Chicxulub meteorite found guilty; Good and bad news for LLMs

It’s the last week before Christmas and not a creature is moving, not even a mouse model bred to exhibit the characteristics of ADHD in in vivo studies of central nervous system stimulants. This week we reported on the discovery of the world’s oldest known mammalian ancestor; the study of 1,300 burial mounds in western Azerbaijan; and a quasiparticle present in all magnetic materials. Additionally, researchers have put to rest another alternative theory about the end of the dinosaurs; LLMs are more like a brain, according to one study, but another suggests they show signs of cognitive decline:

Rock, paper, volcano

We all know that a massive meteorite impact in the Gulf of Mexico killed off the dinosaurs 66 million years ago. But what some researchers assume is… maybe they didn’t? In fact, geologists have hypothesized that massive volcanic eruptions on the Indian peninsula could have expelled enough ash and CO₂ in the atmosphere to cool the planet to the point of making it uninhabitable.

But climate scientists at Utrecht University have now ruled out this scenario, finding that although roughly contemporary volcanoes had a measurable impact on climate, these impacts were different in terms of the degree, type and scale of time compared to a meteorite impact.

The researchers analyzed soil particles in ancient peatlands to reconstruct air temperature over the entire period covering both volcanoes and meteorite impacts. They found that the volcanoes occurred 30,000 years before the impact and corresponded to a global atmospheric cooling of 5 degrees Celsius, likely due to the input of atmospheric sulfur. About 10,000 years later, temperatures were rising, likely aided by volcanic CO.₂ emissions.

“These volcanic eruptions and the associated CO₂ and the release of sulfur would have had dramatic consequences on life on Earth. But these events occurred millennia before the meteorite impact and probably played only a small role in the extinction of the dinosaurs,” explains Lauren O’Connor from Utrecht University.

12 brains agree

Researchers at Columbia University and the Feinstein Institutes for Medical Research Northwell Health report that as LLM models advance in sophistication and functionality, they also become more brain-like. By creating a sort of book club for 12 recent LLMs and training them on a single text, the researchers extracted the text’s internal representations, called “embeddings,” and compared them to the results of a human experiment.

In the human experiment, they recorded the neural responses of neurosurgical patients while they listened to the text. Next, the researchers attempted to predict recorded neural responses to words from word embeddings derived from LLMs.

“The ability to predict brain responses from word embeddings gives us an idea of ​​how similar the two are,” says Gavin Mischler, first author of the paper.

Using computational tools, they attempted to assess which LLM layers most closely aligned with human neural responses, particularly those in brain regions associated with language processing. They conclude that as LLMs become more powerful, their integrations become more similar to the brain’s responses to language.

However

While it is great that LLMs now have their own researcher-run journal clubs, a report in the Christmas issue of British Medical Journal finds that almost all major LLMs show signs of mild cognitive impairment in standard dementia diagnostic tests. According to the researchers, this is the first study of cognitive decline in large language models.

The Montreal Cognitive Assessment Test includes short tasks and questions to assess attention, memory, language and spatial skills; it is widely used to detect cognitive impairment and early signs of dementia. The instructions given by the researchers to the LLMs were identical to those given to human patients; the LLMs were noted and the results were evaluated by a neurologist. They report that all chatbots showed poor performance on visuospatial skills, executive tasks, and clock drawing test. Models tended to perform better in naming, attention, language, and abstraction.

In a final leap of science fiction conjecture, the authors write: “Not only are neurologists unlikely to be replaced by large language models anytime soon, but our results suggest that they may soon find themselves dealing with new virtual patients – artificial intelligence models. presenting with cognitive disorders. »

© 2024 Science X Network