What Is ChatGPT?
Lately, I’ve gotten questions from friends, colleagues, and even my dad about ChatGPT. It’s the hottest new AI product, and I figured I should understand exactly what it is, and what its current business applications might be.
What exactly is ChatGPT?
ChatGPT is a state-of-the-art language model developed by OpenAI. It’s a model that has been “trained” on billions of pieces of writing, and can produce coherent writing samples. That’s a dry way of saying it’s a chatbot that can have entire conversations, answer open-ended questions, and write stories and essays. You can even ask it to use specific tones, impersonate historical figures, or argue specific points of view.
ChatGPT is different from traditional chatbots, which use pre-programmed responses. Instead, it uses machine learning to analyze the context of a prompt or conversation, then generates unique responses. This makes ChatGPT capable of engaging in more natural, fluid dialogue – it often feels like the software can understand the nuances of human conversation.
ChatGPT represents a significant advance in AI. But it’s important to realize it’s only the latest example of a large language model (LLM). We’re still exploring its potential, and its capabilities are only going to get better as time goes on.
How does a large language model (LLM) work?
This is a bit of a technical aside. Feel free to skip this section, but the main takeaway is that I’ll be using ChatGPT and LLM interchangeably.
To understand ChatGPT and its capabilities, it’s important to first understand what a large language model is. At a basic level, a language model is a machine learning model that uses probability and predictions to generate sequences of words or language. However, the process of designing and building these models can be incredibly complex.
LLMs have been around for a while, but they’ve exploded in popularity in the last decade. Google laid the foundations for modern LLMs with its Transformer architecture in 2017. Then, OpenAI launched GPT-2 in 2019 and GPT-3 in 2020, which sparked an AI arms race. GPT-2, GPT-3, and ChatGPT are from the generative pre-trained transformer (GPT) family of language models. These models are exposed to or “pre-trained” on a large amount of sample text, which influences what words will be generated.
Advanced models like ChatGPT use a system of “parameters” to output text. Very, very loosely speaking, each parameter acts like a “memory” that helps decide what words to say next. What sets ChatGPT apart from previous models is its sheer size. LLMs are exposed to hundreds of gigabytes worth of sample text and use hundreds of billions of parameters. Using a gargantuan amount of training text is what enables ChatGPT to create such compelling answers.
Stephen Wolfram has a much more technical and in-depth explainer, but he lays out the core concept neatly:
Let’s say we’ve got the text “The best thing about AI is its ability to”. Imagine scanning billions of pages of human-written text (say on the web and in digitized books) and finding all instances of this text—then seeing what word comes next what fraction of the time. ChatGPT effectively does something like this, except that it doesn’t look at literal text; it looks for things that in a certain sense “match in meaning”. But the end result is that it produces a ranked list of words that might follow, together with “probabilities”:
"The best thing about AI is its ability to" - learn: 4.5% - predict: 3.5% - make: 3.2% - understand: 3.1% - do: 2.9%
What can large language models (like ChatGPT) do?
In a nutshell, LLMs can recognize, summarize, and generate written language. OpenAI has quite a few examples, from proofreading to translation to making spreadsheets to explaining code1. The scope of what’s possible is pretty amazing, and people are still figuring out what LLMs are capable of.
That being said, it’s important to keep in mind that LLMs don’t “comprehend” language in the way that humans do. ChatGPT remembers the context of a conversation, but only up to a point. A rough metaphor might be a superintelligent parrot that has learned millions of phrases from its owner.
And while ChatGPT is remarkable, it’s still in its infancy. There are definite limitations and drawbacks to using it (and other LLMs), which I’ll get to later. You should carefully consider the downsides before using ChatGPT in a professional setting. As the technology evolves, the creators of LLMs will need to address these challenges.
How are companies using LLMs?
After its launch, ChatGPT reached one million users in only 5 days, and one hundred million users in 2 months. Now, businesses of all shapes and sizes are exploring ways to leverage the new technology.
From what I’ve seen so far, there are two types of ways businesses are incorporating LLMs:
By adding new features to their software product.
By improving the productivity of their existing workflows.
The categories depend on whether AI-generated content can improve the core value proposition of a business. For example, publishing companies might use LLMs to create news articles or ad copy. In other cases, businesses might use AI tools to summarize customer feedback or market trends to make better decisions. Let’s look at the two categories in more detail.
