Our colleague and Operations Manager Alex Kania explains how a newer development in Artificial Intelligence, called a Large Language Model, works- and how you can make I work for you
If the first two decades of this millennium were defined by the emergence and increasing relevance of the internet into public life, the next decades may well be defined by the explosion of Artificial Intelligence.
In a few short years, “AI” (Artificial Intelligence) has transformed from sci-fi speculation to a normal part of many people’s personal and professional lives.
Advocates see this as a transformative tool which could usher in the greatest technological change since the industrial revolution.
Opponents see it as an existential risk factor comparable to nuclear weapons. Skeptics see it as a glorified party trick bandied about to boost share prices while failing to offer anything truly transformative.
I’ll leave it up to you to decide who is right.
In the article that follows, I explain how this current crop of (at times) shockingly human-like services like ChatGPT actually works, and how you might use them.
What do we mean by “AI,” anyway?
AI is a fairly loaded term which conjures images of an impossibly intelligent computer system far beyond our fleshy comprehension. The reality is actually much more mundane – “AI” is a fairly nebulous term which simply refers to a machine which can accomplish a task or tasks typically associated with human intelligence.
This leads to a paradox where once a task becomes more associated with machines than human intelligence, it definitionally ceases to be “AI.” Searching a database, for example, is now almost exclusively thought of as a job for computers, but was once (not too long ago, so I’m told) performed by humans.
In some sense, then, AI has been here for at least the past 20 years – but this is not what’s driving the current “AI Boom.”
Rather, concurrent increases in computational power and development of more sophisticated means of representing language has led to the development of tools that can (seemingly) read and write like people, powered by something called a “Large Language Model.”
What is a Large Language Model?
A large language model or (“LLM”) is a computer program which can process natural language. Models designed to generate output text to respond to the user are considered “generative,” since they generate text as a response, although LLMs also encompass models designed for other functions like classifying text or generating code from a text input.
LLMs earn the name “large” because they are trained on *enormous* collections of text – scanning through millions of books, articles, and websites.
By digesting this vast data, an LLM builds up a statistical model of language, which it uses to produce new sentences that sound fluent and human-like. Think of an LLM as a super-advanced version of the autocomplete feature on your phone.
Just as your phone might suggest the next word when you’re texting, an LLM can predict words to continue a sentence in a sensible way.
The difference is that an LLM’s abilities go far beyond a simple text suggestion – it can generate paragraphs of text, answer complex questions, write stories or essays, and even carry on a dialogue.
Modeling Language
A model is a simplified representation of something designed to preserve some key elements.
For example, a model train preserves the appearance of a train while eliminating or changing other attributes (like changing the size and eliminating the ability to carry passengers).
Modeling language in this context means simplifying it to a series of statistical interrelations between symbols with no preservation of the deeper understanding of the world language represents.
In some sense, language is a model of our experience of the world, shaving off qualia but maintaining enough of some underlying essence to allow for communication.
LLMs are effectively working with a model of a model – although this can produce some very human-like results, it’s important to remember these systems have no deep understanding of language beyond their encoded statistical maps.
One way to think of it is like a very well-read parrot that has seen every possible context for words – it doesn’t know facts or truth per se, but it knows how words are usually used together. As a result, it can produce remarkably coherent sentences on almost any topic.
How Does an LLM Decide What to Say?
When you ask an LLM a question or give it a prompt, how does it come up with a response? It all boils down to prediction.
The model looks at your input and internally tries to predict a suitable next word, then the next, and so on, one word at a time. Each word is chosen based on probabilities the model has learned.
Essentially, the LLM asks itself: “Given everything I’ve seen (in the prompt and in my training), what word is most likely to come next?”, and it outputs that. Then it repeats the process for the following word, continuing until it produces a complete answer.
Here’s an analogy: imagine the AI is continuing a text that it “thinks” you want to see. If you start a sentence, it will finish it in a way that fits the style and context. The LLM doesn’t have a database of perfect answers; instead, it generates answers on the fly by stringing together likely words.
