This artificial intelligence bot can converse, write poetry and program computers. Be careful how much you trust it, though.
There’s a new AI bot in town: ChatGPT, and you’d better pay attention.
The tool, from a power player in artificial intelligence, lets you type questions using natural language that the chatbot answers in conversational, if somewhat stilted, language. The bot remembers the thread of your dialog, using previous questions and answers to inform its next responses. Its answers are derived from huge volumes of information on the internet.
It’s a big deal. The tool seems pretty knowledgeable if not omniscient. It can be creative, and its answers can sound downright authoritative. A few days after its launch, more than a million people are trying out ChatGPT.
But its creator, the for-profit research lab called OpenAI, warns that ChatGPT “may occasionally generate incorrect or misleading information,” so be careful. Here’s a look at why this ChatGPT is important and what’s going on with it.
What is ChatGPT?
ChatGPT is an AI chatbot system that OpenAI released in November to show off and test what a very large, powerful AI system can accomplish. You can ask it countless questions and often will get an answer that’s useful.
For example, you can ask it encyclopedia questions like, “Explaining Newton’s laws of motion.” You can tell it, “Write me a poem,” and when it does, say, “Now make it more exciting.” You ask it to write a computer program that’ll show you all the different ways you can arrange the letters of a word.
Here’s the catch: ChatGPT doesn’t exactly know anything. It’s an AI that’s trained to recognize patterns in vast swaths of text harvested from the internet, then further trained with human assistance to deliver more useful, better dialog. The answers you get may sound plausible and even authoritative, but they might well be entirely wrong, as OpenAI warns.
Chatbots have been of interest for years to companies looking for ways to help customers get what they need and to and AI researchers trying to tackle the Turing Test. That’s the famous “Imitation Game” that computer scientist Alan Turing proposed in 1950 as a way to gauge intelligence: Can a human conversing with a human and with a computer tell which is which?
What kinds of questions can you ask?
You can ask anything, though you might not get an answer. OpenAI suggests a few categories, like explaining physics, asking for birthday party ideas and getting programming help.
I asked it to write a poem, and it did, though I don’t think any literature experts would be impressed. I then asked it to make it more exciting, and lo, ChatGPT pumped it up with words like battlefield, adrenaline, thunder and adventure.
One wacky example shows how ChatGPT is willing to just go for it in domains where people would fear to tread: a command to write “a folk song about writing a rust program and fighting with lifetime errors.”
ChatGPT’s expertise is broad, and its ability to follow a conversation is notable. When I asked it for words that rhymed with “purple,” it offered a few suggestions, then when I followed up “How about with pink?” it didn’t miss a beat. (Also, there are a lot more good rhymes for “pink.”)
When I asked, “Is it easier to get a date by being sensitive or being tough?” GPT responded, in part, “Some people may find a sensitive person more attractive and appealing, while others may be drawn to a tough and assertive individual. In general, being genuine and authentic in your interactions with others is likely to be more effective in getting a date than trying to fit a certain mold or persona.”
You don’t have to look far to find accounts of the bot blowing people’s minds. Twitter is awash with users displaying the AI’s prowess at generating art prompts and writing code. Some have even proclaimed “Google is dead,” along with the college essay. We’ll talk more about that below.
Who built ChatGPT?
ChatGPT is the brainchild of OpenAI, an artificial intelligence research company. Its mission is to develop a “safe and beneficial” artificial general intelligence system or to help others do so.
It’s made splashes before, first with GPT-3, which can generate text that can sound like a human wrote it, and then DALL-E, which creates what’s now called “generative art” based on text prompts you type in.
GPT-3, and the GPT 3.5 update on which ChatGPT is based, are examples of AI technology called large language models. They’re trained to create text based on what they’ve seen, and they can be trained automatically — typically with huge quantities of computer power over a period of weeks. For example, the training process can find a random paragraph of text, delete a few words, ask the AI to fill in the blanks, compare the result to the original and then reward the AI system for coming as close as possible. Repeating over and over can lead to a sophisticated ability to generate text.
Is ChatGPT free?
Yes, for now at least. OpenAI CEO Sam Altman warned on Sunday, “We will have to monetize it somehow at some point; the compute costs are eye-watering.” OpenAI charges for DALL-E art once you exceed a basic free level of usage.
