Alessandro Casuccio

In an exciting wave of developments, SAS, renowned for its prowess in data and AI solutions, has unveiled what they're calling a game-changing approach to tackling business challenges. The introduction of lightweight, industry-specific AI models for individual licensing sets a new-era beacon. It embodies SAS's ambition to democratize AI - making it accessible regardless of one's skill set. Some thoughts on why such a move is transformative: • It disrupts the traditional approach by placing ready-to-deploy AI technology directly into the hands of organisations. This could mark the dawn of an era where AI integration into businesses becomes the new normal rather than a distinguished edge. • The idea of industry-specific AI models undermines 'one-size-fits-all' and acknowledges the unique needs of each business sector. It can potentially ensue a wave of diversified innovations in different industry niches. Consider the vast spectrum from healthcare to entertainment to finance - each one will have a model suited to their unique challenges and needs. • It is not just about easy access, but also about efficient productionisation of real-world use cases. Given SAS's expertise in AI, these models could well fill the demand-supply gap in AI technological requirements, setting a new track record in driving productivity with speed and efficiency. As Chandana Gopal, Research Director, Future points out, this could be a big boost towards making AI accessible to businesses of all scales. In light of this, my prediction is two-tier: 1. We would witness an upsurge in the adoption of AI technology across businesses – be it multinational corporations, local enterprises or startups. 2. Given the industry-specific approach, I anticipate AI-related breakthroughs nurturing sector-specific innovations leading to enhanced customer experiences. #AI #Technology #Innovation #businessstrategy #SAS #ArtificialIntelligence

٥

١ تعليق واحد

Episode 1 is live now 🚨 Leading a data transformation ? 🎯 Define a clear vision and an execution plan 🚥 Define the meaning of a data driven organisation 🪡 Elevate your organisation with tailored educative programs around Analytics and AI #breakthemyths And many more insights in the first episode of L'Oréal Tech Express below 👇

٦٠

٧ تعليق

The core data platform is properly the most consequential tech decision a D&A department has to make Your ability to innovate is indirectly tied to your vendor's development pace, and the platforms become hugely consequential in your ability to work in the larger integration landscape of a company, implement solid governance, or define ownership and discoverability through a data catalog. Therefore, I'm also looking forward to participating in a panel discussion about this topic with regard to Microsoft Fabric hosted by Inspari A/S. I like a lot of things about Microsoft Fabric. The rework of the data model engine in Power BI has huge potential for expanding the use cases for real-time reporting and removing unnecessary data duplication across a cacophony of confusing data warehouse layers. I think the commitment to open-source storage formats is incredibly helpful in ensuring a smoother exchange of data across platforms. However, I think Microsoft should reconsider some of the other choices that they have made on the platform. If I'm to focus on two main things, it is the reintroduction of coupling between storage and compute. I thought that the "Spark revolution" over the last couple of years had made that a thing of the past, but I can't see why you would move to a more expensive OneLake setup than the regular ADLS. The second thing they should reconsider is the capacity setup. As someone coming from the Power BI world, I've always been frustrated by the lack of flexibility in the capacity pricing model. Transferring this pricing model to the backend of the platform could give our FinOps colleagues some unpleasant surprises... Over the last couple of years, we have made three significant platform and architectural bets. I think two of them have been extremely successful, and one of them has failed. If you want to hear which is which, you have to attend the debate 😉 Link below 👇

١٢٨

٤ تعليق

🧠 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

٥

١ تعليق واحد

When it comes to generative AI, the more reinforcement learning with human feedback is applied during model training, the higher quality the model’s responses become.   Read Gartner's executive’s guide to understanding, planning, piloting, implementing and maintaining GenAI. #GenAI #AI #evanta #cdaos #cios

