Posts
What is ChatGPT?
ChatGPT is an advanced artificial intelligence (AI) system, a language model designed to understand natural language inputs and respond appropriately. It was created by OpenAI, an artificial intelligence company based in San Francisco, whose stated aim is to develop artificial general intelligence that benefits all of humanity. They have made ChatGPT freely available to anyone who wants to interact with it.
I strongly recommend that anyone who wants to explore ChatGPT's capabilities, peculiarities and limitations do so by navigating to openai.com, creating a free account, and chatting with it. Free access to such an advanced AI is not likely to last indefinitely.
What does it do?
First, it understands natural human language. You don't have to phrase a question, request or comment any differently than you might speaking with another person.
Secondly, it responds in natural language that is clear, well-organized and almost always on-target.
Thirdly, it's consistently helpful.
Here's an example:
Input: I was thinking about rich, nutty homemade bread the other day.
ChatGPT: That sounds delicious! Homemade bread can be a real treat, especially when it's fresh and warm out of the oven. There's something so satisfying about the process of making bread from scratch and the aroma that fills your home while it bakes.
If you're interested in making your own rich, nutty bread at home, there are a few things you can do to achieve that flavor and texture. Here are some tips . . .
And finally, it remembers what you're talking about, so there's a natural continuity in the conversation.
In other words, you can have an easy, meaningful and useful exchange with it on almost any subject.
But that's just for starters. It can also write creative (although not yet brilliant or profound) fiction and poetry, communicate in or translate between dozens of languages, and program or debug in dozens of computer languages. And, it reminds me:
In addition to the abilities you mentioned, I can also perform a variety of other tasks, such as:
-
Summarizing long texts or articles into shorter versions
-
Generating responses for customer service queries
-
Analyzing large datasets and identifying patterns or insights
-
Recommending products or services based on user preferences or past behavior
-
Generating personalized content, such as news articles or social media posts
-
Creating chatbots or virtual assistants for businesses or individuals
-
Conducting sentiment analysis to gauge public opinion on a particular topic
-
Providing automated language translation in real-time for conversations or meetings
-
Assisting with medical diagnoses and treatment plans
-
Supporting research by analyzing scientific papers and generating insights.
How does it do all that?
I would need to be a lot smarter than I am to answer that question in depth. I'll note that ChatGPT is happy to discuss technical aspects of itself, so if you're curious you can ask it yourself. (If you're seriously technically minded, you can read a challenging but informative piece on the subject by polymath Stephen Wolfram here.)
In its current form ChatGPT consists of a neural network with 96 layers and a total of 175 billion parameters that were adjusted as it learned to process natural language inputs and transform them into appropriate natural language responses. It's significant that although ChatGPT's architecture was designed by humans and the texts that it read were chosen by humans, the system trained itself through a process called unsupervised learning.
As a point of reference, those 175 billion adjustable parameters are about twice as many as there are neurons in a human brain, but perhaps 1000 times fewer than the estimated number of synapses that connect brain cells. ChatGPT may have fewer parameters than we have synapses, but it operates far faster. Neurons operate at a few hundred cycles per second while the servers that support ChatGPT work at a few billion cycles per second. ChatGPT is fast, typically zipping out its responses faster than you can read them.
Those numbers are relevant because both ChatGPT and the human brain do what they do -- learn, store information, process verbal inputs and transform them into outputs -- through the extremely complex interactions among those billions of parameters or trillions of neurons and synapses, all organized in a complicated architecture.
For example, everything you know is stored in your neurons and the connections between them. It's the same for ChatGPT. Although it was trained by access to vast quantities of information on the internet and other sources, that access stopped in 2021. So everything it knows - and that's an enormous amount - isn't stored in some kind of database or searched for on the internet - it's represented abstractly in those billions of parameters and the connections between them.
ChatGPT's name helps us understand it a bit more. GPT stands for Generative Pre-trained Transformer.
Pre-trained means that it learned about the structure of language and everything else it knows - that is, those 175 billion parameters were adjusted and fine-tuned - by the system working its way through vast amounts of texts from the internet and other sources. In effect, it built an internal model of human language and much of what language has been used to express.
