The Case For Process-Guided Models And A New Foundation For Machine Intelligence

By Christopher Sheckler

Have We Been Teaching AI To Think In The Wrong Language?

For decades, artificial intelligence has been pursuing an astonishing goal: teaching machines to understand and communicate in human language

The results of these efforts have been nothing short of revolutionary.

Large language models (LLMs) are capable of writing software, analyzing documents, solving math problems, creating art, and holding conversations with astonishing clarity and complexity. Modern natural language processing has opened up new possibilities and transformed the way humans and computers interact.

But should we be teaching AI to think in language? What if human language isn’t the best possible foundation for intelligence?

These are difficult questions to answer, especially when the most successful techniques in modern machine learning are rooted in linguistics. But this article will explore the idea that the development of artificial intelligence would benefit substantially from a shift away from language and toward processes.

Language Is A Way Of Communicating About Things, Not A Description Of How They Work

Human languages evolved to allow people to describe their experiences to each other

The processes driving gravity, cellular biology, and even electrical circuits, on the other hand, do not require or make use of language at all

Language is a descriptive tool; it is not the underlying mechanism by which things change. To illustrate the difference, let’s examine a very basic process: becoming thirsty. When a person says, “I’m thirsty,” that is a statement of condition.

By stating that fact, they communicate a desire to relieve that thirst. That is a motive.

They can continue to describe how they might satisfy that motive, but the statement of condition is not an action. It is a statement and a desire, not an actual change to the world. A model that reasons about processes should not necessarily reason about language.

The Process-Guided Model Asks “What Is Happening?” Rather Than “What Does This Mean?”

At this point, it’s worth noting that most language models do not actually make calculations using human language. They use numbers instead. The weights, activations, and instructions are all numeric, not linguistic. That is an important observation, because it suggests that if language is not an intrinsic part of a model’s reasoning, then some alternative might serve just as well.

In a large language model, the weights and instructions determine what the model “knows”. Those weights and instructions are usually trained using a language. This is very successful, but it leaves an important question open: if language is not an intrinsic part of a model’s reasoning, would an alternative to language provide better results? This line of thinking leads directly to the concept of the Process-Guided Model.

Using Process-Based Representations To Guide Reasoning

A Process-Guided Model is a system which uses process-based representations as its primary reasoning method.

This is not necessarily incompatible with language. But rather than using language as the primary way of structuring the knowledge it uses for reasoning, the PGM would use the process-oriented relations between things. It would distinguish between what is stated, implied, or even intended, and what is actually happening or could actually happen.

This is a very subtle distinction, but it is extremely significant.

If a machine could reason using processes native to the domain it is working in, it could potentially understand what is happening in that domain at a fundamental level. In a language model, when asked to describe how to fix a problem, the model could provide an excellent explanation of the solution, but it might not be able to actually fix the problem.

In a PGM, on the other hand, the model would be capable of doing more than describing the solution. It would be capable of actually executing the solution. The two approaches are not mutually exclusive. A PGM-based model could be combined with language-generation techniques to form a system which could both explain and execute. But the PGM approach has the potential to provide dramatically better results than either a pure language model or a pure process model. Let’s examine some of the reasons for that.

The Benefits Of A Process-Oriented Approach To Machine Learning

If a PGM can be developed which demonstrates consistently superior performance to a language-based model, that would be an important breakthrough. There are several reasons to believe that such a system would outperform a language-based model in most circumstances.

1. Process-native representations may allow more efficient reasoning than language-based ones.

Language is a powerful tool, but it has limitations. It is not precise; it is not deterministic. It was designed for one-on-one communication and not for computation. If an AI has to reason using language, it cannot do so as precisely or efficiently as it could using process-native representations.

Would that lead to significantly improved performance in large-scale multistep reasoning tasks?

That is a question for testing and experimentation.

But it is a reasonable hypothesis to suggest that using a more deterministic and less ambiguous representation format could lead to improvements in performance.

2. Knowledge transfer between dissimilar domains would be significantly improved.

People routinely make use of analogical reasoning to apply knowledge from one domain to another. A machine learning system using process-native representations could be designed to make use of the same kinds of analogies. By encoding the relationships between items in a way which makes those relationships directly accessible, a PGM could enable rapid transfer of knowledge between different but related domains. This would allow a computer to “understand” a concept, and apply that understanding to a seemingly-unrelated task in another domain.

This technique could be considerably more effective than current methods of achieving knowledge transfer in deep neural networks.

3. Autonomous systems would become more trustworthy and capable.

As AI becomes more capable, these systems need to be able to take action in the real world. Explanation is no longer enough; an AI which can take appropriate action needs to make appropriate decisions. It needs to be able to recognize appropriate and inappropriate actions, select the former and reject the latter. It needs to make sure that the results of its decisions are what it intended them to be.

That is a process-oriented capability.

A PGM-based AI will be significantly more capable in autonomous systems applications, including robotics, infrastructure automation, scientific experimentation, and enterprise logistics and operations.

4. Perception and action could be unified within a common conceptual framework.

The next stage in the evolution of artificial intelligence is likely to involve considerably more integration of perception and action. A modern AI application may need to make decisions based on video and sensor data, computer software state, robotics hardware state, and human instruction.

Natural language is only one of these factors. A system which used a common conceptual framework to reason about all of these inputs would be considerably more capable than one which used language as its primary reasoning paradigm. A process-native representation system could allow an AI to take advantage of all available inputs and produce actions which were responsive to all of them.

The Importance Of Distinguishing Between Description And Capability

One of the most important concepts in artificial intelligence is the difference between being able to describe an answer and being able to actually perform the task required.

A language model might be able to describe how to repair a machine, but it might not be able to actually diagnose a malfunction. It might be able to describe project management techniques, but not actually manage a project when conditions change. It might be able to produce an excellent action plan, but not be able to recognize if the action plan has failed.

These are all significant limitations which stem from the fact that an AI which speaks excellent language is not necessarily capable of performing useful work in the real world.

A PGM-based AI will have significantly better capability in these areas. Its primary focus would be on results and their effects, rather than on presentation.

• Can the system recognize a change?
• Can it predict the result of an action?
• Can it select an appropriate action?
• Can it recognize that an action has failed?
• Can it transfer what it has learned to a new situation?

These are all much more important questions to ask about an AI system’s capabilities than whether it can describe an answer in detail.

Nature makes extensive use of processes to evolve and adapt without the aid of language.

Intelligence in its most general sense is not description; it is the ability to affect change in the world.

Watch this humorous clip as to why human language is not necessarily the most efficient means of process. https://www.youtube.com/shorts/mrmqRoRDrFg

Christopher Sheckler | AI Architecture & Research

#ArtificialIntelligence #ProcessGuidedModels #BeyondLLMs #FutureOfAI #DuchessAI