What is Artificial Intelligence?
Artificial Intelligence
Sure, AI is everywhere... but what is it, exactly? We put together an AI primer to bring you up to speed on common terms and concepts in this rapidly evolving field. Our hope is that in understanding some of the groundwork, you can identify elements of AI that could help you in your business.
Artificial Intelligence (AI) has burst on the scene in the last couple years as both the next biggest thing in tech and the newest threat of job security (don’t worry, AI won’t take your job… but someone using AI might).
Artificial Intelligence, broadly defined, is any software system, device, or robot that's trying to mimic human intelligence or otherwise represent human or animal type intelligence. Artificial Intelligence is a broad term encompassing any intelligence demonstrated by machines. This includes the ability to reason, learn, solve problems, and adapt. The goal is creating machines that think and act intelligently.
We have OpenAI to thank for skyrocketing AI into the consumer conscious, but engineers have been building on this technology since the 70s and 80s.
As it gains more attention and becomes more accessible to a wider audience, we want to break down some common concepts within the space. AI is home to various branches of research and almost unlimited applications, and our goal is to share context that can help you understand how it might benefit your business.
There is a lot to unpack within the field of AI, so in the rest of this article we’re going to give you a primer on some common concepts. Understanding these concepts will put you well on your way to envisioning how this technology could help you and your business.
Machine Learning
Machine Learning (ML) teaches machines to improve at a specific task through learning from data without explicit programming. This is achieved through algorithms that analyze data and adjust their internal parameters to improve performance.
AI vs ML
AI is a goal not a technology by itself. By contrast, Machine Learning is a technology (or a family of them) that is most promising for achieving this goal.
It is possible to build AI without ML, but the ML ones have proven to be the most “intelligent” so far. While ML is a powerful tool for achieving AI, it's not the only one. Early chess AIs, rule-based expert systems, and even handcrafted logic-based algorithms fall under the AI umbrella without using ML.
AI is the destination, and ML is one of the vehicles to get us there. It's a powerful vehicle, but there are other paths (handcrafted rules, symbolic reasoning) that could lead us to the same destination of intelligent machines.
Training in ML
ML systems program themselves through feedback cycles of learning. You begin with a system that is as a blank slate. You feed that system an input to the problem you are trying to solve, and it predicts the answer. Then you have what’s called a “loss function” that tells the model how close to or far from the actual answer it was. Based on that feedback, the model updates itself in a way that generally ensures its guess will be better next time.
ML teaches a computer by showing it tons of examples instead of writing exact instructions. You continue these successive feedback loops until the model has a very high success rate, and so we say it has “learned.”
Branches of Research Within Artificial Intelligence
Several key branches or approaches exist within AI, each with its own strengths and applications. AI applications often utilize multiple of these branches of research within a single product. Here are some of the most prominent:
Natural Language Processing (NLP): This focuses on enabling computers to understand and process human language, including tasks like speech recognition, machine translation, and text analysis.
Computer Vision: This equips machines with the ability to "see" and interpret visual information, analyzing images and videos for object detection, scene understanding, and more.
Robotics: This field combines AI, engineering, and other disciplines to create intelligent machines that can interact with the physical world, performing tasks like autonomous navigation, manipulation, and decision-making.
Expert Systems: These are knowledge-based systems designed to mimic the expertise of human specialists in a particular domain, solving problems and making decisions based on stored knowledge and reasoning rules.
Large Language Models (LLMs)
A broad term for describing any generative AI model that is designed to produce coherent strings of text. They are considered “large” because the models themselves contain many values and are trained on very large data sets.
Andrej Karpathy, a burgeoning researcher who recently left OpenAI where he was a founding member, frequently posts in-depth, educational videos about Large Language Models. We highly recommend his talk, Intro to Large Language Models.
Jeremy Howard is the founder of the data science education platform Kaggle and now fast.ai, which provides courses and tools for learning the latest in ML and AI. His video, A Hacker’s Guide to Language Models, is a useful introduction for understanding how LLMs work and the basics of how coders can use them. It’s geared towards coders (”hackers”), but he explains it well even for non-technical people, and it’s helpful for gaining an intuition of how these tools work and are used.
We also love Practical Deep Learning for Coders.
Artificial Narrow Intelligence and Artificial General Intelligence
Artificial Narrow Intelligence (Narrow AI, Weak AI, Specialized AI) is usually built for a single purpose. For instance, a self-driving car might drive intelligently like a human can, but it can’t possibly read a book.
