AI’s trillion-dollar opportunity: Context graphs The last generation of enterprise software created a trillion-dollar ecosystem by becoming systems of record. Salesforce for…

Foundation Capital just published "Context Graphs: AI's Trillion-Dollar Opportunity" and it's the most technically coherent thesis I've seen on where enterprise AI infrastructure is heading.

I’ve been exploring ways to reason more clearly about meaning, responsibility, and criticality in data-heavy organizations — and I’d really value input from others who work with conceptual models or architectures. One line of thinking I’ve been exploring is whether it helps to separate three concerns that often get mixed: • A very small upper ontology that defines basic kinds of things (e.g. object, event, measure, context) • A Business domain ontology that describes operational reality — what exists, happens, and can be validated in the business • A Systemic domain ontology that describes how that same reality is interpreted for finance, risk, reporting, or regulatory purposes The intent is not to add abstraction for its own sake, but to avoid: • business meaning being overwritten by reporting logic • regulatory interpretations being treated as operational facts • debates about “one correct definition” Instead, business and systemic concepts would be explicitly related, but not collapsed into one. I’m curious how others see this: • Does separating business reality and systemic interpretation resonate with your experience? • Is an explicit upper ontology helpful — or overkill — in this kind of setup? Genuinely interested in different viewpoints, especially from those who’ve tried (or rejected) similar approaches. #EnterpriseArchitecture #DataArchitecture #InformationArchitecture #DataGovernance #BCBS239 | 23 comments on LinkedIn

The Procedural Knowledge Ontology, From Theory to Practice

Here is a really nice NotebookLM deck that explains how my Open Source HTML Graph service works

From AI to Knowledge Graphs, more and more companies are discovering the power and usefulness of semantic data to open up new business capabilities and support current programs. Semantic data models are different from other types of data models in that they incorporate both the meanings of the terms

📌 Why Knowledge Graphs Are No Longer Optional for Enterprise AI

Descriptive set theorists study the niche mathematics of infinity. Now, they’ve shown that their problems can be rewritten in the concrete language of algorithms.

In this episode of Founders Discussion, TuringDB founders Adam Amara and Rémy Boutonnet sit down to discuss one of the most important and often overlooked ca...

Stop overloading your contexts!

Discover what’s new this week in graph technology, including a big acquisition, a pre-seed investment, a 360-degree take on Connected Data London, and more.

A Gremlin-based implementation of the LDBC Social Network Benchmark (SNB) Interactive v1 workload for TinkerPop-compatible graph databases.

Learn all about G.V()’s recent $658,000 pre-seed investment round from Techstart Ventures and what’s ahead for the product, company, and graph community.

TinkerBench is a benchmarking tool designed for graph databases based on Apache TinkerPop. It provides an efficient way to measure Gremlin query language performance in an easy and flexible manner. TinkerBench is created and maintained by Aerospike

proposed model GRAPHTRIALS identifies key processes for visually proving an assertion about a given graph in an adversarial setting. The prosecution lawyer, a software or a human (assisted by software), intends to highlight evidence for a graph being accused of satisfying an assertion by usage of a visual certificate drawing. To convince the judge, the human audience of the drawing, the visual certificate must guide the judge’s perception to form a mental model which makes the assertion easy to validate. Further, the visual certificate must be unimpeachable as a defense lawyer, yet another piece of software or a human adversary, and checks for reasons to doubt the validity of the certificate which may influence the judge’s verdict

But what if a broken link could become a moment of discovery?.

An engineer for New York Times Games has been trying to teach artificial intelligence to understand wordplay more like a human.

Every few years, the field rediscovers ontology

What the RDF is a Knowledge Graph? I generated an interactive taxonomy of 120 semantic modeling concepts, including: • Clear definitions • Detailed explanations • Explicit relationships between concepts • One navigable overview instead of scattered articles The result is something you can actually explore, not just read. You can also export the entire taxonomy to: = Excel = Static clickable HTML = Clickable PDF = OWL (work in progress though) I’m curious about your feedback: • Does anyone know a similar easy to read open-source overview like this? I could not find one. • Are there important concepts missing? • Do you disagree with certain classifications, definitions, explanations? == Link and details in the comments #KnowledgeGraph #SemanticWeb #Ontology #RDF #SemanticModeling | 41 comments on LinkedIn

SEOs have been mildly obsessed with knowledge graphs for a while, but mostly because we want to ensure that our websites/brands/clients are well represented in Google’s Knowledge Graph, which makes sense. What not many of us have had experience with (myself included) is building our own knowledge graphs. It is technically challenging and not immediately apparent what the benefit is, and some people aren’t aware that it isn't exclusive to Google. Pioneers like Andrea Volpini and Dixon Jones have long been advocates for approaching this in a more nuanced way. Right now, I’m just learning in public, so stick with me. So what have I learned? - Structured data is IMPORTANT. Not just schema, but having tried to reliably infer semantic triples across a 1000+ document set using an LLM, it’s clear that easily understable/scrapable data is KEY - As a second point, I might not recommend doing this with an LLM at all… depends on your objective (more below). But using a merchant feed (for example) OR schema and graphing that would ultimately be easier. - It’s too easy to forget to provide explicit context for key information. For example, when you strip a page down to JUST its readable content, does it still make sense? Maybe the key question here is, “Why build yourself a knowledge graph, though?” It’s an excellent opportunity to learn, but also quite a painful experience - especially when each full processing run has taken about 18 hours, so you need a good reason. Here are my goals: - Searchable knowledge store/database, particularly looking for “facts”, great for content writing - Trying to document the knowledge an LLM can easily infer by way of a basic simulation of how other models may understand content. Not in a reverse-engineering sense, but to really highlight what information IS NOT clear/specific - A grounded knowledge store for LLM content generation - rather than a model making up key details, we can use this KG as a store of already signed-off details. Creating a basic API for Cursor/Claude Code to use as a Tool is where I am starting This is not an “are you even an SEO if you’re not building knowledge graphs in 2026” post - that would be madness. This process is challenging and clearly not worthwhile for many, many people. But if you have large-scale content challenges and want a really in-depth way to understand them, I can totally see the value. If this isn’t you, but you want to benefit from some elements of this, I’d probably ask Dixon about Waiky 😉 | 15 comments on LinkedIn

