The YottaGraph and Why AI Needs Explicit Knowledge of the World.

Today, most AIs operate in forgiving spaces. They’re creating fun images, writing executive summaries, and delivering clean answers to questions that used to be hidden inside Stack Overflow threads. If it doesn’t have the information it needs to answer these questions perfectly, then it’s annoying, not catastrophic. But don’t believe for a moment that this attitude will work as we bring AI into big serious enterprise organizations.

There are far more important activities with which we hope AI will help us all, such as quickly providing life-saving intel in disaster situations, preventing people from making terrible mistakes with their finances, providing due diligence on multi-billion dollar deals, or accelerating drug discovery during the next global pandemic. These serious problems require the kind of cognition best described as an investigation, which involves piecing together massive datasets, combining multiple sources, and coping with uncertainty. For these serious problems, the tools built for forgiving spaces simply aren’t enough.

The reason for this failure is well articulated in a recent article from Goldman Sachs. Knowledge can be held and processed in one of two ways:

• Implicit knowledge lives inside a neural network. This knowledge is encoded as weights and activations in a way that possibly mirrors how humans think. No one can point to the exact part of a neural net that knows the Beatles came from Liverpool. This is what makes LLMs so impressive at language, summarization, and brainstorming.
• Explicit knowledge lives in spreadsheets, databases and other structured records. Until the emergence of LLMs, explicit was the only game in town for computers. Explicit knowledge are facts stored as simple sentences and logical or probabilistic statements. Lately in regards to AI, this type of information has been referred to as context, and something that provides context to AI models as a context engine.

Reasoning with implicit knowledge gives amazing and impressive answers to language translation tasks, generative image creation, or high level brainstorming. But, as is witnessed by the failure of almost all large investigative AI projects in industry, implicit knowledge is powerful until you need grounded information. That’s when context with explicit knowledge enters the equation. However, when the investigative question requires touching even just thousands of discrete facts, existing context methods like Google search and RAG fall apart.

Why We Built the YottaGraph

Our day job at Lovelace is Elemental: our context engine builder. We help companies stand up their own explicit knowledge infrastructure by taking a customer’s own data and turning it into something an AI can actually use and investigate rather than hallucinate over. Elemental creates secure, enterprise-specific context engines that are coherent and agent-navigable, transforming an enterprise’s raw, fragmented data at scale into a structured graph that agents can query without constraint. That’s a real product solving real problems, and we could have stopped there.

But we saw a gap early on. A context engine built only on a customer’s internal data is powerful but also blind. It doesn’t know that the counterparty in a trade was just sued by a partner yesterday. It doesn’t know that the supplier the procurement team is about to onboard was a subsidiary of a bank that failed last quarter. It doesn’t know that a public figure being considered for a board seat appears in the Epstein files. An enterprise’s internal data is a small, late-arriving slice of what’s actually happening in the world.

So we made a key decision early on. Before Lovelace came out of stealth, we would prove our technology on the hardest dataset possible: the world itself. We would use Elemental to ingest, resolve, and graph an enormous, fast-moving, deliberately heterogeneous corpus of public data, and we would do it at a fidelity high enough that we’d stake our company on the result. We call this work the YottaGraph.

Read full article on our blog on Medium

[Part 2 of 2. Read Part 1: “We Built Something That Didn’t Exist. Today, We’re Sharing It With the World.”]