LLMs are bringing us closer to Artificial General Intelligence (AGI, Strong AI, Human-Level AI) which more closely resembles the general intelligence possessed by humans and some animals.
Narrow AI is well-developed and encompasses many of the active applications of AI today, while AGI is largely theoretical. AGI development faces significant technical and philosophical challenges, but of course, could revolutionize various fields with its ability understand and learn any intellectual task, solve problems across domains, and adapt to new situations.
Generative AI
Generative AI is an application of AI focused on creating entirely new content, rather than just analyzing or understanding existing data. LLMs like ChatGPT, Diffusion Models like Dall-E and Midjourney, and OpenAI’s newly-released Sona video generator are all examples of generative AIs.
Sona Prompt: A flock of paper airplanes flutters through a dense jungle, weaving around trees as if they were migrating birds.
Generative AIs have the ability to create original content like text, images, audio, video, and more. It does this by learning the underlying patterns and structures of existing data (training data) and uses them to generate entirely new examples that share similar characteristics.
Generative AI is being used in various applications like personalized recommendations, creating realistic images for games or design, composing music, and even writing different kinds of creative content.
Traditional AI is like a translator, understanding and interpreting information, and Generative AI is like a writer, using its knowledge to create something entirely new. It's a rapidly evolving field with exciting potential to expand our creative horizons and push the boundaries of what's possible.
McKinsey published What Every CEO Should Know About Generative AI, stating “CEOs should consider exploration of generative AI a must, not a maybe. Generative AI can create value in a wide range of use cases. The economics and technical requirements to start are not prohibitive, while the downside of inaction could be quickly falling behind competitors.”
Behind the scenes scaling ChatGPT provides an interesting look at how OpenAI did not expect their little “research demo” to go viral, but were able to keep up with astonishing demand anyway in late 2022 and ultimately productized far more rapidly than most.
Neural Network (NN)
A neural network is a machine learning approach inspired by the structure and function of the human brain, simulating the neurons and connections between them using floating point numbers referred to as “parameters.” These values are updated throughout the training process until the network has learned to make highly accurate guesses.
Once the training process reaches a satisfactory level of accuracy, the parameter values are "frozen," transforming the network into a trained model. This model can then be used for the specific task it was trained on without requiring further adjustments. Importantly, these trained models can be easily shared with others as files, enabling their application across various environments and platforms.
Some examples of how trained models of neural networks can be used across various use cases and industries:
Fraud Detection: Analyze credit card transactions in real-time, flagging potentially fraudulent activity.
Manufacturing: Predicting equipment failures for proactive maintenance.
Healthcare: Analyzing medical images for faster and more accurate diagnoses.
Finance: Detecting fraudulent activity and predicting market trends.
Transportation: Optimizing delivery routes and predicting traffic congestion.
Machine Learning Models
This is a broad term that may cover the structure of a model, the training process, and the resulting file containing the values embedded in the model. Model files are tested, shared, and licensed in the AI/ML community on sites like HuggingFace.
Here are some examples across different categories:
Classification Models:
Logistic Regression: Simple yet powerful model for predicting binary outcomes (yes/no) based on input features. Used in fraud detection, credit scoring, and sentiment analysis.
Decision Trees: Easy to interpret model that makes predictions by following a series of branching rules based on feature values. Used in medical diagnosis, customer churn prediction, and spam filtering.
Support Vector Machines (SVM): Creates a hyperplane to separate data points belonging to different classes. Used in image recognition, text classification, and anomaly detection.
Random Forest: Combines multiple decision trees for improved accuracy and robustness. Used in financial forecasting, weather prediction, and recommender systems.
Regression Models:
Linear Regression: Predicts continuous outcomes based on a linear relationship with input features. Used in stock price prediction, house price estimation, and sales forecasting.
K-Nearest Neighbors (KNN): Predicts new data points based on the values of their nearest neighbors in the training data. Used in image classification, customer segmentation, and recommendation systems.
Gradient Boosting: Combines multiple weak models sequentially to achieve better accuracy than individual models. Used in fraud detection, credit scoring, and natural language processing.
Unsupervised Learning Models:
K-Means Clustering: Groups similar data points together based on their features. Used in customer segmentation, image segmentation, and anomaly detection.