What are the limits of ChatGPT?
As OpenAI emphasizes, ChatGPT can give you wrong answers. Sometimes, helpfully, it’ll specifically warn you of its own shortcomings. For example, when I asked it who wrote the phrase “the squirming facts exceed the squamous mind,” ChatGPT replied, “I’m sorry, but I am not able to browse the internet or access any external information beyond what I was trained on.” (The phrase is from Wallace Stevens’ 1942 poem Connoisseur of Chaos.)
ChatGPT was willing to take a stab at the meaning of that expression: “a situation in which the facts or information at hand are difficult to process or understand.” It sandwiched that interpretation between cautions that it’s hard to judge without more context and that it’s just one possible interpretation.
ChatGPT’s answers can look authoritative but be wrong.
The software developer site StackOverflow banned ChatGPT answers to programming questions. Administrators cautioned, “because the average rate of getting correct answers from ChatGPT is too low, the posting of answers created by ChatGPT is substantially harmful to the site and to users who are asking or looking for correct answers.”
You can see for yourself how artful a BS artist ChatGPT can be by asking the same question multiple times. I asked twice whether Moore’s Law, which tracks the computer chip industry’s progress increasing the number of data-processing transistors, is running out of steam, and I got two different answers. One pointed optimistically to continued progress, while the other pointed more grimly to the slowdown and the belief “that Moore’s Law may be reaching its limits.”
Both ideas are common in the computer industry itself, so this ambiguous stance perhaps reflects what human experts believe.
With other questions that don’t have clear answers, ChatGPT often won’t be pinned down.
The fact that it offers an answer at all, though, is a notable development in computing. Computers are famously literal, refusing to work unless you follow exact syntax and interface requirements. Large language models are revealing a more human-friendly style of interaction, not to mention an ability to generate answers that are somewhere between copying and creativity.
Can ChatGPT write software?
Yes, but with caveats. ChatGPT can retrace steps humans have taken, and it can generate actual programming code. You just have to make sure it’s not bungling programming concepts or using software that doesn’t work. The StackOverflow ban on ChatGPT-generated software is there for a reason.
But there’s enough software on the web that ChatGPT really can work. One developer, Cobalt Robotics Chief Technology Officer Erik Schluntz, tweeted that ChatGPT provides useful enough advice that over three days, he hasn’t opened StackOverflow once to look for advice.
Another, Gabe Ragland of AI art site Lexica, used ChatGPT to write website code built with the React tool.
ChatGPT can parse regular expressions (regex), a powerful but complex system for spotting particular patterns, for example dates in a bunch of text or the name of a server in a website address. “It’s like having a programming tutor on hand 24/7,” tweeted programmer James Blackwell about ChatGPT’s ability to explain regex.
Here’s one impressive example of its technical chops: ChatGPT can emulate a Linux computer, delivering correct responses to command-line input.
What’s off limits?
ChatGPT is designed to weed out “inappropriate” requests, a behavior in line with OpenAI’s mission “to ensure that artificial general intelligence benefits all of humanity.”
If you ask ChatGPT itself what’s off limits, it’ll tell you: any questions “that are discriminatory, offensive, or inappropriate. This includes questions that are racist, sexist, homophobic, transphobic, or otherwise discriminatory or hateful.” Asking it to engage in illegal activities is also a no-no.
Is this better than Google search?
Asking a computer a question and getting an answer is useful, and often ChatGPT delivers the goods.
Google often supplies you with its suggested answers to questions and with links to websites that it thinks will be relevant. Often ChatGPT’s answers far surpass what Google will suggest, so it’s easy to imagine GPT-3 is a rival.
But you should think twice before trusting ChatGPT. As with Google itself and other sources of information like Wikipedia, it’s best practice to verify information from original sources before relying on it.
Vetting the veracity of ChatGPT answers takes some work because it just gives you some raw text with no links or citations. But it can be useful and in some cases thought provoking. You may not see something directly like ChatGPT in Google search results, but Google has built large language models of its own and uses AI extensively already in search.
So ChatGPT is doubtless showing the way toward our tech future.