For example, if the prompt is “The actress that played Rose in the 1997 film Titanic is named…” the model will recognize this as a question about a known fact. It will consider the patterns in its training data and likely predict “Kate” then “Winslet” as the next words, forming the answer “Kate Winslet.”
The model arrives at this by having “read” many movie articles and knowing that “Rose…Titanic…is named” is often followed by that name. In essence, the LLM is doing an educated guess based on learned patterns.
It’s important to note that the model isn’t retrieving this answer from a stored fact lookup; it’s generating a response from what it learned.
If your prompt were slightly different or if there were ambiguity, the model’s guess for the next word could be different.
The decision process is statistical: the LLM has a sort of internal compass that was calibrated during training to point to likely continuations.
Because it has so many parameters (like millions or billions of “neurons” adjusting to text patterns), it can capture subtle relations – like understanding that “Rose”, “Titanic”, and “actress” together are likely talking about Kate Winslet. That’s how it determines what to say next.
Another way to think about it: the LLM generates text kind of like how we form sentences when speaking off the cuff.
We don’t plan every word in advance; our brains produce words that make sense as we go.
Similarly, the LLM produces one word at a time in a fluid manner. It doesn’t have a conscious plan or an agenda – it’s just following the direction given by the input and its training.
This is why sometimes the outputs can surprise even the creators of the model: it’s not using a simple script, it’s dynamically weaving a response from learned language patterns.
When AI Gets It Wrong: “Hallucinations” in LLMs
Now that we’ve established LLM “predict” instead of “think”, it may be easier to understand how and why they get things wrong.
LLMs generate text by probability, not by querying a database of verified truths. If the prompt leads into territory the model is uncertain about, it will still produce an answer because that’s what it’s designed to do – generate a continuation.
Unlike a human, LLMs don’t say “I don’t know” unless specifically trained to. They just keeps writing something that sounds right. For users, the key takeaway is: LLMs do not guarantee accuracy. They can embed plausible-sounding falsehoods in their answers.
This is why using an LLM can feel like conversing with a very knowledgeable but sometimes overconfident person who occasionally “bluffs” an answer when unsure.
In practical terms, when you use an LLM (like asking for medical advice, legal information, historical facts, etc.), it’s wise to treat the responses with a healthy dose of skepticism. Use them as a helpful draft or a starting point, but verify critical details from trusted sources.
The technology is improving, and newer models are trying to reduce these hallucinations, but no LLM is 100% reliable on facts.
LLM vs Search Engine: What’s the Difference?
It’s easy to confuse using an LLM with using a search engine like Google, since both can answer questions. However, they work very differently.
Search engines (Google, Bing, etc.) are tools that find and retrieve information. When you search, the engine looks through its indexed web pages for your keywords and returns a list of links to webpages, images, or documents that might contain the answer.
Essentially, a search engine is like a librarian – you ask for information, and it hands you a stack of books or articles (the search results) where you might find what you need. It’s then up to you to read and extract the answer. Traditionally, search engines don’t generate new text; they give you existing content from the web, along with its source.
LLMs (ChatGPT/Bard and similar) are tools that generate content. You ask a question or give a prompt, and the LLM directly produces an answer in natural language. It does not give you a list of sources or direct excerpts unless specifically designed to do so.
Instead, it creates a response on the fly. Using the librarian analogy, an LLM is like a knowledgeable person you ask a question to, and they speak an answer back to you in full sentences, as if they’re explaining or teaching.
The answer is synthesized from what the model “knows” (from its training data), not quoted from a specific webpage.
It’s worth noting that the lines are blurring: modern search engines are integrating AI (Google now often shows AI-generated answers or summaries at the top of search results) to give more direct answers, and many LLM-based services can cite sources or even search the web when generating answers.
But fundamentally, an LLM is not searching the live internet when you ask it something (unless explicitly connected to a search tool). It relies on its pre-existing training data and any provided information to craft a response.
This means LLMs might not have the latest information, whereas a search engine is continuously updated by crawling new web content.
What Can LLMs Help With in Everyday Life?
Summarizing Information: If you have a long article or report and you want just the key points, an LLM can summarize the text for you in a few paragraphs or bullet points. This is like having a speedy reader digest content and present the highlights.
