There’s been a dizzying amount of buzz on GPT-3 and -4, especially on Twitter, and lots of people are trying to make sense of the technology and its implications. Here are some of my thoughts as an A.I. practitioner guided by some questions I’ve been asked from folks who are not tech professionals. Technology is like a lot of systems where things move “slowly, then suddenly” and regardless of what you think about this tech it’s undeniable that we’re now sprinting forward after a long crawl.

What does this technology do?

The chat GPT family of models are software neural networks deployed as chat applications. You ask some questions using natural language, and the bot answers. You can ask anything and will get a response. You can ask it to write code, write jokes; speeches; or poetry, draw diagrams, describe events (it needs to know about the events), and more. You can ask it to imitate someone else: “Tell me the Gettysburg Address in rap form, in the style of Eminem.” The answers are realistic and much of the time, accurate, although there are glaring and important exceptions to the accuracy. The bot is engaging enough to have drawn 100 millions users after 2 months of launch in its current form.

What is this technology, really?

GPT is an “LLM,” which stands for large language model. It’s technically an ensemble of different models, but taken as a whole it’s a black box system (meaning it does some stuff but you can’t understand exactly what) that takes in some human language prompts and outputs understandable responses. We can’t discuss this tech without bringing up the Turning Test which was Alan Turing’s famous thought experiment describing how to test if a machine can “think,” although he posed the problem as an “imitation game,” since he didn’t want to strictly define what it meant to “think.”

The test is administered like this: a human judge chats with two unseen people, but one is a machine. If the judge can’t tell which is which, the machine passes the test. Chat GPT seemingly blows the Turning Test away. In fact, Turing specifically indicated this test was to be administered as a text chat because getting machines to speak realistically was too hard at the time and would break the test (Turing proposed this test in 1950). That barrier has now also been broken, and today with adjacent technologies you can realistically talk to and hear responses from GPT and would have trouble discerning if they came from a machine or human.

Is this technology really just a million people somewhere else in the world answering our questions?

Although we can’t be 100% sure, it doesn’t seem to be. This is the essence of the Turning Test, and because we can’t tell, if it is a machine it would pass Turning’s criteria as being able to imitate a human realistically. It’s worth noting that this has been achieved before using much less sophisticated technology like ELIZA, PARRY and other bots although not nearly to the degree where it can fool as many people as often. The addition of other A.I. voice and video technology compounds the illusion significantly.

How is this technology made?

GPT, and A.I. models in general, use software architectures that are based on the way the human brain works. It’s not an exact model but it uses the basic underlying structure of connected brain cells that interact based on specific rules. Our brains actually make these connections themselves, while these models are designed and built by humans. These models are then schooled on the internet; literally by crawling it: reading posts, viewing images, listening to audio, watching videos and they “learn.”

Each of these internet objects are turned into a string of numbers and those numbers are fed into the software. The software learns to predict what comes next, given a sequence of input. Your questions are a sequence of input and the model, having read all of Reddit, The New York Times, and Twitter, then formulates an answer. It knows how Eminem sounds because it has watched hours of Eminem videos. It knows the Gettysburg Address because it read the Gettysburg Address on Wikipedia. It is designed to process language and knows how to make basic connections between concepts. It then predicts what a combination of those concepts would look like in its answer. In essence GPT is all of the data of the internet compressed into a database that you can talk to.

How does GPT learn how to respond?

This is one of the more interesting emergent properties of this technology. An emergent property is a trait or behavior that emerges from something without a specific effort to implement that behavior or trait. There is no “code” in GPT that describes how to respond to each particular question, it predicts what words it should output and those words produce the effect we are all witnessing when using the software.

What can’t it do well right now?

GPT doesn’t do math or logical reasoning well. If you ask it, for example, to do simple math problems it has trouble. This is ironic as it’s supposed to be a computer, but it shows that the underlying models are not optimized to process logical or mathematical content but rather language. What GPT can do, however, is write a program to get the answer it needs that can be run on a computer. This is how a human would solve the same problem, although humans can train their brains to do simple math in their heads (some of us, anyway).

GPT can’t say “I don’t know” without special programming instructions written by its creators. By default, it will hazard a guess at a question you ask it, without regard to the accuracy of the answer. Because OpenAI is, well, closed, we as users don’t know what those guardrails are.

GPT can only answer on knowledge it has been trained on. Like Humans, GPT needs to be learning constantly to stay relevant. If it stops training on the internet it won’t know things that have been added to the internet since it’s last training.

Is this technology dangerous?

Humans can be dangerous, and human tools can be dangerous. Any technology can be dangerous but the degree to which these models can cause harm may cause us to make societal shifts to adapt. GPT, and A.I. tech in general makes the following nightmare scenarios possible:

• No message, video chat or phone call can be believed. Your mother just called and told you she needs money in an emergency. It looked, sounded and spoke just like her. You got so flustered, you weren’t thinking critically and just followed her instructions, but it was a scammer.
• Court evidence that convicts people today can easily be faked. Emails that sounded like they came from you actually didn’t. There’s a video of you breaking into that car, and you don’t have a solid alibi. There’s a phone message that sounds like you admitting to the crime.
• Hate speech text, audio, and video can be generated in bulk, and disseminated widely and cheaply to influence people and elections. The OpenAI versions of this tech has specific code to limit this but leaked versions of Meta LLaMA’s model have appeared on the internet, complete with weights. The weights of a model are the output of the expensive training that would normally be available only to people with the resources to calculate them. This will allow bad actors to replicate OpenAI’s technology without guardrails.

There is a long tradition of science fiction stores being eerily predictive of the future, and ours may be Black Mirror if we don’t take steps to seriously consider and contemplate the implications of this technology. The real danger would be in not making the governance changes we need to make in order to mitigate the inevitable dangerous use of this technology. We need oversight on this technology’s use, development and deployment. We don’t let the private sector build nuclear weapons, and we need to be equally wary of this technology.

Does the technology really understand the content?

This is the 40 billion dollar question. I don’t know. What is most interesting to me is that it really raises questions like:

• What does it mean to understand anything?
• Do humans understand anything or do we just predict responses based on our own training?
• Are any ideas unique or are they all just combinations of prior knowledge? As Steve Jobs famously said: “great artists steal”

What’s so special about OpenAI that they were able to do this?

In short, money. As is the case with most breakthroughs, the knowledge needed to achieve this was developed over the last decades by thousands of A.I. researchers publicly sharing research and having open discussions. What OpenAI did was to apply it at enormous scale, training on huge datasets using millions of computers. The cost was astronomical. But the data they used was created by all of us, and the knowledge was gained through our education and research systems. Google, Facebook/Meta and others have similar models but were wary to release them earlier, maybe rightfully so.

What are some other implications?

• Trust relationships may change: in the near future, digital communications will have to change drastically in order to maintain trust fidelity between users. If you can only believe what you hear in person, more in-person meetings will happen so the participants are sure that the conversation is real. This will have a major impact on personal lives and work.
• Justice systems may change: Evidentiary rules will change to adapt to a world where what is considered “hard” evidence today will be untrustworthy in the future. There will be a push for more oversight into these systems, watermarking to prove trust using mathematical concepts (maybe even a public ledger like the blockchain whose utility has been questionable up to this point), and generally chaos as the systems adapt to these changes.
• Work may change: Some jobs are definitely at risk. It’s unclear how this will evolve, but this tech can definitely replace some human workers today.

Buckle up, it’s going to be a bumpy ride!