Mark Mauerwerk

Earlier this week, I chatted with the experts about the latest version of small language foundation models from H2O.ai – Kaggle Grandmasters Philipp Singer, Pascal Pfeiffer, and Yauhen Babakhin. Today, I share with you their insights into the development and application of Danube2 🧠💡 Dive into our latest blog post where I unpack the features, benefits, and potential of H2O Danube2. Learn how it's setting a new standard for Small Language Models and how it can empower your organization to leverage AI like never before. 📈🤖 🔗 https://lnkd.in/ghf-AArb #H2Oai #Danube2 #SmallLanguageModels #AI #MachineLearning #Innovation #DataScience #KaggleGrandmasters #TechnologyNews #AIforGood

116

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

𝐔𝐧𝐥𝐞𝐚𝐬𝐡𝐢𝐧𝐠 𝐭𝐡𝐞 𝐏𝐨𝐰𝐞𝐫 𝐨𝐟 𝐆𝐞𝐧𝐀𝐈 𝐟𝐨𝐫 𝐄𝐧𝐞𝐫𝐠𝐲 𝐑𝐞𝐭𝐚𝐢𝐥 Was my Keynote topic at CIO Events Generative AI Summit in Munich this week. Here are my Top 3 takeaways: 1. 💶 𝐌𝐨𝐧𝐞𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐃𝐞𝐦𝐨𝐜𝐫𝐚𝐭𝐢𝐳𝐞𝐝 𝐆𝐞𝐧𝐀𝐈: Although more than half of the participants had in some for a <insert company name> GPT or similar platform, it remains difficult to monetize and fund aside from innovation budget. I'm therefore happy that we at E.ON have early on established a very successful license model for our internal "Chat with Your Data" (a LLM based RAG solution for specifc user groups) solution after democratizing GenAI by opening up our "Ask me Anything" (enhanced drop-in replacement for ChatGPT, but safe and secure) service to all 74k employees at E.ON. 2. 📈 𝐅𝐫𝐨𝐦 𝐇𝐲𝐩𝐞 𝐭𝐨 𝐇𝐢𝐠𝐡 𝐕𝐚𝐥𝐮𝐞 Now is the time to push high-value use cases forward, a company replacement for ChatGPT is only the start. NOW is the time to identify the highest value use cases. In Energy Retail the main cases I see revolve around the topic of customer service automation, three topics to highlight: automated email classification and information extraction, enabling customer service agents with natural language based RAG systems to get quick answers from their knowledge bases, and third automatic checks on bills and other documents. All of these are much easier now with LLM's. The marginal cost of asking LLMs to extract information is close to 1ct, this is 20-100 times cheaper than human labor on the same task! The APIs of our GenAI platfrom enable the cases! 3. 🤖 𝐖𝐡𝐞𝐫𝐞 𝐚𝐫𝐞 𝐚𝐥𝐥 𝐭𝐡𝐞 𝐂𝐨𝐏𝐢𝐥𝐨𝐭𝐬?  I'm not sure what your experience is, but the takeaway from the conversations for me was: many have tried especially the Office365 CoPilots, but hardly anyone is really convinced on their value including myself. The only CoPilot that was received positive by a majority were the GitHub CoPilots for development. What do you think about these? Special thanks to the organizers of the event, everyone for the great conversations and especially to Gerhard Holzwart for inviting me to speak! #data_on #welovetechnology #itsonus #ai #genai #generativai #cio

197

27 Kommentare

I attended a fantastic conference yesterday at the University of Bern: Synthetic Data Day (https://lnkd.in/dacTY7Qa). Great and relevant topic for AI in Healthcare, and excellent speakers - this could also be a very big conference in my opinion. #SyntheticData #PrivacyVsUtility #Healthcare #AI

