Big technology firms are spending billions of dollars to build artificial intelligence models that even their creators struggle to explain.
Peering inside that “black box” will be the focus when academics and students from across New Zealand meet at St Margaret’s College in Dunedin on Monday and Tuesday for the Artificial Intelligence Researchers Association’s annual conference.
It is the first time the two-day event has been held in the city.
University of Otago computer science senior lecturer and conference chairman Dr Lech Szymanski said major overseas companies had turned the field into a race driven by huge datasets and computing power.
“There’s people who have these instincts, these engineers that can really create amazing systems, but they fundamentally can’t explain how they work,” Dr Szymanski said.
A technology firm could use a model to solve a batch of mathematics problems, with the programme jumping straight to the answers without showing the mechanics of how it got there.
When hard equations were solved by hand, the prize was the new insight built along the way. Mathematicians had little interest in a verified solution that came with no proof.
Scientists still did not know whether AI programmes could crack complex challenges, or if they were just picking off easy targets that humans had overlooked.
Trying to settle those basic questions was frustrating because the sector kept advancing so quickly.
Models used last year were already obsolete, replaced by more complicated agentic set-ups where multiple AIs collaborated.
“I don’t think we’re going to find out that AI doesn’t work.
“It works, but it might not work in the way that we thought, and therefore our understanding of it might change.”
Debates over whether large language models possessed consciousness were irrelevant to him.
“It’s a tool, and that tool is useful for some things, and I’m almost certainly sure we use it incorrectly in some aspects, and in some aspects we’re probably using it correctly.
“We just don’t know yet.”
Postgraduates and staff at the university were investigating what happened inside AI programmes and how to improve them.
Otago projects at next week’s gathering include teaching a four-legged robotic dog to walk across tricky terrain using reinforcement learning, building recommender tools to predict consumer choices, testing whether chatbots can be trusted and training models to read images in finer detail.
Other local researchers will present findings on predicting wildfire risk, cutting machine learning energy use and tracking sleep disorders with brain signals.
“I would say we are doing good work.
“We have students at the top conferences.”