LLMs – Expert Group Meeting
21.8.2023

Expert Group Meeting on LLMs on 21.8.2023

The Expert Meeting on Monday, 21 August 2023 focussed on various aspects and applications of Large Language Models. 45 participants attended, most of which were on-site.

The following presentations were held at the meeting:

Kim Engels, Converto AG
Large Language Models for Cross-Media Marketing

In this talk, Kim presented how Converto AG use AI and LLMs to improve or speed up their projects.

They use text generation for newsletters, as well as code generation and self-developed solutions to create 3D models for customer campaigns. 

All of these approaches have in common that they can save a lot of time but none of them can be used without revision by a human.

Florian Tramér, ETH
Are Aligned Neural Networks Adversarially Aligned?

Florian discussed the vulnerability of large models to adversarial attacks, which can circumvent their alignment (the tuning that makes them “helpful and harmless”). He showed examples of how image and multimodal text-image models can be tricked into misclassification and misbehaviour. Text-only models used to be relatively robust, but a recent paper found that attacks can be tuned on open-source models and then copy-pasted to closed-source models, where they also work.

Alex Paramythis, Contexity AG
Adapting Large Language Models for Customer Request Handling

Alex’s talk showed that using open-source LLMs for accessing internal or private data is far from trivial. He and his collaborators tried to use a variety of models out of the box and also experimented with fine-tuning them on a variety of data sources. Their evaluations showed that currently none of the open-source models can rival OpenAI’s models in terms of performance and that fine-tuning is not always leading to a performance improvement.

---