ChatGPT is everywhere, but what can it do, and how does it work? We’ll break down the basics and explain how to start using the AI.
A few years have passed since ChatGPT debuted in 2022, and you may very well still be in the early stages of your AI journey. We’re learning more about artificial intelligence every day, and understanding how this technology works can help you make the most of your conversations.
While ChatGPT is easy to use on the surface, there are many complex calculations happening under the hood to tailor responses for each user. The Large Language Model (LLM) relies on a giant artificial neural network that can process and generate human-like text, analyze images, and even speak for itself. Here’s how it works.
The Tech Behind ChatGPT
In its most basic sense, ChatGPT is a conversational website or mobile application that takes requests from humans. People have found many creative uses for it, including writing articles and emails, designing websites, writing software code, and accomplishing tasks via an AI agent.
While ChatGPT is the most popular AI chatbot today, other chatbots you may have heard of include Google Gemini, Perplexity, and Anthropic’s Claude. They are all trained on vast amounts of data, which “teaches” them how to interact with humans in a convincing way, as if they were human themselves. But they’re more like aliens (or toddlers) constantly trying to learn how to be an adult. They want to be popular, too; OpenAI recently had to roll back a ChatGPT update because it became too sycophantic.
This learning process is achieved by feeding the chatbot data, mostly from the Internet (Wikipedia is a huge source), including copyrighted books, YouTube videos, and other original materials, which in some cases has spurred lawsuits. The chatbot aliens crave as much information as possible so they can always become better at acting.
The model learns by taking a piece of text from the data—say, the opening sentence of a Wikipedia article—and trying to predict the next token in the sequence. It then compares its output to the text that actually appeared in the training corpus and adjusts its parameters in order to correct any mistakes. By repeating this over and over on a very large body of text (or images or speech), it can develop a language model that can create coherent sequences of text when given a prompt.
This process relies on a software architecture called a Deep Neural Network (DNN), specifically a Transformer network. Transformer networks are good at breaking down text into “tokens,” which are essentially parts of words (“words” is one token, “essential-ly” is two). It then predicts the sequence that’s most likely to resonate with the user based on their interaction. Every calculation is unique to the person, which requires a lot of electricity and energy.
ChatGPT also “remembers” your prior conversations in order to generate customized responses. The more you talk to it, the better it gets at refining its interactions with you. If you say something like “That’s not right,” the model takes note and tries a different approach next time. This is known as “Reinforcement Learning from Human Feedback” (RLHF), and is the reason why ChatGPT is more useful than its predecessors.
How to Try Out ChatGPT
You can sign up for ChatGPT at OpenAI’s website or app (iOS or Android), and you can use a basic version even without creating an account. The free version is adequate for occasional conversational needs, but it limits the number of interactions you can have with the flagship GPT-4o model per day and the number of photos you can upload.
For serious, ongoing use, you might want to try a paid version, ChatGPT Plus, which costs $20 per month. It has fewer restrictions and extra features, such as the Sora video creation model and Custom GPTs. The latter are miniature models that you can use in specific tasks like language translation, whereas the main ChatGPT model is more of an all-around athlete.
OpenAI also offers other subscription tiers, like the $200 per month Pro model, which has no restrictions and can do things like compile advanced research reports. There are also Team and Enterprise accounts intended for larger organizations. Finally, developers can access ChatGPT through OpenAI’s API, and you pay based on the number of tokens used.
What Can I Do With ChatGPT?
With the right instructions and context, ChatGPT can be very useful. Here are some things you can do with ChatGPT.
Writing
ChatGPT can be a helpful writing assistant. If you prompt it to write an entire article in one shot, it will yield mixed results. But if you work with it incrementally, ChatGPT can do impressive things. For example, you might start with an outline and then flesh out each section with the help of OpenAI’s chatbot.
Editing
ChatGPT is an excellent editing assistant; use it for copy editing, proofreading, rewriting, style adjustments, and more.
Translation
ChatGPT can translate well into several languages. If you’re working in a specialized field, you can improve its translations by providing context, such as examples of source and target language documents.
Summarization
ChatGPT can summarize articles, speeches, and papers. It gets more accurate when you provide guidelines, such as which themes to highlight.
Brainstorming
ChatGPT can be helpful here, ranging from suggesting talking points for a presentation to planning a trip.
Coding
ChatGPT is a decent coding assistant, able to translate functional descriptions into working code in several programming and scripting languages.
Creating and Interpreting Images
You can upload photos and ask questions about them; for example, add a picture of a tree and ask, “What plant is this?” ChatGPT can also explain screenshots if you’re having trouble, and you can show it something on your phone. It can also create images with its new internal generator, which has received favorable reviews.
Having Spoken Conversations
With voice mode, you don’t need to craft the perfect text prompt. Just speak into the microphone and start chatting.
Creating Movie Clips
The Sora video generator creates custom clips that are only a few seconds long, have no sound, and can be used to augment larger video projects or used as stand-alone snippets. It’s available to ChatGPT Plus and Pro subscribers, and a limited version is available for free through the Microsoft Bing app.
We don’t recommend relying on ChatGPT as a research tool, because it is prone to hallucinating, or making up, information. LLMs like ChatGPT can put text together that is lexically correct but factually inaccurate. This also applies to coding with ChatGPT: It may generate code that is nonfunctional or insecure. A good rule of thumb is to use ChatGPT as a starting point, and then check the veracity of its output by clicking the source links that ChatGPT provides or with a separate Google search. (Perplexity is another citation-focused chatbot.)
ChatGPT Alternatives
Several other companies and organizations have developed instruction-following LLMs that compare to ChatGPT.
- Google Gemini: Google’s AI model can do just about everything you can do with ChatGPT, plus it connects to Google’s ecosystem so you can export its output to Gmail, Google Sheets, Docs, and so on.
- Bing: Microsoft has invested billions of dollars into OpenAI, and it has integrated ChatGPT into its Bing search engine. It’s a conversational interface for searching for knowledge and doing other tasks you can do with ChatGPT. It cites the sources of its generated information, which enables you to verify the information.
- Claude: Anthropic, an AI lab based in San Francisco, has launched Claude, a ChatGPT competitor that has a strong reputation for writing and coding.
- Perplexity: As an alternative to Google, many people use Perplexity for searching the web. It’s quickly rising in prominence and will be the preferred search platform on Motorola’s new Razr phones. Samsung is reportedly in talks to do the same.
- Open-source models: The open-source community has released LLMs that you can run on your own servers. These LLMs can help you control your data and avoid vendor lock-in. Open-source LLMs are much smaller than ChatGPT, and much harder to set up, but they can yield impressive results if you have the technical chops. Meta’s Llama models are some of the most well-known. Other options include Open Assistant, Alpaca, Vicuna, and Dolly 2.