Yannick Even

I continue to be amazed by large language models (LLMs) and use them multiple times a day. https://lnkd.in/gvP_DpyD is a comparison of GPT4.o, GPT4, Gemini Advanced, and Claude 3 answering a philosophical question. GPT4.o's answer is a clear step-up from its predecessor, especially in the current thinking section. How are you using LLMs in your work? #LLM #GPT #Gemini #Claude

33

1 Comment

We are standing at the edge of a paradigm shift. Data is no longer just a resource; in the AI future, it's the fuel. But here's the game-changer: organizations with the right data strategy won't just power AI systems—they'll design the future. So, how do we prepare? It’s not just about bigger datasets or faster pipelines. It’s about future-proofing data by focusing on three critical pillars: **Data Ethics at Scale: In an AI-driven future, trust will be the currency of success. Are your data strategies built on fairness, transparency, and accountability? AI amplifies biases; your data must neutralize them. **AI-First Data Governance: Traditional governance won’t cut it anymore. Data flows in real time, across borders, from humans and machines alike. An AI-first mindset demands real-time compliance, adaptive security, and seamless interoperability. **Predictive Ecosystems: It's not about data dashboards but about actionable ecosystems. Tomorrow’s leaders will be those who anticipate opportunities, not react to them. Your data strategy should position your organization to make decisions before the moment arrives. The future of AI isn’t just about tech—it’s about trust, actionability, and foresight. What are you doing today to ensure your data strategy can shape tomorrow’s AI outcomes? Let's share ideas—how are you preparing your data strategy for the AI future? 🚀 #AI #DataStrategy #AIStrategy

111

5 Comments

The AI Index Report 2024 helps exploring the enigma of AI Model Training Costs: definitely one of the best part of the report and please find attached one of the key charts. Despite the veil of secrecy surrounding expenses, it's widely understood that these investments amount to millions of dollars and are on an upward trajectory. Notably, OpenAI's CEO, Sam Altman, disclosed that training GPT-4 exceeded $100 million. This sharp escalation in training expenses has posed a formidable barrier to universities, historically pivotal in AI research. While the importance of understanding these costs cannot be overstated, comprehensive data remains elusive. The AI Index stands out as a pioneer in offering estimates on foundation training costs. These estimates affirm suspicions that recent years have witnessed a substantial surge in model training expenses. For instance, from the original Transformer's $900 training cost in 2017 to GPT-4's estimated $78 million in 2023, the disparity is staggering. This underscores the urgency of informed discussions and strategic planning in navigating the evolving landscape of AI development. Link to full report below: https://lnkd.in/drAhKtvT #AI #ModelTraining #FoundationModels #TechTrends #LinkedInDiscussion

25

How to Vet Information Before Making a Decision by Alex Edmans in Harvard Business Review . "The daily decisions modern leaders face are increasingly complex. But executives have a tool to combat these challenges – information. At the click of a mouse or the press of a thumb, they can call up cutting edge research on virtually any topic. With so much information available, how do we know what to trust? What executives need is a simple taxonomy of misinformation so they know what to look out for. Drawing on the tools of social science research we can categorize misinformation into four missteps. This framework can be useful to leaders of all kinds who need to ask better questions to manage their own information onslaughts. - The first misstep is that a statement is not fact, because it may not be accurate. How often do we take a statement at face value because we like what it says? - The second problem is that a fact is not data because it may not be representative. Case studies are a popular and powerful teaching method. They’re vivid, they bring a topic to life, they inspire leaders much more than any statistics or regressions can. And this applies beyond business school education to best-selling books and viral talks, which typically start with a story to hook the audience. But gurus typically pick the one example that illustrates the point in the starkest possible way, as it makes the story most memorable. A single anecdote — or even a couple of anecdotes — tell us little, because they might be cherry-picked. They’re the exceptions that don’t prove the rule. - The third warning is that data is not evidence, because it may not be conclusive. Large-scale data may be representative, but it is only a first step because data is merely a set of facts. Evidence is a set of facts that allows us to draw a conclusion. Doing so requires it not only to support your theory, but also to reject rival theories — just as evidence in a criminal trial is only meaningful if it pinpoints your particular suspect. - The final caveat is that evidence is not proof, because it may not be universal. A proof is absolute. When Archimedes showed that the area of a circle is pi times the square of its radius, he proved this not just for circles in Ancient Greece in the third century BC, but also for circles throughout the world today. But even if evidence nails a result perfectly, it may only do so in the setting in which it was gathered — it may not do so elsewhere." https://lnkd.in/ggquda5H