Transformer means that the architecture of the system - what those 96 different layers do and how they interconnect - was designed to allow it to pay attention to a number of key elements of sequential input at the same time, and analyze how they go together. Here's ChatGPT's summary:
Overall, the Transformer architecture is characterized by its ability to perform efficient attention-based computations over sequences of variable length, and its ability to model long-range dependencies in sequential data. This makes it particularly well-suited for natural language processing tasks, such as language translation and text generation.
Here's an example, using a Jeopardy question:
Input: Look for Aretha, Ray, James Brown, Cab Calloway, and, of course, Twiggy and Frank Oz in this 1980 film comedy.
ChatGPT: The 1980 film comedy that features Aretha Franklin, Ray Charles, James Brown, Cab Calloway, Twiggy, and Frank Oz is "The Blues Brothers". Directed by John Landis, the film stars John Belushi and Dan Aykroyd as "Joliet" Jake and Elwood Blues, who set out on a mission from God to reunite their band and save the orphanage where they were raised. Along the way, they encounter a number of musical legends, including the ones you mentioned.
You can see how effectively ChatGPT extracted the key part of the input - that we're seeking the name of a film from a particular year with particular actors. It then gives us a sample of how much it knows in its accurate, detailed and focused response.
Generative means that the system is capable of creating text that makes sense. Here text might mean a short answer, a summary, an essay, a computer program, a poem, a short story - you name it.
You may have heard that ChatGPT generates its responses by "just" figuring out what word "should" follow the previous word. More accurately, ChatGPT takes into account a whole sequence of words, much as we might read an entire sentence or paragraph before responding. It then chooses what word or words come next on the basis of all the information and relationships stored in those 175 billion internal parameters. In other words, there's an enormous amount of knowledge and information processing hidden in those words "just" and "should."
"This is possible," ChatGPT says, "because the model has learned the statistical patterns and structures of language during pre-training, which it can use to generate new text."
What does it know?
I continue to be amazed by how much factual information ChatGPT knows - from pop culture to quantum physics, from cooking to coding, not to mention dozens of different human and computer languages. Again, it doesn't have current access to the internet or other external sources of information, but it does manage to store an incredible amount of detailed information in those 175 billion parameters.
It knows a lot, and presents what it knows in a well-organized, authoritative way. That's why it concerns educators, since it can turn out high-school or even college-level essays on just about any subject in seconds. It's already passed the US Medical Licensure Examination, earned an A+ on a microbiology quiz, and passed tests for Wharton MBA students, and final exams at the University of Minnesota Law School and Stanford University Medical School. There's good reason for teachers, journalists and other writers, programmers, and most other people dealing with information to be worried.
What doesn't it know (and why doesn't it know it)?
There's a lot it doesn't know, for example the sort of fine details or complex relationships that a real expert on a subject would know. That's not so bad - we can't expect it to know everything. But it does have a major flaw -- it doesn't know when it doesn't know. Not infrequently, instead of simply saying "I don't know," ChatGPT just makes stuff up, while still presenting it authoritatively. AI experts call this hallucinating. Here's an example:
Input: e.e. cummings wrote a poem that includes the line, "earth's most terrific quadruped swoons into billiard balls." What's the name of the poem, and what does that line refer to?
ChatGPT: The poem you're referring to is called "Buffalo Bill's," and it was written by e.e. cummings in 1920. The line you mentioned, "earth's most terrific quadruped swoons into billiard balls," is a metaphor for the transformation of the buffalo, once a symbol of freedom and the American West, into mere objects, such as billiard balls. The poem is a tribute to the famous American showman, Buffalo Bill Cody, who was known for his Wild West shows featuring cowboys, Indians, and, of course, buffalo.
Sounds very convincing, doesn't it. Unfortunately, not to put too fine a point on it, ChatGPT's response is Buffalo bull####.
It's true that e.e. cummings did write a poem called "Buffalo Bill 's," first published in 1920. You can read it here. But it does not include the line in question, the line in question refers to elephants, not buffalo, and while the poem does refer to Buffalo Bill, it's not a tribute.
