• Christian Dallago, NVIDIA

Large Language Models (LLMs) are a type of neural network architecture that can be trained on massive amounts of text data to learn the patterns and relationships between words and phrases. They evolved from Recurrent Neural Networks (RNNs) to massively scalable Transformers, found in systems like ChatGPT. One of the key drivers of the success of Transformer architectures is the use of self-attention mechanisms to capture long-range dependencies between distant words. However, training LLMs is nontrivial, requiring massive amounts of data and large compute clusters. However, once trained, LLMs have many practical applications, including text classification, sentiment analysis, machine translation, question answering, and more. In the medical setting, LLMs can be used for clinical decision support systems, drug discovery, and patient monitoring. In this talk, we will briefly touch on the fundamentals of LLM architectures, training and applications.

• Alex Bateman

Large language models are one of the most exciting AI developments of recent times, but like many powerful tools they can be used in positive and negative ways. LLMs offer powerful language editing at no/low cost as well as highly accurate translation of technical text. Access to these tools can help level the playing field for scientific authors from around the globe. However, there has also been a growing trend of fake papers that are the output of so-called papermills. Current LLMs are able to generate extremely plausible scientific text based on minimal prompts. While completely AI generated papers are unlikely to make their way past a rigourous peer review, there remains a grey area where papers are partially written by AI. One of the recent findings is that the latest generation of AI are prone to hallucinating incorrect facts in a very confident tone. Checking facts in AI generated texts is extremely time consuming and as an Editor I fear that authors who are pressed for time will not be as careful as they should be.
To provide authors with some guidances on the use of LLMs the ISCB in association with Editors from Bioinformatics and Bioinformatics Advances have drafted an LLM policy (https://www.iscb.org/iscb-policy-statements/iscb-policy-for-acceptable-use-of-large-language-models). This policy outlines the permitted uses of LLMs in the publication process and highlights what is not permitted. This policy is likely to evolve as the tools improve and peoples level of acceptance changes.

• Patricia Palagi

AI is on everybody's lips these days. Large language models (LLMs) like ChatGPT are transforming how we interact with technology and use data. They seriously affect how we teach and learn - LLMs are game changers in this domain. In this talk, we will delve into the impact of LLMs on education, teaching and learning. We will start with ethical considerations such as bias, honesty and transparency and with a few words about some of the regulatory measures trying to mitigate these issues put into place so far by Higher Education Institutions. Additionally, we will explore some of the many challenges teachers face, including evaluating student progression and adapting learning outcomes to ensure effectiveness and achievability within this new context. Lastly, we will discuss the potential benefits of LLMs in education and training, such as personalized learning experiences for students and increased focus for teachers on critical educational aspects.

• David Leslie

The past few years have seen the rapid emergence of LLM applications that promise to accelerate scientific productivity and enhance the efficiency of research practice. These systems can carry out brute force text mining, “knowledge discovery”, and information synthesis and streamline scientific writing processes through the generation of literature reviews, article summaries, and academic papers. More recently, researchers have stitched together multiple transformer-based language models to create so-called “Intelligent Agents” which can perform “automated experiments” by searching the internet, combing through existing documentation, and planning and executing laboratory activities. In this talk, I explore some of the limitations of this new set of AI-supported affordances and reflect on their ethical implications for science as a community of practice. I argue that, amidst growing swells of magical thinking among scientists about the take-off of “artificial general intelligence” or the emergence of autonomous, Nobel Prize winning “AI scientists,” researchers need to take a conceptually sound, circumspect, and sober approach to understanding the limitations of these technologies. This involves understanding LLMs, in a deflationary way, as software-based sequence predictors. These systems produce outputs by drawing on the underlying statistical distribution of previously generated text rather than by accessing the evolving space of reasons, theories, and justifications in which warm-blooded scientific discovery takes place. Understanding this limitation, I claim, can empower scientists both to better recognise their own exceptional capacities for collaborative world-making, theorisation, interpretation, and truth and to better understand that contexts of scientific discovery are principal sites for human empowerment and for the expression of democratic agency and creativity.