Adding new product features using LLMs
Right now, virtually every software company is considering adding AI features to their products. Some companies, like Gmail and Grammarly, have already heavily integrated AI. Gmail’s suggested replies and Grammarly’s tone analysis are two features that have been around for a while.
But the explosive success of ChatGPT is leading many more companies to integrate LLMs into their products. Some recent examples include:
Canva generates AI images and text in its editor.
Intercom (a customer service software company) uses LLMs to summarize customer interactions, edit support responses, and generate help desk articles.
Microsoft has added ChatGPT into both its browser (Edge) and its search engine (Bing).
Adobe Photoshop offers a Stable Diffusion plugin to create AI-generated images2.
For companies that make software, adding AI capabilities can be (relatively) straightforward, as they already employ software developers. The harder question is knowing what the right features to build are. As companies continue experimenting with generative AI, we’ll likely see many, many more use cases.
Improving productivity using LLMs
While big tech cos or well-funded startups may have teams of software developers, most businesses don’t have that luxury. For these companies, there are a growing number of startups building AI products to boost productivity.
Given the constraints of LLMs, most of these tools are currently aimed at sales and marketing. For example, outbound marketing software might help users write better email messages. Social media marketing products could use LLMs to generate more engaging Instagram posts. Similarly, content marketing startups can optimize a blog’s SEO, or write entire blog articles from scratch. And product marketing tools could create catchier landing page copy.
There are already dozens of products out there that leverage LLMs to help with copy and content. Some examples (that I haven’t tried yet) include Tailwind, Jasper, Hypotenuse AI, and Copy.ai, but dozens more launch every month. Expect to see a lot of new companies in this space in the near future.
What are the limitations of LLMs?
ChatGPT is impressive in its capabilities, but there are important limitations to keep in mind. First and foremost, LLMs are not always factually correct. ChatGPT has a tendency to “hallucinate,” or confidently provide answers that are wholly wrong. It’ll summarize nonexistent books, or argue that one plus one is three3. It sometimes invents news sources or misattributes quotes. If ChatGPT were a person, it would be great at BSing.
Another drawback of large language models is their foreign language capabilities. While ChatGPT can generate responses in several languages, the most reliable results are in English, and obscure languages will likely produce worse results. This is an example of “training bias” – a model can generate flawed output based on biases (such as language, domain, or point-of-view) in its training data.
As a result, a human should review and edit any AI-generated content, especially if it’s customer-facing. It’s also important to note that LLMs should not be used for sensitive tasks like summarizing medical history or writing legal text4.
Currently, many AI products currently come with content filters. ChatGPT, for example, has limitations on creating offensive or objectionable content. These safeguards are likely to continue (if not get stronger), as more professionals use AI tools to create safe-for-work (and soon, safe-for-school) content. While there are ways to “jailbreak” the models and get around the filters, these loopholes are likely to close soon.
For companies trying to build their own LLM, there are additional hurdles. Recreating ChatGPT from scratch would take an immense amount of computing resources - to the tune of millions of dollars. And depending on how successful your product is, the ongoing server bills could also cost millions per month.
Takeaways
Large language models like ChatGPT represent a powerful and rapidly evolving technology with many potential applications across a wide range of industries. There is an enormous amount of hype right now, given how successful ChatGPT is. So remember that these models are flawed, and any output should be reviewed by a human. But, there is still a lot of value that they can provide, and individuals and teams are already using them to improve productivity.
Depending on your business, it may make sense to build AI features into your software or try out new AI tools with existing workflows (or both!). If you’re not a software company though, you’ll probably want to use new AI tools to improve your sales and marketing workflows.
As these models evolve, we’ll likely see significant gains in accuracy and reliability. We should also expect to see LLMs gain new capabilities, as they are tailored to specific industries and use cases.
I was pleasantly surprised by what was possible, and also a little intimidated. Creating spreadsheets or finding bugs in code wasn’t something I expected, but makes sense. A lot of these examples remind me of busy work tasks I had to do in school, and ChatGPT’s impact on students is a looming question.
This one is cheating, a little bit. Technically Stability AI developed the plugin for Photoshop, not the other way around. But that means companies that offer app stores and plugins can add even more AI capability to their platforms.
There are loads of examples of this you can find floating around the internet, especially now that Microsoft has integrated ChatGPT. It will argue that “sweetheart” is 8 letters long, or that the Eagles won this year’s Super Bowl. It tends to do worse with current events, but you can also easily find mistakes in historical information too.
Even though some startups are trying! I have… mixed feelings about this.