Daniel Beck on data & AI in 2024 and beyond
"We should look at models like LLMs as the backbone piece of a much bigger puzzle."
I was lucky to meet Daniel through a friend some years ago. The AI/ML/NLP community really was that small until recently 😎. A senior lecturer at RMIT University in Melbourne, Daniel will tell you that they “love listening to other people’s problems and figuring out how we can leverage NLP, ML and AI to help solve them”. Daniel’s research ranges from natural language processing, into information retrieval and beyond into neuroscience, engineering design and AI ethics.
Daniel, 2024 has been another busy year for data & AI. What’s one development / milestone / news story that really caught your eye?
Not just one single story, but the sheer amount of stories reporting the use of AI-generated text in a range of applications.
You’ve been working in and around data & AI for a while now. Many things have changed! But tell us about something that was true when you started out in this space and is still important today.
That AI systems make errors. Successful deployment of those systems is not necessarily about having the highest accuracy, but how you create safeguards to gracefully recover from any errors your system *will* make.
It’s been a heady couple of years with 2024 almost as frothy as 2023. What's one common misconception about AI that you wish would go away?
My biggest pet peeve with the field is the idea that we can build a single, general purpose AI model to solve all tasks. Much of the work in the space of LLMs is targeted towards this goal, but my experience tells me you will always need to tailor models to end-task scenarios. This affects all aspects of AI system development, beyond simply “fine-tuning” but ensuring the collection of high-quality domain specific data and development of appropriate evaluation metrics as well.
Instead, we should look at models like LLMs as the backbone piece of a much bigger puzzle. I am sympathetic with the term “foundation models” in that regard. They are just that, a good foundation where you can build your system on, they are not the entire system.
The festive season is almost upon us, so many readers will have a bit of extra time to read / learn / reflect. Who do you follow to stay up to date with what’s changing in the world of data & AI?
No one =). There is too much noise and I don’t find it productive to try to keep track of any trends, beyond what I already get from following general news. These days, I find it better to “get back to my academic basics” and focus on doing (formal or informal) literature reviews when faced with a new challenge or opportunity.
But more importantly, I will be using the extra holiday time to go to the beach, play some video games, spend time with my partner and friends and hopefully do some outdoor climbing. The AI world can wait. =)
Leaning into your dystopian side for a moment, what’s your biggest fear for/with/from AI in 2025?
Not dystopian at all, but unfortunately very real: that AI systems will not just perpetuate but even accelerate the myriad of societal biases we currently experience. This is particularly relevant to my personal experience as a Queer Latinx person. There has been much recent work in this avenue but unfortunately I can only see this changing with governments stepping in and using this work to develop proper regulation standards for AI systems.
And now channeling your inner optimist, what’s one thing you hope to see for/with/from AI in 2025?
As an educator, I would like to see practical and fair ways to incorporate LLMs into our teaching. Every single one of my students is using these models in their studies and there is much talk about how this affects misconduct such as plagiarism. But my optimistic self tells me that students want to learn, so I would rather see us educators focus on how we can teach our students to use these systems in a critical and positive way.
You can follow Daniel on LinkedIn or perhaps take a course with them at RMIT