٢٥

Exciting news! OpenAI o1-preview has been released, offering a unique approach to AI interaction. Unlike traditional models, o1 is an agentic reasoner, designed to provide insightful responses with minimal input. What sets o1 apart? This innovative model thrives on concise direction, leveraging advanced reasoning to deliver comprehensive answers without extensive prompts. To effectively prompt o1: 🔹 Provide less detail and more direction 🔹 Trust o1 to infer context 🔹 Use clear commands with action verbs like "Explain," "Compare," or "Summarize" By following this approach, o1 can utilize its full reasoning potential, saving time and producing more nuanced, insightful responses. How to Effectively Prompt o1: Less Detail, More Direction: Skip lengthy explanations. Instead of: “Can you provide a detailed analysis of the impacts of renewable energy on global economies, including benefits and drawbacks?” Try: “Analyze the global economic impacts of renewable energy.” Minimal Context Needed: Trust o1 to infer context. Instead of: “Discuss how recent AI advancements are influencing the job market across sectors.” Try: “Discuss AI's impact on the job market.” Use Clear Commands: Start prompts with action verbs like “Explain,” “Compare,” or “Summarize.” Example: “Compare remote work vs in-office setups on team productivity.” Use this new model where depth is important and where context can be inferred, use something like 4o-Mini for tasks, it's faster and more easily controlled. If you do have the opportunity to test it, don't forget to expand the reasoning widget next to the OpenAI logo within the chat experience to view how the model is breaking out the instructions and tasks in response to your questions, the transparency is very much appreciated! #OpenAI #ArtificialIntelligence #Innovation

١٦

٣ تعليق

A few years ago, there weren't many bloggers writing about analytics, unlike today. Substack now lists over 360 newsletters about analytics. However, only a few writers truly stand out, one of whom is Benn Stancil. I am so excited to share my interview with Benn, my favorite writer and analyst. In our conversation, we discuss: - The tricky nature of BI. - The pitfalls of building data analytics tools. - The importance of version control for analysts. - Favorite features of analytical apps. - New trends in analytics. And more - https://lnkd.in/gMbyu_bR

٤٢

٣ تعليق

In the realm of big data analytics, the use of Large Language Models (LLMs) provides a unique opportunity to create and develop additional insights. #DataScience #AI #ArtificialIntelligence https://hubs.li/Q02tt2SC0

٣

🔥 🚨 Breaking news 🚨 🔥 Meta releases Llama 3.1 absolutely destroying the benchmarks. More info here: https://llama.meta.com/ 𝗧𝗵𝗲 𝗔𝗵𝗮! ✅ Includes Tool Use (Upload a dataset and analyze it. Prompt model to plot graphs and fetch market data) ✅ Multi-Lingual Agents (eg. translate photos into video in Spanish) ✅ Complex reasoning (understand and solve logic problems) ✅ Long Context (Upload, analyze, and summarize large documents with up to 128k tokens) ✅ Coding Assistants (for complex end2end programming workflow) Summary of Llama 3.1 405B benchmarks: 🏆 88.6% on MMLU (vs. 88.7% GPT4o, 88.3% Claude 3.5 Sonnet) 🏆 88.6% on IFEval (vs. 85.6% GPT4o, 88.0% Claude 3.5 Sonnet) 🏆 96.8% on GSM8k Math (vs. 96.1% GPT4o, 96.4% Claude 3.5 Sonnet) Rah-rah.. another slightly better model.. why should you care? 🎉 This is an inflection point where for the first time an open weight model has come this close, and in some areas beat the leading closed source frontier models. This is AI's Linux moment where Open is the new standard. 🎉 Llama3.1 license allows for any developer to distill a smaller model optimized for their use cases, use it legally for generating synthetic data, and as a teacher model to improve smaller models. OpenAI, Anthropic etc. don't allow this. 🎉 A 405 billion parameter model being able to outcompete models 3x it's size makes the performace/cost ratio deliver great business ROI 𝗧𝗵𝗲 𝗨𝗵-𝗢𝗵! 😔 Meta models are open-weight not open-source like IBM granite or Mistral. 😔 Meta does not provide any indemnification for their models. Which means a third party sues the enterprise for copyright infringement related to their use of the Llama Materials, the client is responsible for defending themselves and holding Meta harmless. If the enterprise sues Meta for copyright infringement, the license agreement will terminate. IBM's granite models have set a very high bar for indemnification that needs to be the new standard. https://lnkd.in/gVgYT9Yt 😔 Since Meta released Llama under their custom license, it has a lot of constraints, and doesn't come with patent protection like Apache 2.0 (eg. IBM granite, Mistral). 😔 Meta enforces you to include “Built with Llama” on any app that leverages Llama eg. on chatbot UI, app UI, website, blogpost etc. #meta #watsonx #ibmconsulting #ai

