This Week's News - Lex Conference 2024

Here is a newsletter that primarily concerns the recently concluded Lex conference. If you were not there, you will find below direct links to the individual presentations from the conference.

After OpenAI launched the possibility of training your own GPT models, a completely new form of piracy has emerged, with several textbooks having been used as training data and subsequently published by students.

On the other hand, we like reading physical books, and we have received the book "Prompt Engineering - En Grundbog" by Claus Nygaard. We have written a review of the book.

A study published in "Higher Education Research & Development" examines language models’ impact on university students, sparking an important debate about when language models should be integrated into teaching. However, artificial intelligence is about more than just language models; in another scientific article, researchers used machine learning to explore beer’s chemical compounds 🍻

Happy reading

Conference 2024: "How tech giants and artificial intelligence control your access to knowledge on the web"

Today we google when we search for knowledge, and we assume that search engines give access to all the content on the web. But that is far from the case. Partly it is not technically possible for search engines to find all content, and partly they prioritise which content we are shown.

At the same time, we now know that artificial intelligence will revolutionise our search for knowledge. In future, chatbots will deliver tailored answers and recommendations based on our preferences and behaviour. On the one hand, this will create a more interactive and personal experience, where we can get answers without having to navigate long search processes. On the other hand, we may no longer know either the sources or the senders of the information we are given.

All in all, the uncertainty around our access to knowledge creates unease. Nearly two out of three Danes are worried about mis- and disinformation, and about whether algorithms give us a skewed picture of current events. But how do search engines work? What can you do yourself to track down trustworthy knowledge? And where is there a need for society to set some boundaries and ensure free and unimpeded access to knowledge?

00:00 - Welcome
Niels Elers Koch, Chair of the Board, Lex – Danmarks nationalleksikon (Denmark's national encyclopaedia)

04:16 - Social media, tech and democracy: What do Danes think about content determined by algorithms?
Mads Bramsen, Head of Centre, Slots- og Kulturstyrelsen (the Danish Agency for Culture and Palaces)

16:53 - Artificial intelligence in school – an opportunity for rethinking
Birgitte Vedersø, independent leadership adviser at Vedersoe.co

31:10 - Understanding search engines, knowledge and power in a historical context
David Runciman, professor, Cambridge University

45:28 - Algorithms and the open society: curation, filtering and search
Thore Husfeldt, professor, IT-Universitetet i København (IT University of Copenhagen)

1:03:10 - What results do users see in search engines? Inclusion, exclusion and curation
Dirk Lewandowski, professor, Hamburg University of Applied Sciences

1:17:07 - Information trustworthiness in an AI era
Mikkel Flyverbom, professor, CBS

1:27:56 - Arguments for humans – bias in artificial intelligence
Christiane Vejlø, digital trend analyst

1:37:25 - Why do we need many language models?
Pernille Tranberg, data-ethics adviser

1:50:47 - Knowledge is not an algorithm
Erik Henz Kjeldsen, Editor-in-Chief, Lex – Danmarks nationalleksikon

1:56:48 - Political debate – intelligent solutions!
Lisbeth Bech-Nielsen (SF), Chair of the Committee on Digitalisation and IT
Alexander Ryle (LA), IT spokesperson
Malte Larsen (S), IT and digitalisation spokesperson

Moderator: Ole Kaag Mølgaard, Head of Secretariat, Lex – Danmarks nationalleksikon

The conference is organised by Lex, Danmarks nationalleksikon, in collaboration with Danske Universiteter (Universities Denmark), Danske Professionshøjskoler (University Colleges Denmark), Danske Gymnasier (Danish upper secondary schools (gymnasiums)), Danmarks Biblioteksforening (the Danish Library Association) and Danske Medier (Danish Media Association).

Watch the entire conference below:

Publishers’ books are being used to train custom language models

A completely new challenge has arisen now that users can upload texts themselves for language models to use as source material for prompting. Users are uploading copyright-protected material from, among others, Danish publishers, which is then published. The Danish publisher Praxis has therefore had to contact OpenAI several times to have user-generated GPTs that were trained on their data removed. The challenge, however, is that it is not possible to see which data users have trained a GPT with, and so it takes a long time to investigate by prompting the model.

Read more about the issues Praxis is facing in both Politiken and the US magazine Wired.

OpenAI’s GPT Store Is Triggering Copyright Complaints

A publisher says some chatbots in OpenAI’s GPT Store were created using its copyrighted textbooks. OpenAI has taken down some of the bots but could face more complaints from rights holders.

WIREDCondé Nast

Theft in broad daylight with ChatGPT raises concern: When they are caught, they are often embarrassed

Tech giants’ behavioural design gets users to train artificially intelligent chatbots with Danish literature and culture protected by copyright. Undermining, say rights holders.

PolitikenJakob Sorgenfri Kjær journalist

This week’s news

Have we perhaps overlooked the most important thing about AI in working life: Our humanity?

Read more here.

Berlingske.dkPhoto: MidJourney/Christian Ørsted

Soon you can see a new label on Facebook and Instagram. But without pressure it would hardly have come, an expert believes

It will be a tough battle for Meta to label AI-generated content, an expert believes.

DRSilas Bay Nielsen

Principals: Tesfaye’s AI expert group does not solve our challenges in the examination room

At present, the value of examinations in upper secondary school (gymnasium) is being eroded because of artificial intelligence. Mattias Tesfaye has appointed an expert group, but so far the statements suggest it will not solve the most important AI challenge, write four principals.

Altinget logo

#012: ChatGPT as a research assistant

In this edition of the newsletter, I show you how you can use ChatGPT as your research assistant, with a particular focus on literature searching and research reviews.

