As my work focuses on assessing the environmental impact of AI and digitalisation more broadly, I was particularly drawn to this week’s ‘Twosday’ Talks, which featured presentations by Dr Rachel Parkinson and Dr Heloise Stevance. Both offered insights that could expand my understanding of the wider impacts and uses of AI beyond environmental ones. This is important, as not all impacts of AI can be measured in environmental terms.
This Tuesday Talk began with Dr Rachel Parkinson's presentation entitled Rethinking Risk: using AI and multimodal data to predict pesticide impacts on bees. Rachel provided a compact background on the crucial role that bees play in ecosystems and agriculture, guiding us through the work under the ‘BEEhaviour Lab’. Rachel showed the fruits and vegetables that either depend on or benefit significantly from bee activity (and, although somewhat irrelevant, her talk made me even more excited about the ‘Plan Bee’ volunteer program I recently joined with the aim of monitoring and maintaining solitary bee hotels).
Rachel’s research addresses one of the many factors contributing to global pollinator declines: the sublethal effects of pesticides. Using sensors and AI, her team analyses bee behaviour by recording their movements and sounds (though the latter wasn’t covered in this talk). Her findings have identified specific behavioural markers - increased large turns, decreased climbing behaviour, stillness, or sluggishness - that signal toxicity in bees. Additionally, the BEEhaviour Lab approach has potential applications for studying other pollinating insects.
The second talk, delivered after a shared family-style main course, was by Dr Heloise Stevance on How Can A.I. Help Us Find Exploding Stars and Hungry Black Holes? Heloise captivated us with vibrant images of the cosmos and explained why tracking cosmic explosions is vital, with reasons ranging from understanding the origins of chemical elements to probing the physics of black holes and the universe's expansion.
She explained how the Virtual Research Assistant significantly reduced their workload; however, she also emphasised the limitations of AI in some of the tasks. For example, AI often overlooks rare or unusual events as it’s trained on common patterns, yet this area is where Heloise’s interests lie. Other issues are related to the fact that:
- Its probabilistic models give different outputs despite the same input.
- In general, AI’s Large Language Models (LLMs), such as ChatGPT and Gemini, aren’t trained to be truthful, but to sound human.
Heloise concluded the talk with a reminder that just because a technology is new, it doesn’t mean it’s always the right tool.
The Twosday evening wrapped up with a delicious mango mousse. Although the audience wasn't the largest, the discussions and Q&A were rich and engaging. I especially appreciated how both speakers communicated their work in an accessible and lively manner.