Principal Component Analysis (PCA): Reduces the dimensionality of data while preserving the most important information. Used in image compression, fraud detection, and gene expression analysis.
Autoencoders: Learn compressed representations of data that can be used for image generation, anomaly detection, and dimensionality reduction.
Deep Learning Models:
Convolutional Neural Networks (CNNs): Excellent for image and video recognition, able to learn complex patterns from pixel data. Used in self-driving cars, medical image analysis, and facial recognition.
Recurrent Neural Networks (RNNs): Process sequential data like text or time series data. Used in machine translation, sentiment analysis, and speech recognition.
Generative Adversarial Networks (GANs): Generate new data that resembles the training data, often used for creating realistic images, text, and music.
This is just a small sample of the many different types of machine learning models available. The best model for a specific task depends on the data, the desired outcome, and other factors.
Transformer
Transformer is a neural network-based ML and generative AI approach introduced by Google in a 2017 paper titled “Attention is All You Need.” It has become the standard approach for generative AI.
Embedding
Embeddings are like mini summaries of complex things like text, images, or sounds. They translate these things into a list of numbers that capture their key features and relationships. That list of numbers represents how a particular input is interpreted and “understood” by a trained ML model, which helps represent complex data in a way machines can handle. Because the list of numbers has significance in the context of a given model, two embeddings generated by the same model can be directly compared without having to translate them to output.
Vector Database
A vector database is a specialized type of database designed to store and retrieve data in the form of vectors. In math, vectors are represented as long lists of floating point numbers, which is the same structure as an embedding. Therefore, vector databases are ideally suited for efficiently storing and comparing embeddings. Vector databases are designed to perform this sort of query efficiently. A common use case with LLMs is to search the vector database for the “N nearest neighbors” of a given embedding and then feed those into the LLM as context.
Foundation Models
Foundation models are super-powered, super-versatile AI models. Instead of being trained for one specific task, like playing chess or recognizing faces, foundation models are trained on massive amounts of general data. This lets them perform a wide range of tasks, from understanding and generating text to creating images and even writing different kinds of creative content.
These models are big - with billions or even trillions of parameters, they can learn complex patterns and relationships in data. They can be "fine-tuned" for specific tasks, making them adaptable to different situations. Foundation Models have led to breakthroughs in natural language processing, image generation, and other areas of AI.
Here are some examples of what foundation models can do:
BERT “Bidirectional Encoder Representations from Transformers” is one of the first Large Language Models built on transformers in 2018 and trained on text from Wikipedia. It can do translation, answer questions, and compute the semantic similarity between words.
GPT and ChatGPT: GPT is the foundational next-word prediction model developed by OpenAI and trained on a massive amount of Internet content. It was later fine-tuned by Reinforcement Learning from Human Feedback (RLHF) to become the well-known chatbot that can hold conversations on a variety of topics.
Llama, Jurassic-1 Jumbo, Mistral, the list goes on…: Since the release of ChatGPT there has been an explosion in the training of foundational LLM models all vying to be first to beat ChatGPT and maybe first to achieve AGI. It’s especially interesting to watch the progress of open source models. These tend to lag behind those produced by Big Tech, but not by far!
AlexNet and CLIP: These were early image labeling models that laid the groundwork for later image generation models.
Stable Diffusion & DALL-E 2: These models can generate realistic images based on text descriptions.
Foundation models are exciting because they have the potential to revolutionize many different fields. However, it's important to remember that they are still under development, and there are potential risks associated with their use, such as bias and misuse.
AI Agents
What if an AI could prompt itself and “think” in productive loops like a human does? That is the idea of agents. One way to achieve this might be fully autonomous LLMs that can prompt themselves in loops of thought, trial, and error to achieve useful outcomes. Lots of experiments are happening here, but nothing killer yet. One popular direction is “AI Engineers” who write code and attempt to debug it if errors happen. OpenAI’s recent keynote hinted that agents are what they are building towards. The first steps in that direction are the GPTs and Assistants that anyone can build and release using their toolkits.
In Conclusion
Artificial intelligence is a powerful reality to be harnessed. From streamlining operations to unlocking creative endeavors, AI holds the potential to be your company's secret weapon for transformative growth. By embracing AI with a clear vision and a commitment to using it for good, you can empower your team to achieve remarkable things.
Need to chat with someone with AI chops? Contact us to explore any questions or projects.
Enjoy this article? Sign up for more CTO Insights delivered right to your inbox.
