The company OpenAI shows with ChatGPT that computers can write very nice texts these days. Launched over a month ago, the software is based on artificial intelligence (AI) and writes articles, blog posts, stories and even pieces of computer code. Seven questions for experts about ChatGPT.
Since OpenAI launched the ChatGPT software in late November, it has gone viral on social media and in newspapers. Computer scientists, journalists and others interested in technology began experimenting with the software. Columnists investigated whether they could have their column written by the computer and on Twitter circulated the most inspiring and fun examples of computer-generated texts (as well as exciting future predictions from experts about using this software).
What exactly is ChatGPT?
For those who missed it: ChatGPT stands for chat generative pre-trained transformer. It is simple and accessible software that allows you to have a human-like chat conversation. The chatbot can also handle carefully worded questions such as ‘Write a Shakespearean poem about nuclear fusion’. Anyone who types in a question and presses enter will see the result appear on the screen within seconds: an easy-to-read and largely error-free piece of text.
How does ChatGPT work?
ChatGPT is an artificial intelligence tool trained on a huge data set, which is a significant part of the web pages of the Internet. From this, a giant language model has been constructed that, as soon as someone asks the program a question, digs into its memory and puts together the words most likely to provide the answer to that question.
This has nothing to do with real sense of language, it is emphasized computer linguist Tejaswini Deoskar from Utrecht University. ‘Children learn the rules of a language in a few years. They do this with much less data input, so much more efficiently than the software does. It seems clear to me that these huge language models are still missing something. I think that the creators of these huge language models in the future will be inspired by the human way of learning a language.’
What does the launch of ChatGPT mean?
The introduction of ChatGPT to the general public will certainly have major consequences, Deoskar believes. ‘We’ve seen it coming from a scientific perspective for some time, but it’s an added bonus that anyone can now get started with it. Big deal.’
The software’s applications seem endless. Anyone who works with texts can benefit from it. ChatGPT provides a first draft of a story or creates a copy for an ad in no time. In the future, ChatGPT can also be used as an improved search engine that answers in nice sentences. Currently, unlike Google, the software is not connected live to the Internet. ChatGPT is trained with information sources up to and including 2021. The model does not yet have access to the latest information.
Are there any downsides to using ChatGPT?
Besides all the hallelujahs, there are also critical comments about ChatGPT. For example, the software itself does not create new insights, but brews texts from existing information. The software has no idea if these sources are all equally reliable. ‘When I collect information, I decide for myself what can be trusted or not. I will not give up that role’, says Deoskar. ‘The problem is also that ChatGPT formulates texts very firmly. But that suggestion is wrong. And sometimes the software seems to hallucinate.’ Then some nonsense comes out.
The use of powerful software based on a gigantic language model also comes with a price: the significant energy consumption of the servers it runs on. OpenAI CEO Sam Altman admitted on Twitter that each response from ChatGPT costs a few cents in power on average. “At the end of the day, the company will have to make money because of the sky-high cost of computing.” Not only that, it also requires a lot of energy to train a large language model. Deoskar refers to research showing that training a massive GPT-3 language model with billions of parameters equals CO2emissions caused by a round trip from San Francisco to New York.
Should people who write for their profession already fear for their jobs?
It’s an objection that’s always raised with artificial intelligence: that computers will take jobs away from people. Is the soup with ChatGPT eaten so hot? It’s not so bad, say experts. For now, ChatGPT can be a useful tool in many professions. For journalists (‘Find background information on…’), for lawyers (‘Search all court decisions on…’) and even for computer programmers. They spend a good deal of their time figuring out how to formulate a function in a particular programming language. A software tool can help with this. The programmer then only needs to give ChatGPT a command such as ‘write a function that performs these tasks’. This already saves hours of time, programmers said on Twitter.
But in the somewhat longer term, software based on AI can definitely affect people’s jobs. Why would there still be a need for a staffed customer service department when software can search all information in a targeted manner and at the same time handle the most diverse human questions and problems?
Another risk of software breakthroughs with excellent language skills is that criminals also know what to do with it. To write phishing-emails are then no longer necessary, as is the creation of apps that try to get someone to transfer money or click a link. These computer-generated texts soon become indistinguishable from the real thing.
What does OpenAI want with ChatGPT?
Currently, ChatGPT is still free to use – if the website is not overloaded, which often happens – because the developers want feedback from the first users. In the longer term, it is obvious that OpenAI will ask for money for the service. As previously mentioned, running the language model costs a lot of power and therefore a lot of money.
Meanwhile, the model continues to be trained by professional AI trainers who specialize in certain topics. These human AI trainers have conversations where they play both sides – the user and the AI assistant. For example, ChatGPT is getting better at having human-like conversations.
What will happen in 2023?
OpenAI is already working on the successor to GPT-3, the language model on which ChatGPT runs. On the podcast Hard fork from New York Times Tech journalists Kevin Roose and Casey Newton reported that GPT-4 will be released sometime in 2023. “People in Silicon Valley who are into AI research talk about it like it’s magic. GPT-4 is rumored to be trained on 500 times as much data as GPT-3. So it’s going to be much, much more advanced in terms of how many different things it can do,” says Roose. “So I think we’re having a moment of societal wonder about ChatGPT right now, as well as confusion, fear and excitement. But in some ways it has been out of date for some years now.’
Get started yourself
If you want to experiment with ChatGPT yourself, you can register at chat.openai.com.
Opening photo: Shutterstock
If you found this article interesting, you can subscribe to our weekly newsletter for free.