١٥٢

٤ تعليق

Round up time! Four recent articles about data quality and my spicy take 🌶️ Marc Di Gaspero and Adrien Vermeirsch from Potloc have done extensive research on data quality, and now Marc is sharing some key insights! Out of over 26,000 entrants, the average was 22 daily survey attempts (!), with similar results across five major online panels and five sample exchanges/DIY solutions.  The best part: High-frequency survey takers is an easy problem to solve - why don’t we? Because less traffic = less revenue. 🤷 In the article “$urprise! Re$pondent$ are not altruistic,” Vignesh Krishnan at Rep Data discusses the significant drop in incentives over the past 10 years (his rough estimate is 71% 😮). While increasing incentives could potentially attract more fraud, he argues that as more good respondents participate, the proportion of bad respondents would decrease.  Fair enough—so why don’t we increase incentives? Because it will increase the CPI and not the profit margin. 🤔 Chris Atkins at Yonder Data Solutions writes about how “the research industry is focusing too much on managing the symptoms of poor data quality rather than addressing its root causes.” I agree 100%. Our response to the data quality crisis is to build bigger armies. We need fraud detection software and QC checks, but that’s not fixing the root cause of the problem.  😐 Terence R. McCarron at OpinionRoute writes about how ‘conversion rate’ is what dictates the CPI, not the old ‘Incidence Rate.’ This matters because quality removals lower conversion rates, which is generally bad for business if you’re a sample provider. He says, “quality is not a priority that’s assured through the economic model.” Amen to that. 🙌 Links in the comments 👇 Your thoughts? 💬 #rmx #dataquality #surveys 

٩٨

٢٢ تعليق

I like when system architectures make sense as human processes that scale. Fancy distributed protocols like Paxos are mathy and fun. But I am happier when a human model of things works at inhuman scale. OCI Generative AI Agents, launched yesterday, are one of those architectures. Jia Chen and Egor Pushkin produced a great Behind the Scenes blog to introduce them. They use retrieval-augmented generation (#RAG) to produce truth-telling, up to date agents. What clicked for me, working with Jia and Egor, was thinking of these #GenerativeAI Agents as librarians. My mom was a librarian, starting as a medical reference librarian. My mom was brilliant and a voracious reader. Without medical training, her customers did not come to her for her medical insight. They came because she knew how to find relevant answers and write well. So too with our RAG-based agents. You grant them access to your library (folders of documents, #OCI #OpenSearch instance, whatever). Then, ask them questions. The language AI stuff decodes your question into a query. The agent roams the reference material, reading for answers. And generates an answer summary from appropriate reference material (with citations). Like a good reference librarian. Or a great customer support agent (the example Jia and Egor use for improved customer support bots). For as many agent/librarians as you need, at any time, on demand. It's neat and understandable. Implementation was non-trivial, but once you get them, employing them is pretty straightforward.

