What is an LLM?

In the simplest terms possible, an LLM (Large Language Model) is an AI system trained on vast amounts of text to understand, generate, and predict human language.

While that is an LLM at its essence, it's a pretty vague descriptor, so let's dive into more detail. An LLM is a form of Artificial Intelligence (AI) that is fed massive amounts of data, like books, articles, websites, and code repositories. It uses this data to learn patterns, structures, and nuances of human speech. By analyzing the data, it allows it to go beyond just a regular AI. It can develop an understanding of grammar, tone, reasoning, and even complex subject matter. This makes it able to make conversation and be helpful from only a single prompt.

How Are LLMs Trained?

While there's a lot of debate on the exact process, most sources agree that it comes down to these main three.

The Pretraining

The pretraining or reading phase, is where the AI is fed trillions of words from books, articles and websites. But before any of that can happen, the text first needs to be broken down into smaller chunks called tokens. A token isn't always a full word — it could be a syllable, a punctuation mark, or part of a word. For example, "tokenization" might get split into "token", "iz", and "ation". This is called tokenization, and it's basically how the model converts text into something it can actually process.

After that, it can start to gain the first concept of human language, however, it still has to get refined down to become usable. For example they are given a sentence like "Grass is ____" and they are meant to fill in the blank. If they say "blue" then it would be marked incorrect and the system tweaks itself repeatedly. After doing this millions of times it becomes insanely good at predicting the next word in a sentence.

Supervised Fine-Tuning

This is where engineers solve the problem of AI being extremely smart, but not being able to answer questions like an assistant. For example, if you asked "What is inflation?", the untrained AI might respond with "Inflation is a quantitative measure of the rate at which the average price level of a basket of selected goods and services in an economy increases over a period of time". The point is it sounds extremely textbook like, so engineers train it to be more conversational during this phase.

Post-Training Alignment

During this phase the AI is basically being purged of rude or potentially harmful responses. Reviewers go in and use a technique called RLHF (Reinforcement Learning from Human Feedback), which trains an AI to not use these bad answers to regular queries. After this final stage they are generally deployed out into an environment where it can chat and assist safely.

How does it generate responses?

When you ask an AI a question, it first breaks your prompt into tokens (like what it did with the data in the pretraining phase), by analyzing the tokens of your prompt, it can then build an accurate picture of what you are asking. From there it can generate a response doing the same fill-in-the-blank trick from training, but at lightning fast speeds behind the scenes.

LLM vs. Old AI

Before the time of LLMs, AI had to be preprogrammed for every single response. This made it so if you asked a question that wasn’t in its memory bank, it would completely fall apart. A perfect example of this is when old chatbots got completely hung up on a typo or slang. It had no idea what you were talking about because it was only made to use the premade responses, meanwhile LLMs can actually think.

That's what makes LLMs special, their ability to think. You can try it right now, ask a question with a minor or even a major typo, it will still probably give you an accurate response. Old AI followed rules, LLMs can actually understand language.

Real-World Examples of LLMs

While you probably use some sort of AI for work or daily use, here's a list of some of the major players in the AI world.

• ChatGPT: Initially released by OpenAI in 2022, they have been arguably the most famous example of AI. In their first two months post-launch they hit 100 million active users, so it's safe to say they were the pioneers of personal AI usage. Nowadays they are used for everything from writing, coding, research, customer support, you name it.
• Claude: Released in March 2023, Claude was actually built by former OpenAI researchers who wanted to make a better, safer and more transparent AI model. They are mostly known for being an AI people can trust, and they generally give better, less harmful results. It's used mostly in professional and business settings where accuracy and integrity are extremely important.
• Gemini: Announced by Google on December 6th of 2023, Gemini is Google’s answer to OpenAI’s ChatGPT. Its primary advantage is how integrated it is in the Google ecosystem. It primarily excels in research because of how it pulls directly from the Google search engine.
• Llama: First introduced by Meta on February 24th, 2023, Llama is one of the most unique additions to the AI space is that it's open source — meaning anyone can download it, modify it, and run it themselves. This makes it ideal for developers and companies who want full control over their AI without being locked into someone else's platform or paying per use.

How Do AI Toys Use LLMs?

The big question is how are LLMs being used in AI toys? In essence, AI toys are using LLMs to generate a response or provide intelligence to an AI toy. This makes AI toys distinct because old electronic toys would use preprogrammed responses, usually tied to a button. Meanwhile AI toys can actually hold a conversation, provide learning, answer questions, and genuinely interact with a child.

Conclusion / TL;DR

LLMs are in the simplest terms possible, an LLM (Large Language Model) is an AI system trained on vast amounts of text to understand, generate, and predict human language.

LLMs are an essential part of AI toys, and provide the basic backbone of all its conversation and thinking. As technology progresses, the LLMs behind these toys will as well. What that means for the future of play is only just getting started.