Michael Burian

I highly recommend Algorithms for Big Data by Hannah Bast, Claudius Korzen, Ulrich Meyer, and Manuel Penschuck for anyone looking to explore cutting-edge approaches in big data processing. This open-access book dives deep into scalable algorithms tailored to handle massive datasets efficiently, covering topics like data streaming, graph processing, and external memory algorithms. It’s a valuable resource for researchers and professionals in data science, engineering, and beyond. Grateful to see such high-quality knowledge made freely accessible! https://lnkd.in/eSfVM6k5 Look for "Read and Download Links" section to download. #BigData #DataEngineering #OpenAccess #DataScience #DataStreaming #DataProcessing

3

Curtain up for the 11th KnowHow-Meeting (20 - 22 Aug 2024) ... Governing the Data Mesh (Lessons Learned) by Tiankai Feng This keynote after lunch will certainly keep you awake :) Decentralized Domains and Data Products - the Data Mesh promises scalable data value generation but stands in contrast to traditional centralized Data Governance. Tiankai learned a lot from his journey both as Data Governance Leader on the client side, as well as now a Consultant, and would like to share some learnings on best practices of governing the Data Mesh effectively. Very attractive participation fees ... If you book the KnowHow Meeting (21 Aug 2024) plus two workshops (20 Aug & 22 Aug 2024) by 30 June 2024, the participation fee for users, for example, is only €248 (net). Information about the 11th KnowHow-Meeting, registration (early bird discount!) and booking a room at the Steigenberger Hotel Hamburg ... https://lnkd.in/ecwbTzde #metadata #datagovernance #dataquality #datacatalog #datamesh #datastrategy #knowhowmeeting #datacampus #dataculture

20

3 Kommentare

Don't call it a comeback. Tables have been here for years. And they contain trusted facts that will make or break the success of GenAI projects. That's my take after completing our report on retrieval-augmented generation, aka RAG, with my BARC colleague Florian Bigelmaier. We explore three architectural approaches to RAG: vector, relational, and graph, and assert that most RAG implementation should have multiple approaches. While most companies still struggle to govern their tables, generally they are more accurate than unstructured objects such as documents or images. So they are a good source of data for RAG, which injects governed content into GenAI user prompts. (BTW, forgive the pop culture reference to LL Cool Jay. My Florian got me thinking in rap terms with his recent post.) 🎤 Below are excerpts from our report, sponsored by Vespa.ai, Vectara, and unstructured.io, with a focus on relational RAG. What do you think? Relational RAG retrieves and calculates trusted values, such as purchase amounts, product prices or financial results, by querying relational database queries to ensure high levels of accuracy. When customers ask about their bill or insurance subscribers ask about claim status, the language model needs to give them a correct answer – and the correct values often come from a relational database. Relational RAG helps casual users by increasing access to insight compared with visual data analysis workflows and traditional SQL queries. For example, more and more BI vendors include such functionality in their software. Here is the workflow in our diagram, from the bottom up. > Data Relational RAG begins with well-organized relational data such as transactional records or customer data arranged in tables with clearly defined relationships. This requires data teams to have mature processes and governance practices in place beforehand. > Pipelines Pipelines prepare data for consumption in a RAG workflow. For example, they might extract, transform, then load it into the database (ETL). Analysts might also calculate relevant metrics in this step. The pipeline also collects and catalogs metadata to assist subsequent steps. > Database The relational database stores the resulting analytical data, including cleaned operational records and pre-aggregated metrics. Metadata describes the complex data structures and abstracts them into business-friendly terms. > Applications/Agents. The application maps the relevant portions of the user input to a database query language such as SQL. The metadata can also populate a dictionary that supports natural language queries. The application retrieves the data and the language model generates an answer. In addition, the language model can generate the database query itself. Architectural variations are possible; for example, the workflow might use a REST API to retrieve structured data instead of querying the database directly. #rag #data #genai #ai

43

9 Kommentare

What Karl says. There's going to be two amazing talks about Hopsworks based #ai and #ml systems: * Build a personalized Bitcoin (BTC) virtual assistant in Python with Hopsworks and LLM function calling by Javier de la Rúa Martínez (https://lnkd.in/eQfWpFRU ) * Build TikTok's Personalized Real-Time Recommendation System in Python with Hopsworks by Jim Dowling (https://lnkd.in/eXEHePjG) If you are in the #berlin area next week - go and take a look!

