Acurai’s audacious claims to have discovered how LLMs operate are now confirmed by studies conducted by OpenAI and Anthropic.

In March 2024, this present author published “Eliminate Chatbot Hallucinations — Yes, Eliminate Them.” This article made the audacious claim that LLMs self-organize around Noun Phrases; and that the behavior of LLMs can be controlled through Noun Phrase manipulation. Recent studies by Anthropic and OpenAI now confirm these to be empirically truths. This is wonderful news! After all, these truths are the basis for eliminating hallucinations — yes, eliminating them.

Noun-Phrase Dominance Model

In March 2024, I presented the revolutionary discovery of the “Noun-Phrase Dominance Model”:

This present inventor’s Noun-Phrase Collision Model led to the development of the higher-level Noun-Phrase Dominance Model — the model that is the key to using LLM token prediction to consistently generate factually accurate output. The Noun-Phrase Dominance Model is perhaps best understood from the perspective of another type of neural network — CNNs (Convolutional Neural Networks).

CNNs are often used for image identification. For example, CNNs can be trained to distinguish images of people, pets, boats, etc. CNNs consist of multiple layers of neurons. Remarkable, during training, these layers self-organize themselves. For example, the early layers self-organize around detecting simple patterns such as edges and textures. The latter layers selforganize by combining the information from earlier layers into more complex patterns like shapes — shapes including the recognition of eyes, ears, legs, steering wheels, etc.

No one tells the CNN to do this. Even though CNNs are merely a collection of neurons with probabilistic weights and biases, CNNs automatically self-organize in this manner in order to fulfill the training objective. While much is discussed in the literature regarding the selforganizing nature of CNN neural networks, little if anything is discussed regarding the selforganizing nature of Transformer Neural Networks — the type of neural network used to construct the most popular Large Language Models such as ChatGPT.

This present inventor’s Noun-Phrase Dominance Model states that neural networks self organize around noun phrases during the training of Large Language Models.

(emphasis in original)

The article then discusses controlling LLM behavior (e.g. ensuring 100% accurate responses) by manipulating the noun phrases that are sent in the query and passages in RAG-based chatbots.

Anthropic and OpenAI Studies Now Confirm Noun-Phrase Dominance Model

LLMs are constructed from multiple layers. In other words, the input (prompt) passes through many layers to generate the output.

Each layer contains many neurons. Each neuron has various values it has learned during training (such as weights and biases).

The Noun-Phrase Dominance model says that neurons don’t operate on their own, but rather self organize around noun phrases. Both OpenAI and Anthropic recently discovered this to be the empirical truth—the actual way that LLMs operate under the hood.

As reported by Axios AI+ on August 23, 2024:

One way AI researchers are trying to understand how models work is by looking at the combinations of artificial neurons that are activated in an AI model’s neural network when a user enters an input.

These combinations, referred to as “features,” relate to different places, people, objects and concepts.

Researchers at Anthropic used this method to map a layer of the neural network inside its Claude Sonnet model and identified different features for people (Albert Einstein, for example) or concepts such as “inner conflict.”

They found that some features are located near related terms: For example, the “inner conflict” feature is near features related to relationship breakups, conflicting allegiances and the notion of a catch-22.

When the researchers manipulated features, the model’s responses changed, opening up the possibility of using features to steer a model’s behavior.

OpenAI similarly looked at a layer near the end of its GPT-4 network and found 16 million features, which are “akin to the small set of concepts a person might have in mind when reasoning about a situation,” the company said in a post about the work.

[Bolded added]

First, notice that Anthropic and OpenAI now confirm that neurons do indeed self organize—just as the Noun-Phrase Dominance Model stated.

Second, notice that the self-organization is not around verbs, adjectives, adverbs, etc. In stark contrast, the neurons self organize around “places, people, objects and concepts.” In other words, the neurons self organize around noun phrases—just as the Noun-Phrase Dominance Model stated.

Third, noun phrase groupings (i.e. features) cluster “near related terms” affirming the existence of Noun-Phrase Routes—just as the Noun-Phrase Dominance Model stated.

Fourth, notice that Anthropic and OpenAI found that manipulating noun phrases can be used to “steer a model’s behavior”—just as the Noun-Phrase Dominance Model stated.

Eliminate Hallucinations—Yes, Eliminate Them

This is remarkable news. After all, the Noun-Phrase Dominance Model is the key to eliminating hallucinations. However, the research community has somehow ignored this model—all while continuing to proclaim hallucinations to be an intractable issue at the same.

Since the March 2024 article, I created a video that uses real-world demonstrations to document the Noun-Phrase Dominance Model, and explains how this is the key to building 100% accurate, hallucination-free, chatbots.

The Noun-Phrase Dominance Model is real. And so is the solution to finally eliminating hallucinations once and for all.

You can build 100% accurate chatbots … today.