42

The nuance difference in AI-driven software: from features to quality - written by our Data Scientist Edwin Bosscha. At Brainial, as a Data Scientist, I often notice that software development focuses on ‘features’. A feature is a specific functionality that software offers or does not offer. This black-and-white view of software capabilities is less and less relevant as AI plays a bigger role. The reality is much more nuanced. The answer is not always Yes or No AI-powered software often has probabilistic outcomes. For example, an AI model functions correctly 80%, 90% or 99% of the time, but that also means that it is sometimes less right or wrong. When developing and evaluating this kind of software, it is crucial to keep this in mind. There is no such thing as an AI model that is always correct. The challenge is to find a level that is practically usable in a business process. The complexity of generative AI With generative AI, this nuance becomes even more important. The output of a generative model can be correct or incorrect, as well as vary in wording and style. What is perfect for one user may be suboptimal for another. This shows that the value of AI-driven software depends heavily on the context and usage scenario. Quality over quantity The value of AI-driven software is determined by the quality of its output. A model that is correct 80% of the time may have little practical use in some situations. At 90% accuracy, the practical value increases significantly. Therefore, it is essential to judge AI software on how well it performs, not just on what it can do. Away from binary assessments Why do we stick to the binary framework of ‘can this be done?’. Instead, we should judge software by working with it, not by simply ticking off a list. Software and models should be compared based on quality and added value, not on simple questions like ‘can it do X?’. One of the key questions I ask are ‘What defines quality in this context?’ and ‘What is a minimum level of quality to add value?’ The future of AI-driven software To harness the true potential of AI-driven software, we need to focus on quality and context. This means shifting our approach from a focus on features to a focus on overall performance and usability. By testing the software in real-world scenarios, we can better understand and exploit its subtle nuances and real value-added. In short, AI-driven software offers a lot of potential but requires a shift in how we assess its capabilities and performance. Only by looking beyond binary ‘can this be done?’ questions can we fully exploit the true value of AI solutions and deploy them optimally for our specific needs.

11

🧠 The myth: More data makes an #AI model smarter and therefore more accurate 💪 The truth: It’s not just about data—it’s about #parameters With the around OpenAI’s #GPT4o it might seem that feeding more data into these models makes them smarter, the real game-changer is the number of parameters they have like Meta #Llama 3.1 with 450b parameters 💡 More parameters mean these models can process and understand data in more sophisticated ways, leading to increased accuracy and the appearance of being “smarter.” More parameters = higher trustworthiness, even with more focused data like the proprietary #data in your organization In the world of conversational user interfaces becoming the norm, more parameters will be crucial for a seamless experience. While yet too slow and too expensive for real fast interactions between humans and machines, with #quantumcomputing we will move in a new area of responsible #UI #design #AI #MachineLearning #LLM #conversationaldesign #hmi #Innovation

5

1 Kommentar

A very insightful reading: most #graph tools were designed and built long before the generative AI revolution. #GenAI use cases are fundamentally different from traditional graph use cases, and require a different approach, even if some tools can be shared between the two. So, always keep things simple: there is a chance that your Graph #RAG use case doesn't need a graph DB nor a graph query language. https://lnkd.in/eThka3a8

9

Been conflicted for a while about the utility of #Ai in the data space. #Ai rocks compared to #humans for repetitive task automation, pattern recognition and anomaly detection, analytics, multi-variable optimization problems, writing papers and creating art. But we don't see self-driving cars everywhere; there's a reason a driver must be present ... the general public, policy-makers and insurance companies feel more at ease with the #aihuman teamup. Check out the movie by Netflix called "Atlas" starring Jennifer Lopez https://lnkd.in/gMDdJ8p4 ("Atlas Shepherd (Lopez), a brilliant but misanthropic data analyst with a deep distrust of artificial intelligence") Is Ai reliable enough for cleaning #mrodata used in critical business operations? Ai is #aidelusional and #aischizophrenic when the #aiprompts or #aitraining data are "bad" or when the #algorithm isn't well-tuned. The #investment and #roi of Ai and the #losttime can cause #buyersremorse. Worse, especially if #baddata causes a #blackswan event. In that case, who do we blame, Ai or the #executiveteam? #Ai alone cannot solve #everything. If you test it, you'll find mistakes that then must be corrected by humans - this is the #aigap - the investment in Ai is not small, and because Ai makes mistakes, the output needs to be reviewed. Mistakes identified then must be fixed by humans, and the data Ai cannot handle must also be processed by humans ... we can use a process that is time-tested and known to work very well --> https://lnkd.in/emyCncUq (Sphera's Data Quality Optimization Service). Regarding highly critical #MRO #masterdata I'm not ready to fully #trust Ai, because there's too much at stake. There's far too much #risk. For non-critical, simple data Ai is fine - some mistakes are OK; plants will not blow up, supply chain will function (if not without errors), inventory may not be perfect, and sustainability might suffer a little. For everything else (that is #critical), make sure humans are involved with a bullet-proof process for curation, creation and quality control. Employ time-tested #datastandards #datagovernance #datacleansing #mdm and #dataquality. This way, improved outcomes in #supplychain, #inventory, #plantmaintenance, #processsafety and #sustainability can happen. At the end of the day, it's about five (5) things: (1) Quality, (2) Schedule, (3) Experience, (4) Service, and (5) Price. If #baddata is hurting your business, it's time to get it right. Check out @Sphera https://lnkd.in/gEyWmwGB or ping Dave Kuketz for a consultation.

