RelationalAI

On last night’s Q4 earnings call, Snowflake CEO Sridhar Ramaswamy described the impressive results they’ve generated through their focus on delivering “the world’s best end-to-end data platform, powered by AI.” ?? As a Snowflake native application, it was especially exciting for us to hear the compelling description of the AI-enabled results RelationalAI client Blue Yonder is generating: "For example, supply chain leader Blue Yonder leverages Snowflake robust data management capabilities and scale to help companies transform their operations by offering AI powered insight." "The Blue Yonder platform processes over 20 billion AI predictions daily to help retailers, manufacturers and logistics providers better manage inventory, optimize delivery and respond to disruption." "It enables their business and their customers to access powerful supply chain intelligence that are deemed impossible to build on their own." ? Global supply chains are under increased pressure as changes in US tariffs have escalated complexity and unpredictability. ?? Reinforced by Nikolaos Vasiloglou, RelationalAI VP of Research ML during his exceptional 12 Days of NeurIPS talks, having modern AI systems capable of dealing with “large behavioral change” is critical:? "Sometimes your customers or the world has a large behavioral change, like a tariff or fraudsters who discover a new way to attack your system." "So now you have to have new features in order to deal with the new change in the world, or deal with this smarter fraudsters" (12 Days of NeurIPS). ? ?? As a result, legacy data and technology stacks are making it substantially more challenging for leaders to run simulations and what-if scenarios while those who have invested in developing intelligent applications are positioned to manage risk and capitalize on opportunities accordingly. ?? Challenges RelationalAI is purpose-built to solve as described by Max De Marzi, RelationalAI Developer Relations Engineer during a recent LinkedIn Live Session “AI, Tariffs + Supply Chains…Oh My.” "So we're going to offer a new solution, which is we're going to solve our problem by building a relational knowledge graph." "Instead, we're going to take our relational data model in tables, and we're going to create this relational knowledge graph that's going to be made up of entities and relationships that are going to be named so that you have some semblance of how things are connected, and pipe it all together into what most people think of as a digital twin." Links to the content above in the comments below.

Going to #MWC25? Come check out RelationalAI and Snowflake reimagining #digitaltwins via #knowledgegraphs

A fascinating trend that appeared across multiple 12 Days of NeurIPS sessions is the emergence of the Language Model Scientist. This specialized role is increasingly being adopted by enterprises at the leading edge of building and deploying AI. ? "I call it the LLM scientist toolkit…This is the part that made me believe that the Language Model Scientist is coming up as a new profession" (Day 8, 0:18), says Nikolaos Vasiloglou, RelationalAI VP of Research ML. All the dimensions for scaling are working independently and, in some cases, together to bring us closer to not only more useful models but also small enough to fit everywhere "AI will reach new levels of efficiency, approaching natural intelligence and become deployable on pervasive devices driven by core innovations in model architecture, software and hardware" (Day 7, 15:06). Understanding that model performance is dependent on the quality of data selection, the emerging role is creating value by improving resource optimization and driving performance enhancement through techniques such as model merging. "This is a way of dynamically merging language models, and I think that's the reason why it is, is actually very popular" (Day 9, 5:14). These skills are especially critical within businesses executing against their vision of replacing application-centric with data-centric architectures, where Language Model Scientists serve as key enablers of intelligent applications. "Smaller high quality data can offer a higher efficiency...They show with 40% less data, you can actually get the same performance" (Day 8, 3:09). Organizations that develop this specialized capability will gain significant advantages in both efficiency and effectiveness as they navigate the increasingly complex AI landscape. Links to the videos for Day 7, 8 and 9 of the 12 Days of NeurIPS content can be found in the comments below.

Every year, I go through NeurIPS or other major conferences and try to distill the content. This year, the conference was BIG. I spent over a month reviewing the content, mainly tutorials, keynotes, workshops, competitions, and expos. I want to share with you: - A slide deck https://lnkd.in/eEV8KGD3 - A flyover video presentation over the conference https://lnkd.in/eCiuH4gN - A series of 11 talks covering the 11, which in my opinion are the most important: https://lnkd.in/ewWSVMyT - A series of 12 podcasts based on the material https://lnkd.in/ecZ48CG9 created with NotebookLM The covered topics: - What is happening with Agents:?What are the breakthroughs of agentic AI, and what are the risks? - LLMs for Tables:?Getting more than from your enterprise tables. - Scaling Laws:?The laws that describe how AI grows. Learn how openAI made the right decisions for revolutionizing AI and how others adopt this approach to pave the way for breakthroughs - Data monetization:?Getting closer to attribution of predictions and text generation. We are one step away from tracing a generated token to the pre-trained/fine-tuned text that made it possible. We can now assign a price to data. - How much data can my model fit?:?Getting closer to theory and empirical rules that will help us decide how much and which data you need to build the LLM you want - The dimensions of LLM Scaling:?Hardware, Data, Model, and Time. How far are we from a ChatGPT on a chip? - The Rise of the Language Model Scientist:?After the Data Scientist, we have the rise of a new role in the enterprise. Learn about the latest tools and responsibilities that make this role a necessity. - Combining smaller LLMs to build bigger and better ones:?Bringing order into the zoo of LLMs. This revolutionary technology will accelerate team collaboration and bring robustness to building LLM-based applications. - Data preparation and evaluation in the era of LLMS:?Data was, is, and will be the new oil. Preparing your data for a predictive model is more important than the actual model. Learn how this applies to building/tuning LLMs - Teaching LLMs math and coding:?Teaching math and coding an LLM not only makes one a better mathematician/coder but also makes one better in other fields. Learn how much progress we are making and the subsequent barriers to be broken. - GNNs become mainstream:?While research is slowing down, open source packages are becoming more mature, and GNNS are ready to add value to your database. Who is your GNN vendor? I usually spend about 100 hours. This year it took more than 200 hours. I am thinking about crawling the content and actually (pre)training a language model to do the task. I am in the process of writing a proposal to the NeurIPS board. I seek companies like HPC-AI Tech and RunPod to contribute excess GPU hours. Comment below if you are interested.