7

“𝗗𝗼𝗲𝘀 𝗚𝗲𝗻𝗔𝗜 𝗥𝗲𝗽𝗹𝗮𝗰𝗲 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗛𝗮𝗿𝗺𝗼𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻? 𝗡𝗼𝘁 𝗦𝗼 𝗙𝗮𝘀𝘁!” 🤓 The rise of generative AI, particularly LLMs, has sparked a misconception: that these models can bypass the need for semantic harmonization through business information models. However, the reality is quite the opposite—semantic harmonization remains critical, even with GenAI as a powerful enabler. 💡 And here are the reasons why: 1. 𝙀𝙣𝙝𝙖𝙣𝙘𝙚𝙙 𝙇𝙇𝙈 𝙋𝙚𝙧𝙛𝙤𝙧𝙢𝙖𝙣𝙘𝙚: LLMs deliver higher accuracy and efficiency when operating on semantically harmonized data. Without it, they expend unnecessary effort trying to “understand” data meanings. 2. 𝙉𝙤𝙩 𝙀𝙫𝙚𝙧𝙮 𝙐𝙨𝙚 𝘾𝙖𝙨𝙚 𝙞𝙨 𝘼𝙄-𝙙𝙧𝙞𝙫𝙚𝙣: Generative AI isn’t a one-size-fits-all solution. Many tasks still rely on standardized, well-structured data products beyond AI’s immediate scope. 3. 𝙁𝙤𝙪𝙣𝙙𝙖𝙩𝙞𝙤𝙣𝙖𝙡 𝙛𝙤𝙧 𝘿𝙖𝙩𝙖 𝙈𝙖𝙣𝙖𝙜𝙚𝙢𝙚𝙣𝙩: Harmonized semantics play a pivotal role in automating processes like master data management and ensuring data quality—cornerstones of any robust data strategy. 4. 𝙂𝙚𝙣𝘼𝙄 𝙖𝙨 𝙖 𝘾𝙖𝙩𝙖𝙡𝙮𝙨𝙩: While GenAI doesn’t replace the need for semantic harmonization, it can accelerate the process. It’s a game-changer for optimizing models and mapping them to physical data representations. 𝙎𝙤 𝙬𝙝𝙖𝙩 𝙙𝙤𝙚𝙨 𝙩𝙝𝙖𝙩 𝙢𝙚𝙖𝙣 𝙛𝙤𝙧 𝙮𝙤𝙪? To truly conquer your data challenges, embrace the synergy: use semantic harmonization as the foundation and GenAI as the accelerator. Together, they’re the ultimate duo for unlocking the full potential of your data. #DigitalTransformation #GenAI #DataStrategy #SemanticHarmonization #AI

45

7 Kommentare

I'm starting to question the necessity of vector databases as standalone part of the pipeline for building AI applications. With new large-context models like #Gemini 1.5, and new functionalities like `file_search` in #OpenAI's latest Assistant API, the integration of context has become more direct and streamlined. The Assistant API, for example, now allows the creation of RAG setups with up to 10,000 files. This approach simplifies the process significantly compared to the complexities involved in managing a standalone vector database such as Qdrant. It seems the trend is moving towards more integrated and user-friendly AI tools. #RAG #LLM https://lnkd.in/e4ZSPzgF

45

49 Kommentare

vCD RoundTable Episode 28: On vApps with Tobias Paschek (he/him), Matthias Eisner, Sascha Schwunk is online at https://bit.ly/45cjlnY Use vApps to streamline your infrastructure and manage VMs effortlessly! In this episode, we dive into the purpose, management, network types, implicit fencing, and automation of vApps. We further unpack key advantages, expert tips, and potential pitfalls. Available now and join us live next time to get your questions answered! #VMware #Broadcom #vApp #TechInsights2024

