"Broad AI" basics

New type of AI from Linz to outperform ChatGPT

The new AI from Linz should bring with it an "understanding of the world".

(Bild: fadi (computergeneriert))

A new form of artificial intelligence (AI) should be able to adapt flexibly to different situations, think logically and have a deeper understanding of the world than ChatGPT and co. This is what the Cluster of Excellence "Bilateral Artificial Intelligence", which was launched in Linz on Wednesday evening, is working on, explained project leader and AI pioneer Sepp Hochreiter from the University of Linz (JKU).

This is to be achieved by combining two important strands of research in the field of AI. "Sub-symbolic AI" is about machine learning, i.e. learning based on data, such as neural networks or ChatGPT. In the case of "symbolic AI", the focus is on how artificial systems deal with knowledge and conclusions. The aim is to combine the processing of large amounts of data and the use of learning methods that are modeled on the human brain.

Domestic expertise is pooled
The aim is to develop the foundations for so-called Broad AI, which has greater problem-solving capabilities and can deal with abstractions. By pooling domestic expertise in this area, the AI should be able to adapt flexibly to different challenges and environments. It could carry out complex planning while taking the context into account, for example organizing a trip or helping to renovate a house as cheaply and ecologically as possible.

It is foreseeable that other players and countries will also rely on a combination of sub-symbolic and symbolic AI, explained Hochreiter. In Germany, however, preliminary work has been underway for years to bring together the groups conducting research in this area, which should ensure a certain lead. Specifically, scientists from the Technical Universities (TU) in Vienna and Graz, the Institute of Science and Technology Austria (ISTA) in Klosterneuburg (Lower Austria), the University of Klagenfurt and the Vienna University of Economics and Business are working together.

Project manager and AI pioneer Sepp Hochreiter

(Bild: JKU)

Chatbots have no "world knowledge"
Current AI models are specialized in text creation, driving or image processing in medical diagnostics, said the head of the Institute for Machine Learning in the run-up to the launch event, which is taking place today at the University of Linz. They are very well suited for this, but are useless beyond this. According to Hochreiter, AI-supported chatbots based on so-called Large Language Models (LLM) cannot think logically and have no "world knowledge": "They have no sense that a tree exists in reality and not just as a sequence of letters."

They have no feeling that a tree exists in reality and not just as a sequence of letters.

KI-Pionier Sepp Hochreiter über Große Sprachmodelle

A self-driving car is unable to decide whether it can drive over a plastic bag that has blown over or whether it is an obstacle. "It needs to know that something is not so massive if it can be blown away by the wind. It still lacks this understanding of the world," said Hochreiter. Broad AI, on the other hand, should have sufficient knowledge to adapt to new situations. It would not be able to match the capabilities of humans; for that, it would need "artificial general intelligence", which is still a long way off.

DeepSeek swallowed up a lot of money
The Chinese start-up DeepSeek recently caused a stir. It presented a powerful AI model that was supposedly created at a fraction of the cost of competing US models. According to the expert, the company had actually invested a lot of money in advance in order to be able to present this low-cost model at the end: "We would have had the technology, but never the team or the computing resources to come up with and test such a model." There will soon be some groups, probably also in Europe, "who will copy this".

If you build small models for industry, I see opportunities again.

KI-Pionier Sepp Hochreiter

After the upscaling phase, in which ever larger models were built and more data was used, Hochreiter now sees a transition to the industrialization of AI: "Now smaller models that are specialized in certain tasks can also score points again. They don't need to know everything about Goethe or the Renaissance." As soon as AI finds its way into companies' production processes and logistics, there will be many opportunities for Europeans to get involved. "If you build small models for industry, I see opportunities again. And that will happen," Hochreiter is convinced.

As part of the "excellent=austria" initiative, the Austrian Science Fund FWF and the Ministry of Education have so far established nine clusters of excellence - on a wide variety of research topics. "Bilateral AI" is set to run for five years and is being funded with 33 million euros. The FWF contributes 60 percent, while the participating institutions account for 40 percent. Following a positive evaluation, the term can be extended to ten years with a total of EUR 70 million in funding.