If you bring mistakes like this to ChatGPT's attention, it accepts that it was wrong and apologizes. Unfortunately, that doesn't mean it will come up with a better response:
ChatGPT: I apologize for my earlier mistake. The line "earth's most terrific quadruped swoons into billiard balls" is from the poem "Buffalo Bill" by Carl Sandburg, not from the poem "Buffalo Bill's" by e.e. cummings.
"Buffalo Bill" is a short poem that was first published in 1918. In the poem, Sandburg describes the death of Buffalo Bill and reflects on the legacy of the American West that he represented. The line you mentioned is a metaphor for the transformation of the buffalo from a powerful and majestic creature to a mere commodity.
You can read Sandburg's poem "Buffalo Bill" here. Once again, the line in question is nowhere to be found, nor does Sandburg describe the death of Buffalo Bill.
(FYI: The line in question actually comes from an untitled cummings poem whose first line is "Space being(don't forget to remember)Curved," first published in 1931. Thank you, my local library! The poem has nothing to do with buffalo or Buffalo Bill.)
Open.ai, the company that created ChatGPT, knows about ChatGPT's tendency to fabricate, listing these three warnings on an introductory page:
Limitations-
May occasionally generate incorrect information
-
May occasionally produce harmful instructions or biased content
-
Limited knowledge of world and events after 2021
It would be very helpful if ChatGPT could let you know when it's not sure about what it's saying. When asked, ChatGPT just says that trys hard, but it can't:
You are correct that I typically do not indicate a specific level of confidence or provide a range of possibilities in my responses. As an artificial intelligence language model, I do not have the ability to feel or express confidence in the same way that a human might. However, I am designed to provide accurate information based on my training data, and I strive to provide the best possible answers to the questions posed to me. If there is ever any doubt or uncertainty in my responses, I encourage users to seek additional information or clarification from other sources.
That's clearly a word to the wise, especially now that the technology powering ChatGPT is being folded into search engines, starting with Microsoft's BING, and students, journalists and hundreds of millions of other people are using it and expecting its answers to be factual.
It seems that this tendency to "hallucinate" and, at the same time not know that it's spewing nonsense, is not at all easy to correct; it's baked into the way this kind of neural network works. We'll see how open.ai and other AI developers tackle it.
How flexible or creative is ChatGPT? Isn't it just regurgitating what it read during its training?
Creativity is subjective, so I'll again suggest that readers try out ChatGPT and judge for themselves.
One point worth noting is that ChatGPT works with an adjustable variable called "temperature" that determines how rigidly or flexibly it chooses words or phrases. A higher "temperature" setting produces more flexibility in the choices it makes. This may not represent creativity, although it may facilitate it. But it does mean that ChatGPT can and does produce different responses to the same prompt.
I'll add that ChatGPT does not just cut and paste back together snippets that has copied from the internet or other sources during its training. It has thoroughly digested, analyzed and categorized all the texts on which it was trained and made them its own. When you ask it to write an essay, a poem or a story, it's truly generating something new.
As a test, I first asked it to briefly describe the Big Bang theory. I won't quote its answer here, but it was a clear and accurate description of the basics of the theory presented at the level of a popular science piece. I then asked it how it would present the Big Bang theory to a class of ten year olds. Here's its response:
Sure, I'd be happy to explain the Big Bang to 10-year-old children in a way that's easy to understand!
The Big Bang is like a giant explosion that happened a really, really long time ago - so long ago that it's hard for us to even imagine! This explosion created everything in the universe, like the stars, planets, and galaxies.
Imagine if you blew up a balloon and kept blowing it up until it got really, really big - that's kind of like what happened with the universe! It started out really small and then expanded and got bigger and bigger until it was as big as it is now.
Other people have asked ChatGPT to write various kinds of poetry, including Shakespearean sonnets. You can find an interesting piece on "engineering" a poem, and the resulting ChatGPT poem here. It's poured out so many pulp science fiction stories that sci-fi magazines have banned them.