3

Want to get started with AI vectors in the 23ai Oracle database ? Have a look at the post we wrote, my colleague Ulrike Schwinn and I ! In this blog post we cover a few key tasks using SQL, PL/SQL, and ORDS: - Loading an ONNX embedding model into the database - Generating vector embeddings using the loaded vector embedding model - Storing vector embeddings in a table with a data type VECTOR column - Using similarity search in SQL and ORDS Feel free to give it a try with the 23ai Free version and have fun! #oracledatabase #23ai #23aifree #oraclecloudinfrastracture #vector #similaritysearch #AI #ORDS #SQL #tutorial #embeddings Blog post: https://lnkd.in/eZBDRcfu

27

Not all digital transformations are fully integrated into existing processes. Often, gaps remain where old methods and new technologies don't quite align. What strategies do you employ to ensure that #digitaltransformation is fully integrated into your business processes? #DigitalTransformation #ProcessIntegration #TechAdoption #ContinuousImprovement #InnovationBlend

1

💡 According to #OpenAI GPT-4o is supposedly 50% faster than GPT-4. Here at #CapgeminiInvent we decided to verify this claim, some initial results below👇 🏭 During the industrialization phase of the #GenAI solutions we have implemented, we discovered token generation speed is much more critical for our clients than actual performance of the inference. 📉 We conducted stress tests in June on our Microsoft Azure endpoints (all MaaS in the US-East-2 region) to measure the latency of the different models we use for our clients. Below is the distribution of inference speeds (in tokens per second) for 13 different models, ranked from best to worst according to the lmsys leaderboard. Here are some interesting insights: ➡️ On average, at 33 tokens/s, #GPT-4o is 80% faster than #GPT-4-Turbo and 20% faster than #GPT-4. #GPT-3.5 still holds the first place with 46 tokens/s. 🥇 ➡️ OpenAI models exhibit a wider distribution in latency compared to other models, with significant daily and weekly variance. ⏱️ ➡️ For other models, throughput is much more predictable over time, especially for #CommandR+, #Mistral-Large, #Phi3-Medium, and #Mistral-Small. This is likely linked to the underutilization of the hardware provisioned to host these endpoints. ➡️ Very small models, like #Phi3-Mini, are fast but not as fast as one would expect. Microsoft has probably provisioned smaller hardware accordingly to run these endpoints. 📖 These results are preliminary, of course, and warrant a more comprehensive analysis, but we can already make the following recommendations: ✅ Be mindful of possible fluctuations in response times. Test various regions and models to identify the most suitable endpoint for your requirements. ✅ If speed is of utmost importance, aim for a speed/performance tradeoff. I’d recommend Cohere’s Command R+, which offers the highest speed/predictability ratio. ✅ For small or very small models, consider self-hosting your language model (LLM), as the existing infrastructure for these models as a service may be insufficient. Harmanvir Singh Valentin Werner Moïse Tignon Volker Darius Mathieu Dougados Jérôme Coignard Philippe Cordier Pierre Demeulemeester Kary Bheemaiah Nicolas Gaudilliere Philipp Wagner Conor McGovern Ian Davies Steve Jones Robert (Dr Bob) Engels #LLMLatency #InferenceSpeed #Azure #MaaS

