Foundation Models: Beyond Text Generation, Unleashing the Power of Reasoning

Why business leaders should see foundat models as “reasoning engines” when considering their AI strategies

iSolutions
6 min readMay 21, 2023

TL;DR:

  • Transformer-based language models like GPT are not only capable of generating text but also possess remarkable reasoning abilities that are often overlooked by businesses decision makers.
  • Given the potential of these models, businesses should focus not only on the text outputs they generate but also on their reasoning capabilities for efficiency, cost reduction, and improved customer satisfaction.

TL;DR Podcast: (2 min)

Original artwork created by AI

Language models based on transformer technology are widely recognized as sophisticated engines for generating text. These models have been utilized for various applications, including email copy, blog posts, poems, and marketing text.

However, their true potential lies in their remarkable reasoning capabilities, which often go unnoticed by many businesses.

When considering AI strategies, business decision-makers must shift their focus from the text generation aspect and start recognizing the power of reasoning offered by these models.

A recent survey conducted among more than 1,500 physical operations leaders across nine countries revealed that 84% of these organizations plan to leverage generative AI and advanced technologies to optimize efficiency. This growing interest in generative AI reflects the widespread belief that language models can provide valuable text outputs.

However, it is crucial to question whether this technology should be solely labeled as “generative AI,” as this term may overlook its long-term value and potential.

Prompt Engineering using the Structure of Language

Transformer models undeniably possess the ability to “generate” text due to their comprehensive understanding of language rules and structures. By studying the principles of language, these models can proficiently complete sentences with remarkable fluency.

For instance, techniques taught in middle school English, such as subject-verb agreement, active vs. passive voice, and paragraph structure, are effectively employed in prompt engineering to create clear and coherent instructions for language models.

Effective prompts leverage the rules of language

Studies conducted by Google AI Language and OpenAI have further demonstrated that incorporating pragmatic language and employing specific and actionable prompts significantly improve the accuracy and fluency of text generated by these models.

Prompt engineering techniques align well with language models because transformer training encompasses a form of semantic grammar. This training is grounded in an understanding that human language exhibits certain regularities and possesses a specific grammatical structure. These regularities include the order of nouns, verbs, adjectives, and other parts of speech.

However, language models, such as GPT, go beyond these surface-level patterns. They capture additional regularities associated with the meaning and semantics of language.

Logic

One notable example of such regularities is logic.

Aristotle’s discovery of logic was a result of personally analyzing patterns in speeches and extracting a formal structure that extended beyond the specific details of each sentence.

Logic provided a way to abstract and generalize the structure of arguments, independent of the content.

What’s interesting is that today’s language models operate at the Aristotelian level, by essentially dealing with templates of sentences using these language structures.

This video demonstrates GPT-4 capabilities with Boolean logic using Truth Tables as a test

Just like George Boole built upon Aristotle’s concept of logic by exploring nested collections of “ands,” “ors,” and “nots” with arbitrary depth, language models discovered logic within language as well.

GPT discovered logic by looking at a lot of sentences effectively and noticing the patterns in those sentences. GPT in a sense, “discovered” the laws of semantic grammar that underlies language.

Logic capabilities enable these models to interpret and understand meaning, marking a significant advancement in computation.

Reasoning

The logic capabilities of transformer models contribute to their ability to reason by enabling them to analyze and process information in a structured and systematic manner.

Transformer models employ attention mechanisms that allow them to attend to different parts of the input sequence and capture dependencies between words or tokens. This attention mechanism helps in recognizing patterns and relationships within the data.

When it comes to reasoning, transformer models can perform logical operations such as conjunction (and), disjunction (or), negation (not), and implication (if-then).

By understanding these logical operations and their combinations, the models can draw inferences and make logical deductions.

For example, if they encounter statements like

“If it is raining, then the ground is wet,”

and

“It is raining,”

the model can logically infer that

“The ground is wet.”