33

2 Comments

If you haven't yet, make sure to read the latest article from McK on optimizing your generative AI strategy & implementation roadmap moving away from the Honeymoon & Pilot phase. Here are some key takeaways:   Eliminate the Noise: Focus on what’s truly effective. Be honest about successful pilots and reduce unnecessary experiments. Channel your efforts into solving critical business problems.   Holistic Approach: It’s about integration, not just individual components. Prioritize how the parts of a gen AI engine work together securely, rather than assessing them in isolation.   Cost Management: Models are only about 15% of the overall cost of gen AI applications. Identify hidden costs and apply the right tools to manage them efficiently.   Tool and Tech Proliferation: The abundance of infrastructures and tools can hinder scaled rollouts. Focus on the capabilities that best serve your business, leveraging cloud services while maintaining flexibility.   Build Value-Driven Teams: Success at scale requires teams with diverse skills to not only build models but also ensure they deliver value safely and securely.   Prioritize Data: Seek out the most impactful data and invest in its management for quicker scaling, rather than chasing perfect data.   Reuse Code: Reusing code can boost development speed for generative AI use cases by 30 to 50 percent.   Dive into the full article to get more insights!   #AI #GenerativeAI #BusinessStrategy #CostManagement #DataManagement #TeamBuilding

21

🔐 Understanding Encryption Key Hierarchy in Snowflake 🔑 Ever wondered how Key Hierarchy works to secure your data in Snowflake? 🌐 Snowflake uses a powerful multi-layer encryption architecture to ensure your data remains secure at every level. To learn how the Key Hierarchy operates and gain a clear understanding, check out this video 📽️: 👉 https://lnkd.in/gnsCM5Uc This video simplifies the concept and explains the hierarchy in detail. Don't miss it! 🚀 For more such content Follow Enqurious (formerly Mentorskool) Enqurious Academy Subscribe to https://lnkd.in/g9M_ZZDV #Snowflake #DataSecurity #Encryption #CloudData #KeyHierarchy

5

I really struggle with pithy data storytelling titles and would love advice on how to approach it with this chart. It's a silly, dirty and quick #30DayMapChallenge inspired by Ansgar Wolsing https://lnkd.in/eeMfiPHf I've filtered maps::world.cities for cities with a population of at least 1 million and drawn a circle from each city to its closest partner - that's what the purple area represents https://lnkd.in/e3V8cyqw

10

2 Comments

Smart people know that #decisionintelligence is way more than "AI for decisions" : it's an integration discipline that connects multiple kinds of knowledge - including human knowledge, randomized controlled trials, econometrics, behavioral economics, ML, AI, and much more - into the pathway from actions to outcomes. Listen to The Decision Intelligence Handbook by L. Y. Pratt, N. E. Malcolm on Audible: https://lnkd.in/grpask9v

22

11 Comments

I've seen some pretty cool demos of using ChatGPT 4o (paid) for exploratory data analysis. Here's how my test went: --- Me: What formats can I provide you data in for exploratory data analysis? ChatGPT: You can provide data in various formats for analysis. Some of the most common formats include CSV, TXT, Excel, or JSON. Me: Can you read .RData files? Chat GPT: Yes, I can read .RData files. Please upload the .RData file, and let me know what kind of analysis you would like me to perform. Me: my_file.RData. Please perform a standard exploratory data analysis. ChatGPT: It seems I don't have the pyreadr module available in my current environment. I'll use an alternative approach to load and explore the data in the .RData file. ChatGPT: Let's use the rpy2 package, which allows Python to interact with R. I'll proceed with this approach to load and inspect the data. ​​ ChatGPT: It appears that the rpy2 package is also not available in the current environment. Me: Can you read parquet files? ChatGPT: Yes, I can read Parquet files. Please go ahead and upload the file, and let me know what kind of analysis you need. Me: my_file.parquet ChatGPT: It appears that I don't have the necessary libraries to read Parquet files directly. However, you can convert the Parquet file to a CSV format, which I can then analyze. Me: my_file.csv ChatGPT: <Analyzing> ChatGPT: It seems there was an issue with the execution environment. Let's try reloading the CSV file and conducting the exploratory data analysis again. ​ Me: my_file.csv ChatGPT: It looks like I'm currently unable to process the CSV file due to technical issues with the execution environment. Please retry again. Me: my_file.csv (smaller this time) ChatGPT: It appears that there is a persistent issue with the current environment preventing the loading and processing of the CSV file. As an alternative, I can guide you through conducting an exploratory data analysis (EDA) on your dataset using Python on your local machine. Here is a step-by-step guide: Me: 😢 😂 ---- Will try again later!