139

10 Kommentare

🇪🇺 What are your biggest pain points when operationalizing the #EUAIAct? Over the past couple of years, we at appliedAI Initiative GmbH have been diving into the complexities of the #EUAIAct for large organizations. As regulatory requirements and standards continue to shape the way we develop and deploy AI solutions, we have observed recurrent pain points among companies at the organizational, process, and use case-level. Understanding these challenges is the key cornerstone to develop the right strategy to not only comply but also continue scaling AI in production. 💡If you want to learn more insights about these challenges and discuss best practices join me and my colleague Ronnit Wilmersdörffer at the ai:NOW conference, in the panel: 'Pain Points of the EU AI Act & Business Implementation', on October 9th in Munich. Reach out if you need a ticket. Looking forward to an enriching exchange with our panelists: Heiko Joerg Schick Frank Stadler Michael Reinle Veaceslav Driglov, MBA Julia Reznitcaia Ronnit Wilmersdörffer #AIAct #AICompliance #AIGovernance #BusinessStrategy #MLOps #EUAIAct Akhil Deo Manuel Jiménez Mérida Claudia Baumgartner Dr. Till Klein Hendrik Scherner Anish Pathak Dr. Philipp Hartmann Scott Raven Charmaine Baker

46

10 Kommentare

SpreadsheetLLM 📈📊 https://lnkd.in/eCwx5aTv "Spreadsheets are characterized by their extensive two-dimensional grids, flexible layouts, and varied formatting options, which pose significant challenges for large language models (LLMs). In response, we introduce SpreadsheetLLM, pioneering an efficient encoding method designed to unleash and optimize LLMs’ powerful understanding and reasoning capability on spreadsheets."

15

3 Kommentare

🆕Announcing the Cloud adoption framework for AI🆕 1️⃣ Key Azure AI services 👉 2️⃣ Six-Phase AI Adoption Process  3️⃣ Tailored Checklists for Startups and Enterprises Learn more 👉https://lnkd.in/e3mUGdfQ #ai #microsoft #programming #development #openai #msft

212

6 Kommentare

💥 Misconceptions and the path to your Data Science career! - Tomorrow on The Data For Good Podcast, with Data Science Leader Evgeniy Savin from neuefische GmbH - School and Pool for Digital Talent 🔎 👩‍💻 #dataforgood #berlin #datascience #zuma

8

3 Kommentare

#rag #genai and now #tag In the context of tables, RAG involves strategies like passing semi-structured documents with tables into LLMs, extracting tables for targeted retrieval, and preserving table structures during document chunking. This approach is crucial for applications involving complex documents with embedded tables, such as financial reports or technical manuals

10

In this article, we will talk about the capabilities of #OpenAI DALL-E models in generating images based on the description provided by users in our #Delphi #FMX app by means of OpenAI #API. https://lnkd.in/eXT6BK9g #Embarcadero #DelphiDevelopers #AI #ArtificialIntelligence #DelphiApp #DelphiModernization

16

GeoSphere Austria & Österreichische Hagelversicherung VVaG watch out: IBM and NASA have unveiled a groundbreaking open-source AI foundation model for weather and climate applications on Hugging Face, offering weather/climate insights beyond traditional forecasting: • the data are pre-trained on 40 years of NASA - National Aeronautics and Space Administration Earth observation data (MERRA-2) • Applicable for short-term weather and long-term climate projections • The model can depict both weather and climate data at up to 12x resolution, generating localized forecasts and climate projections cause of data downscaling and gravity wave parameterization • As a foundation model, it has a unique architecture which allows it to be fine-tuned for global, regional, and local scale fine-tuning This would have been great to predict the recent flood we have been suffering in lower Austria and generally speaking in understanding meteorological phenomena, improve climate risk assessment, and enhance severe weather predictions. Looking for a partner for implementation and integration? The CANCOM Austria Digital Makers are there for you #OpenSourceAI #ClimateScience #IBMResearch #NASATechnology #DataScience #SustainableTech #ClimateTech #ArtificialIntelligence #digitalmakers

8

1 Kommentar

Meta Llama 3, the next-gen open large language model, is now available on the watsonx AI and data platform. This partnership between IBM and Meta will advance open innovation for AI. With pretrained and fine-tuned models featuring 8B and 70B parameters, Llama 3 supports a wide array of use cases, from summarization to content creation. IBM's collaboration with Meta and engagement with enterprises showcases the real-world impact of these models. #AI #Innovation

9