Capabilities of AI-Driven Innovation

This is the second post in our series on Innovation using current AI technology. In this one, we will delve into particular cognitive capabilities, where AI can replace humans or even do a better job.

The revolution concerning large language models (LLMs) took the world by storm, augmenting tasks performed by humans to new heights of consistency, quality and professionalism. The focus has been mainly on amplification of communication skills or their automation especially in terms of personalization using particular contextual information.

It is clear though, that communication by far is not the only potential use-case for the general knowledge-rich reasoning capabilities of LLMs. Furthermore, imagining how these zero-shot machine learning skills, where no training data is required to be able to analyze a task at hand, can be combined via composite AI with prediction, optimization, reinforcement learning and other state-of-the-art machine learning approaches, presents a compelling vision of ground-breaking innovation opportunities.

Focusing on the characteristics of modern AI, we have identified following properties. Lets dive in on LLMs first.

We should start with a disclaimer though: LLMs are at their core just predicting the next word in the sentence. However by compressing the human thought expressed in petabytes of text, they have achieved mindfulness similar to how language is used by humans. While listing these capabilities, it must be clear that every prompt asking the model to do something is a separate action, and therefore we cannot for example ask the model to reveal its reasoning used in the previous answer, since it does not have any consciousness to do so.

Here are some of core LLM properties:

• Breadth of Information: LLMs can cover an incredibly wide array of subjects, from scientific terminology, to culture, to grammar, to programming syntax, to company information and opinions of great thinkers, largely surpassing the breadth of knowledge of any individual human.
• Connecting the Dots: LLMs, by processing all the concepts used for training them in a large neural network, have the ability to connect the dots and not only summarize, but also synthesize, abstract principles, give examples, and find analogies.
• Reasoning: LLMs have shown the ability to perform complex reasoning tasks, such as solving puzzles or explaining scientific concepts in particular contexts, which largely mimics human-level reasoning.
• Common Sense: LLMs can make logical inferences based on general knowledge. When asked to explain the reasoning they are able to provide additional evidence to their claim.
• Creativity: They can generate creative content such as poetry, stories, and even simulate dialogue in different writing styles, showcasing an unexpected level of creativity.
• Problem-Solving: LLMs can assist in solving technical problems, from coding challenges to complex mathematical equations, although especially in producing original sequence of steps, they tend to make mistakes.
• Emotion Recognition: LLMs can detect and respond to the emotional tone of the input, offering empathetic or contextually appropriate responses, a nuanced aspect of human communication.
• Simulating Personalities: They can simulate conversations with historical figures or fictional characters, creating immersive and educational experiences.
• Learning from Interaction: While not self-learning in a traditional sense, LLMs can tailor responses based on the ongoing conversation, showing a form of adaptability.

How lets have a look at more conventional machine learning methods like regression, prediction, classification, clustering, dimensionality reduction, vector embeddings, association rules, anomaly detection, recommendation systems, or reinforcement learning.

• Choosing Best Option: Whether it is next best action in reinforcement learning, new coordinates in swarm intelligence, a type of thing in classification or value of a property in regression, ML is very capable at this kind of task.
• Pattern Recognition: The ability to identify and categorize patterns in data, crucial for methods like classification, clustering, and deep learning.
• Predictive Analytics: Involving forecasting future events or trends based on historical data, as seen in prediction and regression methods.
• Data Categorization: The capability to sort and label data into predefined categories or groups, fundamental in classification and clustering.
• Decision Making under Uncertainty: Demonstrated by reinforcement learning, where the algorithm makes decisions in dynamic, uncertain environments.
• Outlier Identification: The ability to identify data points that deviate significantly from the majority, as seen in anomaly detection methods.
• Insight Extraction from Large Datasets: Techniques like clustering and association rules are adept at uncovering insights and relationships in large volumes of data.
• Adaptive Learning: Machine learning models, especially in reinforcement learning, adapt their strategies based on feedback or changing environments.
• Complex Problem Solving: Employed in deep learning and neural networks, where the models tackle intricate problems like image recognition or natural language processing.
• Data Simplification: Reducing the complexity of data while retaining essential information, crucial for efficient processing and analysis.

