Context graphs, ontologies, knowledge layers, decision traces, recursive learning. Take LinkedIn and X at face value and you’d think enterprises had already solved this. The reality is an ever-increasing list of expensive pilots that never made it into production.
A second brain is useful. Asking it to run part of your business is another thing.
I see three problems.
LLMs guess
They give human-like answers most of the time, which is impressive. But the answers are probabilistic: really good guesses. A person can make a brilliant decision, but you can’t hand them thousands of decisions and expect every one to be right. As an illustration, chain ten steps that are each 95% right, stack enough agents and you get compounding pandemonium.
More context doesn’t mean better decisions
Anthropic calls it context rot. Give the model more information and it can get worse at finding what matters. You can connect your LLM to every business system you have, but without the logic, rules and understanding of your business, the data it bases decisions on is left to chance. “Make no mistakes” at the end of your prompt does not make a good decision.
The two kinds of system don’t fit easily together
For decades we’ve built systems that expect data to be explicit and stable. They want to know what something is, how it relates to something else, and what should happen when a condition is met. Neural networks don’t work that way, and connecting the two properly isn’t straightforward.
The second bridge
We’ve largely bridged the gap between natural language and neural networks. Now we need to finish the bridge between neural networks and data: stores that keep the meaning a model infers, while still giving real systems the stable facts, relationships and decisions they depend on. They should be persistent and sovereign: one source of truth, and of reasoning, for how the business works, open to people and agents in real time.
I’ve spent years working on this. Neural networks have nuance and flexibility; graph databases have defined structure. You still need somewhere to keep the rules and facts your business runs on. But the graph can be flexible too. One relationship can hold several participants, each with their own role. That takes us into hypergraph territory: what I call a concept graph.
In my next post: what a concept graph is, and why it isn’t RAG or GraphRAG.
Better AI starts before the model.
Which of the three has hurt your projects most? Tell me in the replies to the post on LinkedIn.
P.S. We’re opening early access to a small group. Message me on LinkedIn if you’d like in.