RAG Was Supposed to Fix Hallucinations. Instead, It Added New Ones After implementing several RAG systems, I’ve realised something uncomfortable: the solution is creating its own problems. RAG (Retrieval-Augmented Generation) has became the industry’s answer to hallucinations. The logic seemed bulletproof: if models hallucinate from lack of knowledge, give them external documents. Retrieve facts, feed to model, get grounded outputs. Except it doesn’t work that way. I’ve seen this repeatedly in production. Query: ‘What’s Conpany’s stock price’? Retrieved document states $234. Model output: ‘Around $180, showing steady growth’. The correct answer is in context, but the model ignores it. For months, I couldn’t understand why. Then I read the ReDeEP paper from ICLR 2025, which used mechanistic interpretability to investigate RAG systems. Researchers discovered hallucinations occur when Knowledge FFNs, layers storing parametric knowledge from training, overemphasise outdated patterns, while Copying Heads, attention mechanisms extracting retrieved information, fail to integrate external knowledge. This isn’t configuration. It’s architectural. These components fight for control, both writing to the same residual stream. When Knowledge FFNs win, the model ignores retrieved facts and generates from training data. There’s no ‘trust the retrieval’ mechanism, just probabilistic mixing where training patterns dominate. So we added rerankers. Then verification layers. Then confidence scoring. Then fact-checking modules. Each fixing previous failures. Now our ‘solution’ is: ✅ Retrieval ✅ Reranking ✅ Generation ✅ Detection ✅ Fact Checking ✅ Scoring The absurdity: RAG was supposed to reduce hallucinations, but now we have two types: original hallucinations when models fabricate, plus RAG hallucinations when models ignore correct retrieved information and fabricate anyway. We tried solving architectural problems with procedural workarounds. I’m not saying RAG is useless. It helps in many scenarios. But we need honesty. Each layer adds failure modes. We’re building Rube Goldberg machines where every component is probabilistic. For high-stakes applications: healthcare, legal, financial, nobody trusts LLM outputs without human verification. Which defeats automation entirely. I’ve accepted this: hallucinations aren’t bugs to patch. They’re consequences of using probabilistic text generators for deterministic fact retrieval. RAG doesn’t solve this. It moves problems around while adding complexity. The field needs to stop pretending we’re one component away from reliable factual AI. We’re building complex systems to work around fundamental architectural limitations that can’t be worked around. Reference: Sun et al., “ReDeEP: Detecting Hallucination in RAG via Mechanistic Interpretability,” ICLR 2025. #ArtificialIntelligence #MachineLearning #RAG #LLM #AIResearch | 48 comments on LinkedIn

Why SQL Schemas Aren't Ontologies (And Why That Matters for Trustworthy AI)

Analyzing graph data is becoming increasingly important in the modern era, where relationships between people, systems, and transactions matter more than ever. To uncover these hidden connections, we rely on powerful graph analysis tools—and two standout technologies are NetworkX and GraphFrames. These tools help transform complex relationships into meaningful insights, especially in fields like fraud detection, cybersecurity, recommendation systems, and social network analysis. NetworkX vs. GraphFrames — Two Powerful Tools for Different Needs. NetworkX: -A flexible, easy-to-use Python library. -Ideal for small to medium graphs and quick experiments. -Rich algorithms for centrality, clustering, and pathfinding. -Great for research, education, and local analytics. -But not designed for massive or distributed datasets. GraphFrames: -Built on Apache Spark using DataFrames. -Designed for big data and distributed computing. -Can analyze graphs with millions or billions of edges. -Integrates with the full Spark ecosystem (SQL, MLlib, streaming). -Perfect for enterprise-scale, real-time analytics In the financial industry, fraudulent transactions often hide within complex webs of relationships. A financial institution modeled account holders and transactions using GraphFrames—treating people as vertices and transactions as edges. This graph-based model revealed unusual patterns such as clusters of accounts frequently connected through suspicious transfers. Graph analytics provides a new way to understand data—not just as rows and columns, but as a network of connected insights. As data relationships continue to grow in complexity, tools like NetworkX and GraphFrames become essential for data engineers, analysts, and AI specialists. #GraphFrames #NetworkX #GraphAnalytics #ApacheSpark #BigData #DataEngineering #DataScience #MachineLearning #FraudDetection #AI #FinTech #Python #Analytics #BusinessIntelligence #Spark