It’s useful for skimming news, research papers, or even simplifying a dense legal document into plain language.
Explaining or Tutoring: LLMs can explain complex topics in simpler terms. Curious about a scientific concept or a piece of history? You can ask an LLM to explain quantum physics as if you’re 5 years old, or to summarize the causes of World War I in a concise way.
Planning and Advice: While they’re not perfect, LLMs can help generate plans or give advice on everyday matters.
For example, trip itineraries (“Plan a 3-day visit to Paris with a focus on art museums”), meal planning (“What’s a healthy dinner idea with chicken and broccoli?”), or personal to-do lists (“Help me create a weekly schedule for my study routine”).
The LLM will generate a structured suggestion that you can then adjust to your needs. They can even play role-based scenarios – like acting as a personal coach giving you motivation tips or a historical figure answering in character, which can be both fun and educational.
Keep in mind though, while LLMs can assist with tasks, they are not perfect and can occasionally produce odd or incorrect results (remember the hallucinations). So for critical tasks, you wouldn’t rely on the LLM alone.
But for every-day, low-stakes tasks, they can save you time and effort by handling the heavy lifting of drafting text or searching through information. In fact, studies have shown that using LLMs can streamline many routine tasks, freeing up time for more important things.
What LLMs Can I Use Today? Do I Have to Pay?
Numerous LLMs are publicly accessible, either directly or through products built on top of them. Some (like ChatGPT, Gemini, Copilot) are directly aimed at end-users and have easy chat interfaces.
Others are more behind-the-scenes but might surface in apps you use. The good news is you don’t need to be a programmer or a tech guru to try them – if you can use a web browser or install an app, you can likely access an LLM.
For the big names, just visit their official websites and they’ll guide you on how to start. (As with any online service, be mindful of official links to avoid scams.)
One great thing about the AI boom is that many LLM services offer free access, at least for basic usage. For instance, ChatGPT has a free version that anyone can use, as does Google’s Gemini and Anthropic’s Claude.
You might wonder why they’re free – often it’s because companies are gathering feedback, improving the AI, or integrating it with their services, so they want as many people as possible to use it.
That said, there are usually paid options or subscriptions for heavy users or for accessing more advanced features.
Using ChatGPT as an example: OpenAI offers a subscription called ChatGPT Plus (roughly $20 a month) which gives access to their most advanced models (GPT-4, which is more powerful than the free version’s GPT-3.5), and also provides faster responses and priority access even when demand is high.
But if you’re a casual user, the free ChatGPT service is quite capable on its own for light use and trying things out. The paid tier is more for enthusiasts or professional use where the small improvements in quality and speed matter.
Is Your Personal Information Safe with LLMs?
If by “safe” we mean confidential – you should assume that what you type into an online LLM might be seen by humans running the service or at least by the AI company’s algorithms.
It’s not broadcast publicly (other users won’t see your specific chats), but it’s also not 100% private like talking to a lawyer or doctor under privilege.
A good rule of thumb is not to share anything with an AI chatbot that you wouldn’t share in an email to a stranger.
Keep your inputs fairly generic and non-sensitive. Asking it to draft a generic business plan is fine; asking it to analyze your personal medical records is not a good idea.
On the flip side, companies are aware of these concerns. They do implement security measures: for instance, OpenAI claims conversations are encrypted in transit and at rest, and only authorized personnel can access data.
They also have policies against the AI requesting personal data from users. So, it’s not that using an LLM is dangerous, it’s just that you are your own best gatekeeper – you decide what to divulge.
In summary, personal information is as safe as you make it when using an LLM. The AI itself isn’t malicious or trying to steal info, but the infrastructure around it means your data isn’t totally private.
Treat an AI chat like a semi-public space: enjoy the conversation, get help with tasks, but don’t spill secrets. If you stick to that guideline, using LLMs can be very safe.
More Reading on AI and LLMs
Alex utilized a range of sources when preparing the article on Artificial Intelligence and Large Language Models. You might enjoy diving deeper into the subject, so we’ve included links to the sources below.
