Tech Explainer: What are the different types of AI? And how do they differ?
What’s the difference between machine learning and deep learning? LLMs and generative AI? Natural language processing (NLP) and computer vision?
Now that we’re swimming in the deep end of AI, it’s important to draw a distinction between the various types of AI you might encounter.
For example, what’s the difference between machine learning and deep learning? Between generative AI and agentic AI? Between NLP and computer vision?
Making these distinctions can be tricky because AI is really an umbrella term. It generally describes a growing panoply of computer-based functions.
But not all AI systems are created equal. Some are good for certain workloads, but less good for others. To help you sort them out, here’s your guide to today’s main AI varieties.
Machine Learning (ML)
Machine learning lets us create automated systems without having to hand-code thousands of rules.
Previously, users had to painstakingly create a software’s decision logic. This involved attempting to account for every possible eventuality.
But all that changed with the introduction of ML. Now, instead of creating code to define how a system must operate, a user can begin by teaching a system how to interpret data in real time. Then the AI can employ its own logic to determine the most appropriate action.
How does this process work in the real world? Consider the example of office climate control. The old way of doing things required office managers to create fixed logic. This logic would instruct the HVAC system to maintain, say, 68 degrees during business hours and 75 degrees at night and on weekends.
Machine learning enables a smart thermostat to keep a constant watch on the office’s ambient temperature, outside weather conditions, weather trends and current energy prices. Using this information, the system can autonomously make a series of intelligent decisions that keep personnel comfortable while also meeting corporate cost and sustainability requirements.
To paraphrase an old adage, instead of giving the system a fish, we taught it how to fish.
This high-level autonomous functionality comes in handy in many ways. A great example is how banks use ML to fight fraud.
First, a bank trains the ML model by feeding it millions of past transactions, each labeled as either “fraud” or “legitimate.” Then the model uses this data to learn the patterns that separate the two outcomes. Finally, ML scores each new transaction in real time to determine whether it’s fraudulent.
Deep Learning (DL)
Deep learning is a subset of ML. But to go, well, deeper, DL relies on multilayered neural networks that mimic the architecture of the human brain.
Neural networks contain layers of interconnected math functions—the “neurons.” These neurons assign a trust score (referred to as a “weight”) to the information that was passed to it from the previous neuron.
These weights flow downstream through a huge network of layered neurons. Finally, the system tallies the total weight, which leads to a final decision about the data.
That decision could be assigning an identifier—for example, “this is a picture of a car.” It could be a trust score, such as “This transaction is fraudulent.” Or, if the model’s training data is sufficient, it could be just about anything else.
A system enabled by DL doesn’t need manual input from data engineers. Instead, by passing data through layers of artificial neurons, DL can learn directly from raw, unstructured data such as images, audio and text, much the way humans do.
This ability to address unstructured data problems makes DL the best tool for dealing with patterns that are too complex for traditional ML-based systems. For example, major retailers including Walmart rely on deep learning to predict consumer demand for various products. In this case, the AI is making an intelligent guess. But the guess is very intelligent, based on years of previous sales data.
Walmart’s demand-prediction systems also consider a number of other factors. These include seasonality, holidays, weather, local events, promotions, pricing and external market signals.
Armed with that information, Walmart’s purchasing staff can more accurately anticipate demand. That means having the right products available in each retail location, along with additional stock available in nearby warehouses.
Large Language Models (LLMs)
Large language models—also known as chatbots—are a type of ML designed to predict and generate language. LLMs are first trained using vast collections of text, then fine-tuned for specific use cases.
LLM tech is older than many people realize. Sure, ChatGPT only dates back to 2022. But ELIZA—generally considered the very first chatbot—came to life at MIT 60 years ago, way back in 1966.
Today’s chatbots, including Claude and ChatGPT, excel at tasks that require understanding, summarizing, transforming and generating text. These core skills make LLM-based AI particularly good at helping software developers review millions of lines of code.
In this situation, the LLMs can find tiny errors and suggest improvements. These superhuman coding skills can save users countless hours of development time and help lower their costs in the long run.
Organizations that want to self-host open-source models can also opt to use a vLLM. Here, the “v” unofficially stands for virtual.
A vLLM is an open-source library that efficiently serves LLMs by dynamically allocating memory per request. This process is called Paged Attention because large blocks of memory are split into sections called “pages.”
Think of it like the difference between renting a whole warehouse, versus renting only the shelf space you need, when you need it.
Generative AI (GenAI)
Generative AI goes one step beyond the functionality of traditional LLMs. Instead of just classifying and making predictions based on existing data, GenAI can also create novel text, images, audio, code and video.
Training GenAI involves introducing massive data sets. The model uses this data to learn the structure of the content it’s designed to generate.
Then GenAI predicts the most probable next element of its creation, whether that’s pixels, audio samples or anything else. To complete the task, the system repeats this process until it has created content that satisfies a user’s initial prompt.
Businesses use GenAI such as Microsoft Copilot to quickly create emails, reports and meeting summaries. And some marketing departments rely on GenAI creation tools like Adobe Firefly and Suno to produce graphics, video and music for advertising.
However, GenAI users must be prepared to deal with mixed results. GenAI is prone to “hallucinations”—errors that occur when the AI model is not equipped to reasonably satisfy a user’s prompt.
Hallucinations usually happen when a user asks the GenAI to create something that wasn’t accounted for in the model’s training. Instead of just saying “I don’t know,” the AI makes something up.
Natural Language Processing (NLP)
Natural language processing is used to teach machines to read, interpret and generate human language.
But NLP goes deeper than simply reading and responding to language. It can also decode semantic meaning. Then NLP passes on its interpretation to other systems designed to classify, translate and score sentiment.
The data these systems examine are often common business texts. Examples include support tickets, contracts, emails and transcripts.
Some insurers now use NLP via commercial AI-powered software like Ironclad. These applications analyze clauses and alert staff to risks hiding between the lines of dense contracts.
Similarly, Zendesk, a popular vendor of CRM software, uses NLP to create automated sentiment analysis and routing for support tickets. For this use case, an automated system can identify a high-priority ticket submitted by an angry customer. The system would then alert support personnel who can solve the customer’s problem.
Computer Vision
Computer vision models are trained to visually detect objects, classify scenes, track movements, and read text within images and videos. The resulting data is typically collected by digital cameras, medical imaging systems, radar, and other sensors. It’s then passed on to other systems.
The process is designed to replace humans who would previously have had to inspect, monitor and identify images the old-fashioned way—with their eyes.
Where does computer vision come in handy? In one instance, Amazon uses this tech to automatically sort packages and track inventory in its vast warehouses. Other retailers use computer vision to monitor shelves, aiming to reduce lost revenue from theft, breakage and poor inventory control.
Agentic AI
Agentic AI is the latest buzzword, and for good reason. This AI-powered tech could change the way we interact with the world around us.
Agentic AI systems are often built using LLMs. They can plan, use virtual tools (referred to as agents), and enact multistep solutions that satisfy a user’s goal—all without human intervention.
Businesses are already wielding this powerful new tool in myriad ways. For example, some financial institutions use Agentic AI to reconcile complex transactions.
Leading airlines are building agents that let a personal AI agent autonomously book travel based on a customer’s preferences, such as timing, budget, preferred airports and seat selection.
In addition, some corporate sales staff rely on Agentic AI to research and evaluate new sales leads via web-based research and personalized customer outreach.
Will there be more types of AI?
You bet. AI-based systems have already become adept at creating new AI-based systems. In other words, the robots are making more robots, often without much human intervention.
Is that scary or exciting? Perhaps a bit of both.
Do More: