Bacterial warfare, self-programming language model, passive cooling in the big city

There’s a lot of science news in seven days, so just because a new study isn’t cited here Saturday morning doesn’t mean it hasn’t happened. AND lots of more happened. But also check out these four stories:

Bacteria use phages as weapons

Under agricultural conditions on cultivated land, one variant of the bacterium Pseudomonas viridiflava will spread and become the dominant microbe, but this will not happen on uncultivated land. Researchers from the University of Utah wanted to find out why, but early in their study they observed something so unexpectedly strange that it reoriented the entire focus of the study.

Studying the genomes of bacterial pathogens, they found that one sample had picked up a phage—a virus that attacks bacteria—and transformed it to kill its own bacterial competitors. Specifically, the bacteria acquired non-self-replicating clusters of altered phage called tylocins, which penetrate the outer membranes of pathogens and kill them.

Lead author Talia Backman speculates about the finding that tylocins could potentially lead to new antibiotics to address the antimicrobial resistance crisis: “While tylocins have previously been found in other bacterial genomes and studied under laboratory conditions, their impact and evolution in the wild. the bacterial population was unknown. The fact that we found that all of these wild plant pathogens have tylocins, and these tylocins evolve to kill neighboring bacteria, shows how important they can be in nature.

The language model is programmed to make sense

Large language models are capable of producing text. I was about to finish this thought with “that’s damn convincing” or something, but I had to stop and put my head on the table because the AI ​​hype cycle broke and completely disconnected from reality and I tired.

Existing LLMs are nothing more than text predictors without any knowledge of semantics or logic, so they can only help people generate text that contains zero percent symbolic reasoning, and they will now be on every device. Ok, I’m not here to complain about the AI ​​hype.

In an effort to improve LLM performance, MIT researchers have designed an innovative technique to perform natural language, math, and data analysis tasks by generating Python code. They call this approach embedded natural language programs and report 90% accuracy on a wide range of reasoning tasks.

The technique is a four-step process. In the first step, NLEP calls the packages needed to perform the task. In a second step, it imports natural language representations of the knowledge or data that the task requires. In the third step, the model generates a function that calculates the answer. And in the fourth step, the model generates the result in natural language. It is also more efficient for certain tasks, especially those in which the user has many similar questions; instead of generating a new Python program for each query, NLEP can generate one base program and change the variables for each query.

T-shirt cool

As summer heat domes settle over regions across the Northern Hemisphere, researchers at the UChicago Pritzker School of Molecular Engineering report a new wearable fabric that can protect city dwellers in the specific conditions of urban heat islands. Existing cooling fabrics work by diffusing direct sunlight. But sunlight is visible, while thermal radiation from building materials, pavements and infrastructure is infrared.

Engineers sought to create a textile with dual optical properties that could reflect both sources. And because it works passively, it can have cooling applications in areas with increasing energy consumption. In addition to heat-reflective clothing, this material has potential as a structural material to reduce internal temperature or as an insulator for automobiles. The researchers also suggest it could be used to transport perishable food, reducing the demand for active refrigeration systems.

Children immature, study finds

Researchers at the National University of Singapore report that a lower ratio of neural excitation (E) to neural inhibition (I) is a positive sign of brain maturation and found that children with lower E/I ratios perform better in school and on cognitive tests. . Previous studies have found that too much excitation or inhibition carries a higher risk of brain disorders including autism, Alzheimer’s disease and others. The study examined how E/I changes in youth by analyzing MRI brain scans and cognitive test scores in 885 children, adolescents, and young adults.

Remarkably, the team created a non-invasive technique for studying neural excitation and inhibition responses. In the first part of the experiment, subjects took the anti-anxiety drug Xanax or a placebo before an MRI brain scan – this allowed the researchers to find that Xanax increases neural inhibition, so the overall E/I ratio decreases. In the second part of the experiment, the team demonstrated the link between the E/I ratio and cognitive functions using cognitive tests. Participants with lower E/I outperformed those with higher ratios.