٥٠

٥ تعليق

AI is not part of the data stack and belongs in the consumption layer. AI is a data consumer with its own set of clients to delight. It's a customer of the data stack. Our job as data developers is to ensure AI is served the right nutrients to compute and calculate all it wants without spitting out the wrong things that have held several organisations accountable so far (legal, reputation, cost). The more AI becomes relevant for industries and replaces iterative jobs, the more relevant it makes the roles of data engineers, platform developers, governance stewards, and essentially all data enablers. 𝐁𝐮𝐭 𝐇𝐨𝐰 𝐭𝐨 𝐂𝐨𝐨𝐤 𝐭𝐡𝐞 𝐃𝐚𝐭𝐚 𝐑𝐢𝐠𝐡𝐭? 👨🏽‍🍳 ----------------------------------------- Feed large datasets and watch AI do its magic? On the menu, it's like ordering the "𝘙𝘦𝘤𝘪𝘱𝘦 𝘧𝘰𝘳 𝘋𝘪𝘴𝘢𝘴𝘵𝘦𝘳 ⚠️": Tastes delicious at first, and then burns your pockets with hidden charges you never imagined 💸 AI cannot simply ingest large quantities of data carbs, it needs its protein and vitamins from 𝐂𝐨𝐧𝐭𝐞𝐱𝐭💊. Data on data that helps AI "understand" data. So let's leaf through to the "Context-inental" section on the menu. This is where you find authentic context sourced from the depths of the semantic seas. 𝐁𝐮𝐭 𝐡𝐨𝐰 𝐭𝐨 𝐩𝐥𝐚𝐭𝐞 𝐭𝐡𝐢𝐬 𝐂𝐨𝐧𝐭𝐞𝐱𝐭? While you can record and serve semantics in a hundred different ways, there's nothing better than a Graph that combines semantics from, say, underlying data products, independent (bronze) databases, and even downstream glossaries populated by real businesses. It replicates how humans understand data by connecting dots, like neuron links. The article "GraphRAG Manifesto" by Neo4j is about a new approach to AI that combines retrieval-augmented generation (RAG) with graph technology. 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐑𝐀𝐆? A technique which improves AI performances by integrating existing knowledge. RAG enables AI to retrieve relevant info from extended resources beyond the training data to generate more accurate and context-aware responses. 𝐖𝐡𝐚𝐭'𝐬 𝐆𝐫𝐚𝐩𝐡𝐑𝐀𝐆? 𝘛𝘩𝘪𝘴 𝘸𝘰𝘳𝘬𝘴 𝘫𝘶𝘴𝘵 𝘭𝘪𝘬𝘦 𝘢𝘯𝘺 𝘰𝘵𝘩𝘦𝘳 𝘵𝘺𝘱𝘦 𝘰𝘧 𝘙𝘈𝘎, 𝘦𝘹𝘤𝘦𝘱𝘵 𝘸𝘪𝘵𝘩 𝘢 𝘤𝘢𝘭𝘭 𝘵𝘰 𝘢 𝘬𝘯𝘰𝘸𝘭𝘦𝘥𝘨𝘦 𝘨𝘳𝘢𝘱𝘩 𝘪𝘯 𝘢𝘥𝘥𝘪𝘵𝘪𝘰𝘯 𝘵𝘰 𝘢 𝘷𝘦𝘤𝘵𝘰𝘳 𝘪𝘯𝘥𝘦𝘹. 𝘖𝘳 𝘪𝘯 𝘰𝘵𝘩𝘦𝘳 𝘸𝘰𝘳𝘥𝘴, 𝘎𝘳𝘢𝘱𝘩𝘙𝘈𝘎! ~Philip Rathle Here's an incredibly well-written article I'd highly recommend by Philip Rathle that dives into the depths of an optimised approach to adding context for AI, optimising AI costs, and improving "understanding" by bridging the dots that RAG falls short in. https://lnkd.in/dNCEemGC #AIStrategy #GenAI #datastrategy

٤٩

٢ تعليق

It was exciting to see Kjell Carlsson, Ph.D. from Domino Data Lab rock the stage of the #AdoptAI event in Paris hosted by Artefact Kjell highlighted the crucial components of a successful Responsible AI (RAI) strategy.: 1️⃣ People: Empower and equip leaders and practitioners. 2️⃣ Processes: Develop and manage models effectively to mitigate risks. 3️⃣ Technology: Use governed platforms to automate and orchestrate RAI activities. This holistic approach ensures accountability, efficiency, and scalability in AI initiatives. 🌟 #ResponsibleAI #AI #GenerativeAI #Governance #Technology #Leadership #AIatscale #EnterpriseAI

١٢

How fast is fast when it comes to commercial open source software traction? Here are some stats to serve as benchmarks that Matthieu Lavergne put together among many other great data and insights in Serena's first report on commercial open source software. It's available for free here: https://lnkd.in/d5uBvjAU

١٨

٣ تعليق

🚀 The economics of GPUs: How to train your AI model without going broke! Latest article by Ksenia Se helps us to dive into strategies for companies, especially those with tighter budgets, to train AI models efficiently without the staggering costs associated with GPUs. From mixed precision training to activation checkpointing and multi-GPU setups - on how to optimize hardware and software for cost-effective AI development!💡

٥

١ تعليق واحد

#ObsessivelyEnthusiasticInvestor #UnreasonablyEnthusiasticInvestor #IrrationallyEnthusiasticInvestor at REFASHIOND Ventures: The Industrial Transformation Fund. We look for the people we believe are doing the best work in Industrial Transformation & Supply Chain Technology; Attempting the impossible. Then we go all-in on them. With obsessive and unreasonable enthusiasm. Join us. #IndustrialTransformation #SupplyChain #Innovation #Technology #Startups #VentureCapital #DemandTheImpossible #UnreasonableEnthusiasm #ObsessiveEnthusiasm #IrrationalEnthusiasm

٢