27

5 Comments

In most cases, you have to get information from an 8.5x11 sheet of paper into the brain of a decision maker. A major barrier to that transfer is the vison of the decision maker. So understanding how the eyes function is the first step to a successful brief. #governmentwork #militarylife #datavisualization

9

2 Comments

The value of vector indexes is rising for Generative AI ... why and what are vector indexes? As Gen AI gains traction, Retrieval-Augmented Generation (RAG) is emerging as a key approach for adding value. RAG enhances language models with external knowledge retrieval for improved accuracy and domain-focused applications. At the core of RAG lies its most critical component, efficient knowledge retrieval, which happens via vector indexes. In other words, how fast can the solution find the meaningful relevant information from the additional data provided? With vector databases, there are a number of options for indexing, including: 🌲 Tree-based indexing 📊 Graph-based indexing (e.g., HNSW) #️⃣ Hashing-based indexing (e.g., LSH) 🗜 Quantization-based indexing Each has its own set of benefits and challenges, which is why selecting the right vector index is crucial for RAG solutions. When selecting options, it is a challenge of balancing factors like dataset size, dimensionality, and accuracy-performance trade-offs. As Generative AI evolves, vector indexes will play a pivotal role in enabling accurate, knowledge-grounded AI applications. Read the full post on Vector indexing here: https://lnkd.in/egDjCxYy What has been your experience working with a RAG solution? #GenAI #AI #RAG #GenerativeAI #VectorIndexes #VectorDatabase #Data

12

Combining the strengths of RAG Systems with Ontology-Based Approaches can create a powerful hybrid system for question-answering. 🔍 RAG (Retrieval-Augmented Generation): - Utilizes broad retrieval capabilities. - Effectively handles unstructured data. 🔗 Ontologies: - Ensure precision and relevance. - Provide explainable and reliable information by leveraging structured knowledge. In this hybrid model, initial information retrieval is performed using vector-based methods. The final answers are then refined and validated against the structured knowledge provided by ontologies. This approach leverages the expansive data handling of vector systems while ensuring the precision and transparency of ontology-based solutions.   Interested in how this works? 🧬 Check the comments to learn more about how SciBite Chat integrates ontology-backed semantics with a RAG architecture to support researchers in life science-related industries. Feel free to share your feedback or thoughts on these insights!

14

1 Comment

AI is reshaping third-party risk management. But here's what most are missing: It's not about replacing humans. It's about augmenting our capabilities. Key insights from my recent research: 1. AI excels at pattern recognition across vast datasets 2. Human judgment is crucial for context and nuanced decision-making 3. The sweet spot? AI-powered insights + human expertise Example: Our AI flagged a vendor's unusual network activity. Human analysis revealed a previously unknown vulnerability. The future of TPRM is collaborative intelligence. Are you leveraging both AI and human expertise in your risk strategy? #ArtificialIntelligence #RiskIntelligence

14

2 Comments

Lets be clear - LLMs are new software infrastructure. The path to creating value remains still as hard and will take as long as usual. You still need to get these things right and AI won't accelerate it much, - finding right use cases, - building great customer experience, - getting the pricing strategy right, - executing and iterating on GTM, - creating the right org with the right kind of people, and - building a smooth delivery experience Its still a 5-15yrs of hard work.. no amount of VC money will help accelerate it.

44

4 Comments

"The Big Book of GenAI", interesting and useful e-book published in May 2024 by Databricks. 👇 This report compiles best practices for building production-quality #GenAI applications. It provides technical guidance and code samples for deploying applications, fine-tuning and training #foundationmodels, evaluating and monitoring #LLMs for integrity, and exploring various #generativeAI architectural patterns, offering a comprehensive resource for #AI development. Link: https://lnkd.in/dF23qk67 #ailaw #machinelearning #deeplearning #artificialera #web3 #web4 #web5 #fintech #digitalera #digitaltransformation #generativeAI #trustworthyAI #explainableAI #aiethics #artificialintelligence #explainableAI #responsibleAI #foundationmodels #GPAI #GAI #LLMs #NLPs #GenAI #frontierAI #narrowAI #AGI #ethicalAI #adversarialAI

20

At Integrated Special Purpose AI (iSPAI) we are building a decentralized knowledge network and marketplace for Agentic AI. The topic below is deeply relevant to the future of our business and of AI more broadly. If you’re interested in top decentralized knowledge graph companies out there, check out OriginTrail. They’re doing some pretty cool stuff. And if you’re interested in the details below or quantum computing (another light topic…) Larry Bowden Jr, Dr. h.c., P.E. is as deep as you’ll ever need to go.

6

1 Comment

McKinsey released this article on GenAI - https://lnkd.in/gXxvmgTn 65% of companies now use #generativeAI regularly, but many have not seen significant impact yet. Companies need well-defined #operatingmodels to scale #genAI effectively, focusing on #datamanagement, #governance, and #compliance. Implementing #genAI in components allows for flexibility and alignment with business goals, though only a few companies have adopted this model. Q: Where are you in your #journey to #genAI adoption and what #benefits have your realized? Quinte Financial Technologies #dataanalytics #banks #creditunions #financialinstitutions

8

Calling out 2 key points : 1. Data is fundamentally additive 2. It's the marginal lift that matters The value of any given data element is not just in it's application/use case but is also a function of the orthogonal lift over the status quo data used in a given workflow. Great essay Abraham Thomas https://lnkd.in/gWxRBr8h

2