The power of reasoning demonstrated by language models is not infallible, as critics often emphasize their occasional mistakes. However, when these models are provided with the appropriate data, their capabilities are drastically transformed.

Theory of Mind

In the paper, “Sparks of Artificial General Intelligence: Early experiments with GPT-4” in a section titled “Interaction with Humans” there is a discussion about the model’s sophisticated understanding of the “Theory of Mind” (ToM), a fundamental human cognitive process. This understanding, when incorporated into AI systems like GPT-4, allows these systems to reason about human behavior in highly nuanced ways.

GPT-4 demonstrates an understanding of two levels of skills in Theory of Mind:

  1. Basic Skill: GPT-4 can reason about individuals’ beliefs and intentions, a capability highlighted by its ability to answer questions like “What does Alice believe?”
  2. Advanced Skill: GPT-4 can also reason about layered mental states, a more complex skill that enables it to answer intricate questions like “What does Bob think that Alice believes?”
Theory of Mind experiment with GPT-4

These capabilities show that GPT-4 isn’t just repeating patterns in the data — it’s reasoning about the mental states of hypothetical individuals.

GPT-4 can infer human goals, preferences, motives, and expectations, adjusting its responses accordingly. This shows a level of dynamic reasoning about human behavior that goes beyond simple pattern recognition. This learning capability is just one example of evidence of GPT-4’s powerful reasoning engine.

Using GPT’s reasoning capabilities to predict the decisions of jurors after reviewing case details

The paper, “Boosting Theory-of-Mind Performance in Large Language Models via Prompting”, discusses the importance of the capacity of LLMs to reliably perform ToM.

ToM is considered a complex cognitive capacity which is most highly developed in humans. Models that work with social information and with humans will benefit from being able to reason about the mental states and beliefs of agents. ToM tasks often involve inferential reasoning.

For instance, for successful ToM performance, LLMs need to reason based on unobservable information (e.g. hidden mental states of agents) that must be inferred from context rather than parsed from the surface text.

Hence, leveraging these models’ proficiency in ToM tasks could offer valuable potential for a wider range of tasks that require inferential reasoning.

Focusing on Reasoning Capabilities

It is crucial to recognize that the outputs generated by language models represent only a fraction of their true potential.

Business decision-makers should focus on the reasoning abilities embedded within these models to fully harness their power.

AI systems will quickly evolve to access real-time knowledge databases, enabling personalized, context-specific, and highly accurate insights.

Static knowledge bases will no longer constrain AI; instead, businesses will have access to fluid, adaptable systems capable of dynamic interactions, profound understanding, and effective problem-solving.

Embracing the potential of reasoning in AI may seem challenging, given the skepticism surrounding the technology. However, history has repeatedly shown that with the right knowledge and reasoning, daunting challenges can be overcome.

As business leaders tap into the potential of AI and drive its integration into our organizations, considering even the reasoning capabilities of language models today, they can unlock insights, strategies, and superhuman predictions and problem-solving that will shape the future of their businesses.

If you are interested in witnessing the current power of language models as reasoning engines or exploring how they can provide valuable insights, contact us at iSolutionsAI.com

We have created AI solutions that help our clients differentiate themselves from their competition and leverage AI as reasoning engines.

Overall, large language models can help CEOs improve efficiency, reduce costs, and improve customer satisfaction.

By choosing iSolutionsAI to build AI solutions, businesses can harness the power of these tools and achieve their business objectives more effectively.

Trusted AI partners for Fortune 100 enterprise, world’s largest learning platforms, biomedical companies and software creation platform providers along with years of machine learning solutions for business of all sizes.

CONTACT US NOW TO START YOUR JOURNEY

Visit iSolutionsAI.com to start a conversation about how AI can help your organization strategize and safely adopt AI.

--

--

iSolutions

Multiple award-winning experts in custom applications, machine learning models and artificial intelligence for business.