16

"AI needs data more than data needs AI," as noted in a Forbes article last October - and Kai Müller at EVACO GmbH couldn't agree more. Join us in a conversation with Evaco to understand why exceptional AI cannot exist without good, clean, and well-prepared data: https://lnkd.in/eUq_fCic #AI #DataQuality #timextender

28

1 Kommentar

Defining the Meaning of Your Data – A Practitioner View Last Thursday, we had what could be described as the most exclusive data event of the year — the Semantic Layer Symposium. It brought together some of the most renowned data professionals, all gathered in a spectacular setting — the private library of the "Bayerische Hof Hotel” in Munich. Surrounded by hundreds of books, we dived deep into the heart of a essential topic: how to make data more meaningful, interconnected, and accessible to the people who use it. Fittingly, the profession of librarians — now being redefined by AI — served as an inspiration for this work. But that’s a story for another post. I had the privilege of joining the panel alongside Helen Lippell and Daniel Canter, expertly led by Sara Nash, where we shared practical insights on implementing semantic enrichment. Throughout my experience, I’ve learned a few simple tricks that significantly improve the readability, discoverability, and meaningfulness of a data estate. Here are the Top 3 that come to mind: 1. Master Metadata, Taxonomy, and Ontology   These are the building blocks of all the data you have — the “fishbone” that gives your data its core meaning. By defining and connecting referential data, you enable organizational units to work toward common goals and execute advanced analytics. It’s both a one-time setup and a continuous effort, as data evolves alongside business needs. So, plan accordingly for this dynamic process. 2. Build Simple Data Products for User-Centric Abstraction   Often, we have limited control over the semantics of source data (the infamous “systems of record”). Despite their continued relevance to the business, these systems can be challenging to decipher. Data products offer a great solution, allowing you to abstract from the source and add meaning tailored to the needs of your users. This makes your data estate far more user-friendly and impactful. Check excellent post by Karthik Ravindran about why Data Products matters. 3. Discover Data Governance for Unstructured Data   Historically referred to as “Data Lifecycle” management, this approach helps enhance the discoverability of unstructured data by focusing on policies, frameworks, archival strategies, and more. Take a closer look at your unused or partially used data — tools like the “Unused Data Framework” that Tiankai Feng and I presented a few weeks ago can help. The goal is to distill what’s meaningful and leave behind what’s not, improving the overall discoverability of your data estate. A big thank you to the organizers — Zach Wahl, Lulit Tesfaye, Benoit Gaussin, and Sara Nash — and kudos to all the presenters, including Malcolm Hawker, Daniel Canter, Tiankai Feng, and Norbert Gergely. Looking forward to the next year – with vectors winning over relational data, the conversation about semantics will received entirely new spin.. ;) #Metadata #SemanticLayer #DataGovernance #InformationManagement #EnterpriseKnowledge #genAI