Is it possible for an AI system to learn how to play soccer simply by reading a book and then orchestrate a team of specialized agents to use similar teamwork to design breakthrough drugs? In Day 2 of his 12-part series covering the most recent NeurIPS, RelationalAI VP of Research ML, Nikolaos Vasiloglou delivers another insight packed presentation including the answer to the above question… The answer is “yes.” At NeurIPS, we saw how combining imitation learning with reinforcement learning and macro-agent architectures is making this possible, with applications on how enterprise workflows of the future could be transformed. "They basically created a virtual lab of biomedical scientists where they decide how to create the agents and they have team meetings, individual meetings" (40:29). A hierarchical system of AI agents, each with specialized roles, collaborates using both structured and emergent protocols. "In the future there's going to be agents at different levels, the way that you have a company and you have the CEO, the board, the executives, the different departments all the way to technical work" (24:51). Instead of pure reinforcement learning through trial and error, researchers developed a hybrid approach. Language models first ingest domain knowledge, perform internal simulation and reasoning, and then use that foundation for practical execution. We're seeing the convergence of several key technologies: large language models, multi-agent systems, and hierarchical learning architectures. This isn't just incremental progress - it's a step change in enterprise AI capabilities. ? AI systems are evolving to prioritize operational efficiency and practical deployment over raw performance metrics, enabling faster and more cost-effective enterprise solutions (1:31) ? The emergence of new agent evaluation frameworks now allows us to measure AI progress similar to how we assess human performance on complex tasks (6:45) ? AI teams can now handle complete end-to-end processes, from creative design through practical implementation, mirroring human organizational structures (24:51) Link below to the “Day 2” video presentation - “The Return of LSTM & Evolution of AI Agents: Technical Deep Dive” - along with complete chapters and timestamps.

We’re proud to announce that RelationalAI's Hung Ngo has been named one of ACM, Association for Computing Machinery's 2024 Distinguished Members for his outstanding contributions to query evaluation and optimization algorithms. Congratulations Hung on this well-deserved recognition. We appreciate what you do and your innovative work continues to inspire us all! Link below for more information.

When Shopify implemented foundation models across their e-commerce platform, they didn't just add another tech layer – they reimagined their entire business operations. This wasn't an isolated case… At the most recent NeurIPS, we saw how quickly enterprise AI has transformed from buzzword to tangible business reality. Closing the gap between cutting-edge research and practical implementation faster than ever. We’re excited to share these Shopify insights along with many others from RelationalAI's VP of Research ML, Nikolaos Vasiloglou, who is delivering "12 Days of NeurIPS" - a comprehensive series breaking down the conference's most impactful developments via 12 insight packed presentations you don’t want to miss. In the first of 12 videos - “Day 1” - sets the stage for understanding how enterprises can build more effective AI roadmaps, compressing the time to value ratio accordingly. Nick highlighted a number of especially interesting developments helping make this possible, including: ?? Emergence of "Language Model Scientists," a new role bridging AI research, business applications and domain expertise ?? Acceleration of cost reductions through innovative AI model architectures and composition of small Language Models ?? Advanced agentic capabilities such as Chain of Thought, Self Reflection and Grounding, enabling AI systems to evaluate and improve their own outputs ?? Practical tools for implementing AI across business functions, including those that democratize access to those outside of engineering and data science teams For technical leaders:? The frameworks for building and deploying enterprise-grade AI are becoming more sophisticated yet more accessible. For business leaders:? These advances provide clearer pathways to AI implementation, with more reliable results and measurable ROI. Wherever you are in the crawl, walk, run or fly stage of enterprise AI adoption, watch this NeurIPS overview to understand how pioneering leaders are pulling the future forward. “Day 1” timestamp highlights: 14:24 - The rise of Language Model Scientists 23:25 - Shopify's practical implementation 34:18 - Self-reflection in AI systems 56:42 - Cost-effective AI breakthroughs A link to the video and a complete chapter / timestamp list can be found in the comments section below.

?? When tech gremlins strike… ??? Well, that didn’t go as planned. Today we went live with our fireside chat, "Navigating the New Era of US Tariffs with AI-Enabled Knowledge Graphs", but the tech gremlins had other plans. While we battled LinkedIn glitches, 50+ of you were left staring at a connecting screen instead of hearing from Max De Marzi and Alec Coughlin about how AI can simplify the chaos of tariffs, supply chains, and global trade. The good news is that we recorded the whole thing—minus the tech hiccups. ?? Catch the full recording here: https://lnkd.in/g6sFJVSw, and let us know your thoughts! Thanks for your patience and sense of humor. We promise the knowledge drops more than make up for the rough start. #Tariffs #AI #KnowledgeGraphs #SupplyChains

Navigating the New Era of US Tariffs with AI-Enabled Knowledge Graphs

How do we make AI more reliable—and where is the field headed next? Our VP of Research ML, Nikolaos Vasiloglou, joins Tony Hoang to break down the evolution of AI, from early neural networks to today’s advancements in RAG, symbolic AI, and agentic systems. You'll also learn practical solutions for tackling AI hallucinations and what recent models like DeepSeek’s R1 mean for the industry. Listen to the full conversation below #AI #MachineLearning #RAG #DeepSeek