questions in natural language. Think of asking, "Where are my keys?" and getting an accurate reply like "On the kitchen counter." OpenEQA explores two critical scenarios: - Episodic Memory (EM-EQA): AI uses a history of visual observations to answer questions. It’s like a smart assistant recalling past experiences. - Active Exploration (A-EQA): AI actively gathers information by exploring an environment, similar to a robot searching your home to find those keys. Why OpenEQA is a game-changer: - Open-vocabulary: No predefined question-answer sets, making it more realistic and challenging. - Real-world environments: Based on actual data from homes and offices, unlike earlier synthetic benchmarks. - Automatic evaluation: Uses large language models (LLMs) to score open-ended answers efficiently. The results? Even the most advanced models, including GPT-4V, struggle to match human performance, highlighting the complexity of EQA. Yet, OpenEQA is setting the stage for future AI that can interact with and understand the physical world in unprecedented ways. OpenEQA isn’t just a step forward—it’s a glimpse into the future of AI assistants. _____________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news.

Event-driven agent building is the game-changer you’ve been waiting for. Here’s why it stands out: - Cyclic, Multi-Agent Systems: Perfect for complex communication patterns and looping agents.    - Flexible Workflows: Easily implement advanced agent reasoning systems like ReAct and function calling.    - Native Tracing: LlamaTrace (Arize Phoenix) provides full visibility into your agent’s behavior, ensuring you know exactly what’s happening at every step. Dive deeper with these resources: - Blog: https://lnkd.in/ga_NWUPn - Video: https://lnkd.in/gHhnCi6m - LlamaTrace: https://llamatrace.com/ _____________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: LlamaIndex

Unraveling General World Models with Sora Explore the latest insights into general world models with the comprehensive survey paper "Is Sora a World Simulator?" Key Highlights from the Paper: - Sora Model Overview: Investigates the remarkable simulation capabilities of the Sora model, demonstrating an initial comprehension of physical laws. - Generative Methodologies: Examines the role of generative methodologies in video generation, showcasing how these models facilitate the synthesis of highly realistic visual content. - Autonomous-Driving World Models: Discusses the indispensable role of world models in reshaping transportation and urban mobility through autonomous driving. - Autonomous Agents: Explores the significance of world models in enabling intelligent interactions within dynamic environmental contexts. For a detailed exploration, read the full paper here: https://lnkd.in/d5Huewj7 ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Nianchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, Yang You, Zhaoxiang Zhang, Dawei Zhao, Liang Xiao, Jian Zhao, Jiwen Lu, Guan Huang

These Viral AI Videos Are Taking the Internet by Storm AI-generated videos are captivating audiences across the globe, quickly becoming the next big trend on social media. From stunning visuals to mind-blowing animations, these viral AI videos are not just eye-catching—they’re reshaping how we think about content creation. But how are people making these videos? The good news is, it’s easier than you might think. You don’t need a team of experts or hours of work; you can create your own viral AI video in just 2 minutes. _____________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. Credits: Zain Kahn

Taming Complex PDFs for Advanced RAG 🦙 Building Retrieval Augmented Generation (RAG) systems on complex documents like research papers with tables, charts, and images can be challenging. LlamaParse extracts various components, including: - Text - Tables - Images - Graphs - Custom parsing instructions By structuring data effectively, LlamaParse enables: - Improved data quality: A crucial factor for building reliable and accurate RAG systems. - Reduced hallucinations: LLMs are less likely to fabricate information when provided with well-structured data. - Enhanced understanding: LLMs can better comprehend the relationships between different data elements. Check out the workshop and see how LlamaParse can elevate your RAG projects! Workshop video: https://lnkd.in/gjZrKFGQ ___________________ ✔️ Click "Follow" on the Cohorte page for daily AI engineering news. #LlamaParse #RAG #PDF #DataExtraction #AI #LLMs