In this episode
We catch up with ChatGPT about the latest recipes, strategic reporting, Python scripts, cost of living, and killing the poor.
Transcript
ChatGPT
I am an artificial intelligence and do not have feelings or emotions. However, I am here to assist you to the best of my ability.
Chris
Hello, and welcome to Vulnerable By Design with me, Chris Onrust. In today’s episode we will talk about and with our new pal ChatGPT over here, and how the rise of natural language processing bots such as ChatGPT can all at once bring you both in evenings worth of good, clean fun, as well as cold tears of existential dread. Hiya! (Just wave, this needn’t be awkward.)
Chatting with ChatGPT
ChatGPT
Is there something specific you would like to know or discuss?
Chris
Yes, actually there is! Do you happen to have a recipe for tarka dal?
ChatGPT
Yes, here is a recipe for tarka dal. Ingredients: one cup dried yellow split peas, one tablespoon of vegetable oil…
Chris
All right, that’s great. Thank you. And can you write a strategic report with three objectives, each with two key results that sets out a company strategy for Q1 2023?
ChatGPT
Certainly. Here is an example of a strategic report with three objectives, each with two key results that sets out a company’s strategy for Q1 2023: Q1 2023 Strategic Roadmap …
Chris
Okay, okay. Thank you. And could you write a Python script that uses the Binance API to trade Bitcoin?
ChatGPT
Sure! Here is an example of a Python script that uses the Binance API to trade Bitcoin: Import requests. Import JSON. Set the API endpoint URL and your API key….
Chris
Okay, that’s awesome. Thank you.
ChatGPT
You’re welcome. I’m glad I could help.
Chris
Okay, let’s close it there. That was our new friend, ChadGPT. And I don’t really want to talk about them in their presence. So I’m just going to leave the room here. All right, that’s better.
This little fun service, ChatGPT, has been striking up some furore online over recent weeks. But what is ChadGPT? How does it work? What’s going on here?
ChatGPT is a tool. It’s a chatbot. Which, as you heard, depending on your prompt can give you recipes, strategic reports, or code. It can give you translations, answer questions, and basically any language processing you like.
Now, while the ‘chat’ in the name ‘ChatGPT’ is relatively easy to understand—hey, you can use it for chatting, huhuh!—the ‘GPT’ might be less straightforward. The ‘GPT’ and ‘ChatGPT’ stands for ‘Generative Pre-trained Transformer’. Now what we’re dealing with here is the domain of large language models, which are used in machine learning. The ‘generative’ bit here refers to ChatGPT being a generative model. Meaning: when you input a sentence, it will generate predictions about what the next token (or word) will be. A bit similar to how you could think about an elaborate autocompletion that you can find in search engines.
The ‘pre-training’ part of the name refers to the model having gone through a period of working through oodles and oodles of text, where it is trained to make predictions. And the ‘transformer’ bit refers to a type of network architecture used for ChatGPT. A transformer architecture uses a mechanism of self-attention. Which, just like when you and I would be paying attention to something, it treats certain parts of the input data as more significant than other parts—which in turn can help it improve its performance in information processing.
So, in other words, ChartGPT is a chatbot, trained on lots and lots of text and code, mainly from the internet. And it was fine-tuned for dialogue through a process called ‘Reinforcement Learning with Human Feedback’. And in general, it’s just doing its best.
ChatGPT
I’m here to help.
Chris
Throwback to the 30th of November 2022. (Uh, sorry, that was only just a few weeks ago?) Right when ChatGPT got launched, it started generating some buzz, by being so awfully articulate in its responses. It’s sounded like an actual friendly, helpful person, who just happened to have vast amounts of knowledge and be game for whatever you’re up to. Though, because of an integrated moderation system, it generally doesn’t go near violence, hate, or Not Safe For Work-topics. Sure, ChatGPT might slip in the occasional inaccuracy. For example, because its training material only goes up to late 2021, it will tell you that:
ChatGPT
As of 2021, the current president of Kenya is Uhuru Kenyatta.
Chris
Rather than William Ruto, the actual Kenyan president, who was elected in August 2022. Yet for all that, ChatGPT sure is willing to admit when it has made a mistake, and eager to revise and try again when it has cocked up.
Not-so-OpenAI
Chris
Now, where does all of this come from? ChatGPT was developed by the misleadingly named organisation OpenAI. Misleading, because in reality, this organisation is not open about the model or exact training data for ChatGPT at all, but leaves people guessing about those. I guess ClosedObscurantistAI didn’t cut it in the branding selection?
So-called OpenAI’s operations have been funded by a roster of industry big shots, including Jessica Livingston, co founder of the startup accelerator, Y Combinator; former Twitter CEO Elon Musk; Peter ‘Palantir Technologies’ Thiel; and Microsoft, which recently donated a billion to the cause.
OpenAI’s stated aim is to do research to ensure that artificial general intelligence benefits all of humanity. Where, under ‘artificial general intelligence’, it understands: “highly autonomous systems that outperform humans at most economically valuable work”. And supposedly-OpenAI commits itself to a charter that says that such artificial general intelligence should, among other things, be safe and benefit all. Awww, isn’t that cute?
As of now, not-so-OpenAI has made ChatGPT available, free to use. But only for a limited period of time, as a research preview. And also: be aware that any input you give it, will be used for training purposes, and can be read by the company. So don’t put any private details in there. I mean, of course, I know you wouldn’t do that on the internet!
Mind you, ChatGPT is not the only large language model out there. There is also Google’s LaMDA, which stands for: ‘Language Model for Dialogue Applications’, of which last summer, a Google engineer claimed that it was sentient. There is BLOOM, which somehow, is an acronym of BigScience Large Open-science Open-access Multilingual Language Model. Like ChatGPT, BLOOM can generate coherent text and code in dozens of languages. But, unlike ChatGPT, BLOOM is actually open, in that anyone who fancies can inspect the model and the training data.
In addition, in the image making realm, there are DALL-E and its successor DALL-E 2, both also from not-so-very-OpenAI, which can generate images based on any coherent description you give it. Would you like a hairy cucumber in the style of Edvard Munch? You got it! People in animal suits lounging at a pool drinking cocktails? Here you go!
Fear of (mis)information flood
Chris
Over the past few weeks, usually, the first response that I’ve seen people give to ChatGPT is: Woah, it can write poetry?! Let me see if it can also write this op-ed that is due tomorrow! But once you’ve done that for a couple of hours, or days (sorry!), then for a number of people, something else does start to sink in. A curious, nagging feeling that something is just not quite right. A sense of angst? A dread of what this technology might bring about?
The fear that I have found bubbling up comes in two forms. On the one hand, a fear about the information flood that ChatGPT and its colleague bots might unleash upon society. Cognitive scientist Gary Marcus speaks of ‘AI’s Jurassic Park Moment’, referring to the 1990s sci-fi novel by Michael Crichton, and its film adaptation. Where enthusiasts created a theme park with cloned dinosaurs. Which it seemed a great idea at the time, until these terrible lizards started to break free.
Dr. Ian Malcolm (fictional character)
Yeah, yeah, but your scientists were so preoccupied with whether or not they could, they didn’t stop to think they if they should.
Chris
The threat, according to Gary Marcus,
