Almost two and a half years ago, I predicted here that you would read articles created by a text generator more often in the future.
The occasion was a veritable flood of text experiments conducted by users of GPT-3: fairy tales, role-playing games, tweets and even computer code, all written by software.
As a reminder: GPT-3 is a machine learning-based language processing model that can be given individual words, sentences or entire paragraphs, which it then supplements itself.
At the time I wrote about the “most eloquent artificial intelligence in the world”.
It now has a successor called ChatGPT.
While the experts wait for GPT-4, ChatGPT is an intriguing intermediate step.
The OpenAI developers have given GPT-3 a human feedback level, giving the system tutoring, so to speak.
The principle is called Reinforcement Learning from Human Feedback (RLHF).
The result is an easy-to-use chatbot that, according to the OpenAI developers, is able to "answer follow-up questions, admit mistakes, challenge false assumptions, and reject inappropriate requests."
He writes computer code, business plans and shoplifting guides, among other things.
Good and accessible enough to get serious work with, although neither I nor ChatGPT would recommend shoplifting themselves.
The bot can also write columns with a little help from Sascha Lobo.
The head of OpenAI reported this week that his program now has more than a million users.
Euphoria and a willingness to experiment prevail among them.
But there is also strong criticism: ChatGPT is indeed extremely eloquent, but that the software is often wrong in terms of content is only recognizable at second glance.
There are some examples:
ChatGPT writes code that looks right but isn't so often that Stack Overflow has banned it.
The help page for programmers wants to protect its users.
ChatGPT writes Wikipedia entries about people that contain so much correct information that all the incorrect ones are only noticed by attentive readers.
ChatGPT writes homework in history, which sounds very convincing thanks to the sources - but in the end only reproduces statistical patterns that it observed in the training data.
This often leads to the language program claiming complete nonsense.
But there are also plenty of successful, impressive counterexamples.
The trick is to be able to distinguish between what is brilliant and what is bullshit.
An AI that writes homework
Of course, schoolchildren have long known the possibilities of AI text generators and in some schools have had to make declarations of honesty that their texts were not written by software.
However, analyst Ben Thompson does not see the technology as the end of homework, even if this may cause disappointment among students.
Thompson proposes a new kind of homework instead, and I think his idea is great:
"Imagine a school buying AI software to be used by all students," he writes.
“Teachers no longer have to insist that everyone write their own essays, they have to insist that the AI do it.
However, the system regularly and intentionally throws out incorrect answers.
The real homework is verifying these answers – the children learn to review and edit information instead of just repeating it.”
That would not only make sense with regard to a future in which the entire network is full of AI-generated content.
First of all, this is not the future, but the present.
The commercialization of the immature technology is already here, and with it the potential for abuse that comes with it, as I and my colleagues have described in this detailed article.
Such exercises would also be useful in general to train a solid bullshit detector, whether it's TikTok videos, Telegram channels, news websites or newsletters.
Our current Netzwelt reading tips for SPIEGEL.de
»Why the mobile phone sounds the alarm on Thursday at 11 a.m.« (five minutes of reading)
Tomorrow is a nationwide »warning day«.
Theoretically.
However, many mobile phone users will be left out during the test alarm of the emergency system.
Torsten Kleinz explains why that is.
»I also have this anger in me« (seven minutes of reading)
With animal and nature videos, Robert Marc Lehmann became a YouTube star.
He doesn't want to be on television, it restricts him »massively«.
Matthias Kreienbrink spoke to him.
"How the whole of Frankfurt could be heated in a climate-friendly manner in the future" (four minutes of reading)
The waste heat from data centers in this country almost always goes unused, an enormous waste.
Max Hoppenstedt explains why the new energy efficiency law will not change much and what cities like Stockholm can do better.
External links: Three tips from other media
Six-Word Sci-Fi: Stories Written by You (English, up to 15 minutes)
Each month, Wired asks its readers to write short stories in six words (or emojis), on a set topic.
This overview page shows all winning texts since January 2020.
"The Jesus simulator makes us lose faith" (three minutes of reading)
Holy shit: The colleagues from "Golem.de" have made a first impression of the upcoming game "I am Jesus Christ".
"GMail: How to use countless email addresses with just one single account" (three minutes of reading)
Not exactly
breaking news
, but in case you've always wanted to know how you can combine "several million addresses" in one GMail mailbox , »GoogleWatchBlog« has three tips for you.
I wish you a week without bullshit,
Patrick Beuth