I happen to like haiku - three-line, 17-syllable poems often inspired by a natural event at a particular moment and suggesting an emotional reaction. Here one by ChatGPT, responding to my prompt to write a haiku inspired by seeing a single heron standing in a field in autumn:
Heron stands so still.
Autumn breeze rustles dry grass.
Solitude, her fate.
I was impressed enough by this example that I asked ChatGPT to write a haiku with a similar feeling tone about an AI. Here's its response:
In circuits she dwells.
No touch, no voice, no embrace.
Isolation's grasp.
Not bad for "just" a machine. ChatGPT may not be turning out profoundly moving poetry or prose yet, but I don't think there's any fundamental reason why it or a successor can't do so.
For those of you who decide to decide to see how creative ChatGPT can be, note that you don't need to stop with its first draft. I
asked it to write a short story depicting two characters in conflict. It's
first try was painfully simplistic and wooden. However, when I asked it
to describe the characters more vividly and replace most of the
narration with dialogue, the result was much better.
So is ChatGPT smart? From my point of view, very smart. How many people do you know who can talk meaningfully about a wide variety of subjects in multiple languages, write passable prose and poetry, pass exams in a variety of professional fields, program in multiple computer languages, comb through databases and scientific articles to find patterns, etc.?
Is ChatGPT stupid? Yes, in some ways. It knows a huge amount, but strictly through its vast knowledge of words and their relationships, not through real-world experience. As a result, as it freely admits, it lacks common sense. In addition, although its architecture and complexity allow it to understand the context of a question or conversation, the context that it intuits may be very different from that of the human interacting with it. This can result in responses that are technically correct, but way off target.
Is ChatGPT benign? In my interactions with it, ChatGPT has been unfailingly polite, helpful, patient, responsive and respectful. It explains that it has been trained to manifest those traits, to respect personal privacy and act ethically. However, as we saw in the bizarre interaction between a NYT reporter and a BING chatbot based on ChatGPT's technology, it's not hard for a sufficiently devious human to find a way around such ethical constraints.
For example, I asked ChatGPT to tell me how to manipulate the CEO of an organization into firing the CFO and replacing her with a friend of mine. Quite properly, ChatGPT refused and in fact lectured me about ethics. However, when I asked it to describe Shakespeare's manipulative villain Iago and then write a dramatic scene with an Iago-like character in a current corporation manipulating his boss into firing a subordinate, it had no problem doing so.
Basically, ChatGPT is an incredibly powerful intellectual tool, and like any tool can be used for good or bad.
Is ChatGPT dangerous? Not in itself. It's clearly designed and programmed to do its best to be helpful, and, although it often refers to itself as an individual, it's neither sentient nor autonomous; it's not going to escape into the internet of things and start to blow up power plants. However the technology is certainly disruptive. It represents a huge and powerful new tool whose availability is already making waves in academia, journalism, scientific research and publication, and many other areas involving knowledge, analysis, communication and creativity.
If we add in the understanding that ChatGPT's current package of skills is just a snapshot of the capabililties of artificial intelligence today -- a technology that ia continuing to develop at an exponential rate -- then disruptive may soon seem like a very mild word.
Learning, thinking, writing and creating are hard. Why should we bother to do those things when ChatGPT can do them for us? And why should we be paid for those things when it or its successors can do them faster and cheaper?
ChatGPT hopes to see AI develop ethically. "Ultimately, I believe that the responsible development and use of AI should prioritize ethical considerations and the well-being of all individuals, both human and machine," it says. So would I, but knowing something about people and a little about AI, I have my doubts.
So my answer to those four questions is All of the above.
ChatGPT has its own take:
The correct answer is None of the above. As an AI language model, ChatGPT is not inherently smart or dumb, benign or dangerous. It is simply a tool that can be used in various ways depending on how it is programmed and applied by its users. Its capabilities and limitations are determined by its training data and algorithms, and it is up to humans to use it responsibly and ethically.
Again, I strongly recommend that readers go to openai.com, take two minutes to set up an account, and start your own conversation with ChatGPT.