Let us now move to the main part of the article, aggregating the previous properties and conceptualizing more nuanced cognitive skills, where AI can replicate human thinking skills and how they can be taken to next level using composite AI that combines traditional ML methods with LLMs:

Analytical

1. Interpretation: LLMs can take information and interpret it from a perspective of a certain profession, stakeholder or given context, thus replacing a human in telling what outcome, resolution or impact given information has on the situation.
2. Traditional ML methods can recommend data or pieces of information that can enrich the LLM context.
3. Essence Understanding & Simplification: LLMs can use their abstraction powers to take information and pinpoint essential message it brings to the table. This can be useful especially when large sums of documents need to be analyzed.
4. Traditional ML methods can preprocess, chunk, and highlight the information analyzed in order to speed up the abstraction process.
5. Integration and Synthesis: LLMs can consolidate various information into a clear structured result.
6. Traditional ML methods like vector databases can be leveraged to search in the shelves of documents and find only the information that is most useful.
7. Labeling & Diagnostics: LLMs can leverage their common sense to put labels on data items in order to create partly synthetic datasets, for which traditionally human labelers would be required.
8. Traditional ML methods like classification can be the next step in order to move to a production ready system, where the LLM produced datasets would be used to train a final model.
9. Identification: LLMs can identify things in data points or information matching given properties that can be described without formal syntax in plain language. Traditional ML methods can however take LLM-labeled datasets and find interesting dimensions, instances that are outliers or ideal in some sense, without any assignment at all.
10. Exemplification & Contextualization: LLMs can take an abstract or theoretical description of a phenomenon and supply an instance and context that fits such properties. This can be a school-book example, a real-world entity like a person or company or a fictional persona or story that manifests some important qualities in order to understand better.
11. Traditional ML methods can supply additional evidence in order to make the example more realistic and complex, as well as filter from the generated examples such that best suit the task at hand.
12. Inference: LLMs can form conclusions derived from supplied evidence (induction) and abstract principles (deduction) in order to support decision-making, analysis, label datasets and create expert systems.
13. Traditional ML methods can recommend the information to be used in the reasoning process as well as use the conclusions.

Explorative

1. Discovery of Relationships: LLMs can be used to find and call meaningful relations between concepts or entities in an effort to highlight certain kind of significance, yet limited by the size of the context window.
2. Traditional ML methods can do such discovery on a massive scale learning with almost certainty to translate one property into another by using the learned relationship.
3. Property Discovery: LLMs can identify and describe any aspect of any property from any angle required.
4. Traditional ML methods can look at the mass of data about an object and uncover even non-obvious properties that show any kind of pattern within the dataset. It might be interesting to use LLMs to improve the quality of the dataset and then let the ML discovery run on richer data.
5. Mapping: LLMs can build new features of the objects analyzed and thus categorize their properties into predefined or on-the-go-created ontologies, adding dimensions to the dataset.
6. Traditional ML methods can then be used to layout the distribution of the object properties in multidimensional space and discover trends about the overall population.
7. Monitoring: LLMs can look at logs and label them in terms of specific properties in order to build a dataset looking at previously unknown properties and train a brand new ML model.
8. Traditional ML methods like classification can thus be working with richer features and used for arbitrary use-case in a matter of weeks, until new dataset is built.
9. Grouping: LLMs can assign group labels to objects according to their properties and descriptions.
10. Traditional ML methods like clustering can then discover meaningful trends in the population and assign group labels from property distribution point of view.

Management and Optimization

1. Property Optimization: LLMs can be used evaluate quality of new properties for which no ML evaluation exists, using reasoning, common sense understanding and possibly very complex prompts.
2. Traditional ML methods can then use these evaluations to discover the relationship among individual properties in the population and thus figure out causal relationships which property drives another. Such criteria can then filter candidates or according to the type of object described determine how such object should be systematically adjusted in order to reach better results in the goal property.
3. Iterative Testing and Refinement: LLMs can be used to test quality of the outputted objects in case no reliable classification models are present. This is useful especially when with the evolution of the model, new requirements are formed in an effort of continuous improvement.
4. Traditional ML methods can run batches of automated standardized tests that make sure that the products maintain certain level of quality.
5. Adaptation and Evolution: LLMs can playfully connect to the previous state of development (for example a conversation) and continue building it forwards. Likewise they can determine the next best action in any possible context without having to rely on a predetermined and limited scope of options.
6. Traditional ML methods like reinforcement learning can make such adaptation decisions with more reliability, as long as the context is well known (based on extensive training on tens of thousands of examples).
7. Coordination: LLMs can evaluate the state of system components or needs of ecosystem players on an ongoing basis and inform the eventual decisions the agents will make.
8. Traditional ML methods like multi-agent systems and swarm intelligence can then use unheard level of information complexity in order to coordinate together. It is likely that replacing humans in agent-like behavior will require a digital twin-like approach where hundreds of factors are being evaluated.
9. Resource Allocation: LLMs can be used to setup the proper way how to look at the overall problem context and identify in which role or influence different elements are acting.
10. Traditional ML methods can then be launched upon the more standardized representation of the situation where similarities to previously known examples form the past can be drawn. That way concrete numeric recommendations can be predicted in order to adjust the presence of certain strategic options in the overall system.

We hope, it is clear from this analysis, what are the brand new never before seen use-cases for a composite combination of LLMs with traditional ML methods and their potential can get us closer to the goal of AGI.

In the next article of the series, we will have a look at particular tasks performed in large organizations. We will evaluate their features and properties and attempt to determine their importance, role and also potential to be fully automated.

‍