54

5 Kommentare

🚀 Let`s talk about #DataStrategy & #DataGovernance in #Manufacturing 👨‍💻 Data strategies build the foundation on which #AI can thrive and reach its full #potential, as companies leveraging AI in their #manufacturing processes could experience productivity gains of 20%. 🤖 So .. .how to setup a value-added and human-centered Data Strategy? Tune in to our new Siemens AI podcast 🎧: "Human&AI: Mind, Machines, & the Gradient Descent" and join Arberie and me in this interdisciplinary 🏳️‍🌈 virtual format on the interfaces of what it means staying hashtag #human and being #passioned about technology and #IndustrialAI. This time, we are super excited to meet the genius mind Erik Schwulera IoT Lead for Manufacturing at Siemens Industry. Discover the world of AI integration in industrial settings, we explore all things data strategy and AI implementation Gain valuable insights on: 👉 What differentiates a good #datastrategy from a great one?  👉 How do you balance the importance of #domain know-how and #datascience expertise when developing AI solutions for industrial settings? 👉 What are some of the practical #challenges Siemens has encountered while implementing its AI strategy in real-world #shopfloor environments? 👉 What is the hype and specifics on #dataproducts and it`s marketplace charactersitics? 🙏 #teamawesome Alexander Ulbrich Anna Gagliano Bernd "Benno" Blumoser Robert Lohmeyer Boris Scharinger Cynthia Springer Jacqueline Burkert 🏳️‍🌈🤖🧡 #bold #committed #openminded #OneSiemens #SiemensAI #SiemensAILab #AIoT Siemens Advanta https://lnkd.in/dBrVuwuk

114

7 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

Clarifying Misconceptions around RWE in Pre-market Submissions: https://lnkd.in/dhTB6XFV #realworlddata #realworldevidence #rwe #rwd #evidence #value #data #dataanalytics #marketaccess #heor #FDA #clinicaltrials #pharma #biopharma #dataquality #bias #transparency #healthpolicy #realworldoutcomes #patientjourney #patientaccess #access #regulatory #regulatorysubmissions #healtheconomics #medicalaffairs #clinicaldevelopment #regulatoryaffairs

25

2 Kommentare

5 questions with Marcello Cacciato (Metyis) on bringing Generative AI to life Marcello Cacciato, Director in the Data Science practice of Metyis, is an expert Generative AI, having advised multiple organisations on the development and implementation of their Generative AI stra https://lnkd.in/efSSGj3g to-life

1

Wrapping up an amazing #PyConDE & #PyDataBerlin! Looks like our team really enjoyed having a discussion about #Streamlit with Lukas Masuch and Arnaud Miribel. Want to learn more? Missed out PyCon? No worries! We're hosting a meetup in Mainz next week to discuss key trends & takeaways. Stay tuned for details and our write-up. #Germany #DataScience #predict42, #NLP

25

1 Kommentar

Evolution of technology is evident and speed sometimes (read: often) astonishing. Another tangible reminder of the latter: OpenAI just released its new flagship model, GPT-4o (o for omni) and API. New model truly embraces multimodality - It welcomes text, vision and audio as inputs. Even with the previous GPT versions multimodality was enabled, but newly released capabilities make a hefty jump forward. Now you can talk with the model using a voice interface and latency in audio inputs will be on the level of human (as low as 232 milliseconds 😳). New model can also interpret pictures almost in real-time, so approaching human type of perception on this front as well 👀 Also announced: Much waited desktop application (first for Mac computers only) . 🔈Customer service use cases might again be embracing audio (instead of text chat) with these new capabilities, among many other possibilities! New model capabilities will be available to everyone, however paid users get more capacity. In addition, the desktop app will only be available for Plus subscribers (and in the beginning only for Mac users).

37

3 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

Today we report a significant advance in understanding the inner workings of AI models. We have identified how millions of concepts are represented inside Claude Sonnet, one of our deployed large language models. This is the first ever detailed look inside a modern, production-grade large language model. This interpretability discovery could, in future, help us make AI models safer. We mostly treat AI models as a black box: something goes in and a response comes out, and it's not clear why the model gave that particular response instead of another. This makes it hard to trust that these models are safe: if we don't know how they work, how do we know they won't give harmful, biased, untruthful, or otherwise dangerous responses? How can we trust that they’ll be safe and reliable?

5