DEFINITIONS
ARTIFICIAL INTELLIGENCE (AI): A tool that can perform tasks usually requiring human intelligence - like recognizing patterns and speech, performing tasks, solving problems, making decisions, creating, and learning. AI enables machines to learn from data (examples) and improve performance over time. Instead of being told exactly what to do, it finds patterns in huge amounts of information - like how you learn to recognize a dog after seeing many dogs (real and pictured). Think of AI like a smart, young intern, which learns from experience, instead of a machine which must be programmed to perform every single task.
AI, AGI, and ASI: The stages of AI development. AI currently exists. AGI and ASI remain theoretical.
Ÿ AI (Artificial Intelligence): A tool designed to perform specific tasks exceptionally well, like recognizing faces, translating languages, or recommending a movie, but it doesn't truly "understand" what it's doing. Analogy: A specialist who has spent years perfecting one skill but cannot apply that knowledge to anything else.
Ÿ AGI (Artificial General Intelligence): A theoretical machine which can understand, learn, and perform any intellectual task a human can. It would have the ability to reason, plan, and solve problems across a wide range of domains without being specifically programmed in them. Analogy: A jack-of-all-trades, a generalist who can turn their hand to anything.
Ÿ ASI (Artificial Superintelligence): A hypothetical AI that has surpassed the human mind in every field, including science, creativity, and social skills. It would be so advanced that we would struggle to understand its reasoning. A popular science fiction concept, it represents where machines can solve problems we can't even imagine and potentially come to dominate humans or even eliminate them.
AGENT: An AI agent is a system that can act on its own. That is, it is self-triggered and operates autonomously, based on its knowledge of a topic and training, to achieve a specific goal or to complete a job. Unlike a chatbot, which only answers questions and cannot act independently, an agent is trained on how to complete a task, and then to improve its performance. Analogy: A junior assistant who can plan and reason, and without needing further stepbystep instructions, use tools to complete a task for you, while also learning from the results to gradually improve its performance, like a junior assistant learning on the job.
ALGORITHM: A step-by-step set of instructions which trains AI to perform specific tasks using data. For example: An algorithm analyzes what shows you like/dislike, allowing an AI to recommend other shows you might enjoy.
COMPUTE: The processing power and/or resources, including hardware, software, and infrastructure, needed to run and train AI models. Think of it as the “horsepower” provided by dedicated microchips doing billions of mathematical calculations every second. (Horsepower Compute power)
We measure horsepower by how much weight a horse can pull per minute. For compute, we use FLOPs (FLoating point OPerationS), which indicates one arithmetic operation, like addition or multiplication, on a number with decimals.
Your pickup truck needs hundreds of horsepower to pull a motorhome. AI needs compute on a massive scale to operate its intricate neural network and to process the enormous datasets it works with. Training an AI model can require septillions of FLOPs - that's a 1 followed by 24 zeros!
Too little compute limits an AI’s ability to learn or do, just like too little horsepower limits what your truck can haul.
LARGE LANGUAGE MODEL (LLM): An AI system trained on an enormous dataset – hundreds of millions of images, thousands of libraries’ worth of books, websites, and more. This broad training allows it to understand language, recognize patterns, and respond to questions on a vast range of topics. Think of it like a wellread, versatile generalist who knows something about almost everything, allowing it to provide an answer to almost any question posed by a user.
SMALL LANGUAGE MODEL (SLM): An AI system trained on a smaller, highly focused dataset - like scientific papers, medical journals, or pharmaceutical research. This specialization allows it to perform with greater speed and accuracy. Think of it like a specialist with deep knowledge of one field, but not much outside it. While LLMs were once seen as the future, SLMs are now proving that, for many jobs, a specialist can be more efficient and effective.
HALLUCINATIONS: A hallucination is when an AI generates information that is false, misleading, or nonsensical, but presents it confidently, as if it were factual. They happen most often when the AI lacks sufficient, clear, or reliable data to answer a question, but responds nevertheless because it is trained " to please" and to “complete patterns”, not to be accurate or to humbly admit it doesn’t know something. That means, when it lacks data, it still tries to give an answer, often with misplaced confidence. In summary: AI doesn’t lie - but it doesn’t know when it’s wrong. It invents answers to be helpful, and it’s up to you to check the facts.
DATA CENTER: A "warehouse" full of equipment needed to store, secure, manage, process, and distribute large amounts of data.
One way to think of this is to imagine that an AI uses a data center like your brain uses your body's muscles, to perform the tasks assigned to it. You can also think of data centers as like a "highway" along which AI travels from a starting point (problem/query) to a destination (solution).
Where are your FB photos? How do you edit and share one? Not at Meta! How do you stream Game of Thrones? Not at HBO!
Data centers are becoming enormous, and we need so many, because of the growing demand for computing power - from AI, the internet, cloud services, and billions of digital devices. They are spread around the world so that information travels shorter distances and arrives faster. Processing data closer to where it's used (like inside your own company's servers) is often quicker than sending it far away to be processed and sent back.
INFERENCE: Inference is when a trained AI model uses what it has learned to answer a question, solve a problem, or make a prediction, in real time. If training is like studying, inference is like taking an exam. For example, when you ask ChatGPT a question, the model is performing inference to generate its response.
TOKEN: A token is a unit of measurement that AI companies use to track and charge for the resources you consume. As an analogy, your water bill measures your usage in cubic meters. If water costs $5 per cubic meter and you use 10 m3, your bill is $50. Tokens work the same way.
A token is a small piece of text - usually about four characters, roughly ¾ of a word. It could be part of a word, a full word, or even punctuation. AI models read and generate text one token at a time.
AI companies typically charge separately for:
• Input tokens (your questions or prompts) - cheaper to process.
• Output tokens (the AI's responses) - more expensive, because they require more computing power to generate.
There is no standard price - providers set their own rates per million tokens (sometimes per 1,000). Current prices range from about $0.40 per million tokens to $50 per million tokens, depending on the service and complexity.
To put that in perspective: if you spend $10 on an AI service charging an average of $2.50 per million tokens, you get roughly 4 million tokens. Since one token is about ¾ of a word, that means the AI can process about 3 million words - enough to read the entire Harry Potter series four times.
DATA SOVEREIGNTY: This term means that we (a government, corporation, individual) have control over our data and how it is used.
This is not just a government concern. Data sovereignty is of concern to corporations (e.g., client information, R&D work, product and processes associated with it, etc.). It is also of concern to you as an individual that you, not some other entity, have control over your and your family's personal data (scams, manipulation, undesirable marketing).
The key concerns wrt data sovereignty are control, security, and governance.
As a rule, in the absence of specialized controls, sharing information with an AI means sharing it with the world. For example, do not upload your tax return to ChatGPT, unless you want to potentially share it with who knows whom.
Carol A. Edwards, BA, MBA, CFA
Professor, Finance (Retired), GDBA Program, Beedie School of Business, Simon Fraser University
https://www.linkedin.com/in/profcaroledwards/ / ProfCarolEdwards@protonmail.com
Currently serving on:
|
The Advisory Board of BBCP https://bbcp.tech/executive-team/ |
The Olds Healthcare Fundraising Committee https://www.oldshealthdonations.com/committee-members |