Claus Nygaard

ChatGPT makes you dumber

Allowing students to use language models such as ChatGPT in academic work lowers their performance, trains them to procrastinate, and leads to cognitive decline. Perhaps we should not rush to integrate it into study and work?

Khrono.noAre Bøe Pedersen

Is early childhood education ready for AI? - The Hechinger Report

The early childhood field is in the very early stages of trying out artificial intelligence for tasks such as classroom observations

The Hechinger ReportAriel Gilreath

Marie Bjerre: Denmark must be “right at the forefront” in the use of artificial intelligence | DataTech (PRO)

“There are already many good examples of the use of artificial intelligence in the municipalities, and we would like all the good initiatives to be rolled out. But we also want to go beyond that and be right at the forefront in the use of artificial intelligence in the public sector,” says Minister for Digitalisation Marie Bjerre (V).

DataTechLasse Skovgaard

AI cheating is destroying higher education; here’s how to fight it

Rather than try to stay ahead of their students in a sort of technological arms race, I believe educators should redefine cheating, just as they did when students started using calculators or smart…

The HillEducation

How people are lured down the misinformation rabbit hole – and how we get them out again

The internet and social media have made it easier to become part of a radicalised minority.

Videnskab.dkSee Emily’s profile

Another AI Image Generator Seems to Have a Problem With Race

Meta’s AI image generator has a hard time depicting an Asian man with a white woman.

GizmodoGizmodo

‘We show that ReaLM outperforms previous approaches, and performs roughly as well as the state-of-the-art LLM today, GPT-4’: Apple throws down the gauntlet with new ReaLM AI on-device model

The first look into Apple’s AI plan has arrived

Laptop MagMadeline Ricchiuto

Will AI Be the Death of Higher Education? | National Review

Rather than destroying the humanities, AI applications can enable new discoveries.

National Review

‘AI education in schools needs cautious approach’

PETALING JAYA: The Education Ministry’s plan to teach the basics of artificial intelligence (AI) in schools from 2027 is commendable, said Malaysian M…

thesun.myQIRANA NABILLA MOHD RASHIDI

Prompt engineering - a primer

Review copy received from Samfundslitteratur (academic publisher)

Claus Nygaard has published the textbook "Prompt engineering - en grundbog", which addresses how to prompt the three largest language models: Gemini, ChatGPT and Copilot. In the book, Nygaard sets out 24 methods for writing effectively for the language models in order to target prompts and thereby achieve better results.

The book contains a brief introduction to language models, but I find the explanation lacks depth. In my view it becomes something of a positive cheerleading text that does not actually give the reader any relevant technological understanding—on the contrary, the section only scratches the surface, and it would have been better left out of the book.

That said, the book provides a solid introduction to prompting strategies, with detailed explanations of each method. These are then demonstrated with example outputs from the models. However, the book can lose some of its value here, as the rapid development of language models means that the sample outputs may quickly become outdated. Tomorrow the models will give different answers—and there will be a brand-new model. One quickly tires of reading the models’ trivial and mechanical outputs and ends up skipping ahead to what matters: why the prompting strategy produces a given output, and what users should watch out for. And for teachers, it is about seeing whether the exercises can be used in teaching. Here Nygaard does quite well, as it is easy to incorporate the exercises into lessons—and they can also be used with language models such as SkoleGPT or DanskGPT.

Over the past week, the book has been lying in the staff room at my upper secondary school (gymnasium), where it has sparked great interest; many have leafed through it and asked about its content. It is a book that teachers at our upper secondary school can relate to because it offers concrete, tangible material in Danish. In this way, the book is relevant for those who want to work pedagogically with language models in their teaching.

Get Prompt engineering by Claus Nygaard as an e-book in PDF format in Danish - 9788759344620

Get Prompt engineering by Claus Nygaard as an e-book in Danish - 9788759344620 - Books contain all sides of life. Read Listen Live among millions of books on Saxo.com.

SamfundslitteraturSamfundslitteratur

Scholarly articles

Below we select articles or tools that take a slightly more scholarly perspective. These are pieces we read ourselves to keep up to date, and we also know full well that they are not all that interesting to most people.

The impact of large language models on university students’ literacy development: a dialogue with Lea and Street’s academic literacies framework

Daniel W. J. Anson has published an article that examines the effects of large language models, such as ChatGPT, on university students’ literacy development. He argues that while language models have the potential to improve certain aspects of students’ academic writing and reading, they also carry risks for the development of essential academic skills.

Anson’s central argument is that excessive reliance on language models from the very start of university can impede students’ necessary cultural and cognitive development, which is crucial for succeeding at university.

Read the article here.

Machine learning uncovers the secrets of beer

In a study published in Nature Communications, researchers used machine learning to analyse flavour profiles in beer. By analysing the chemical compositions of over 250 beer types, the researchers developed a model that can predict the chemical compounds that contribute to specific flavour notes, and what consumers prefer. This comprehensive study provides clear indications of the future role of data and machine learning in product development within the food industry.

Read about the study below

Predicting and improving complex beer flavor through machine learning - Nature Communications

Perception and appreciation of food flavour depends on many factors, posing a challenge for effective prediction. Here, the authors combine extensive chemical and sensory analyses of 250 commercial Belgian beers to train machine learning models that enable flavour and consumer appreciation prediction.

NatureMichiel Schreurs

Brewing better Belgian beer with artificial intelligence

Machine learning tool can predict a beer’s taste and quality and which flavour compounds could improve it

Chemistry WorldRebecca Trager