What we’ve been reading
Training a single AI model can emit as much carbon as five cars in their lifetimes
MIT Tech Review
Training an AI model has the equivalent carbon footprint as five American cars, including fuel usage, according to researchers at the University of Massachusetts, who performed life cycle assessments for training several large AI models. While that figure relates to a neural net with more than 200 million parameters, the study highlights the unbelievable efficiency of the human brain. The bigger question now is whether we will build machines that rival the brain for efficiency.
To be energy-efficient, brains predict their perceptions
Quanta Magazine
Many neuroscientists view the brain as a “prediction machine” which, through predictive processing, uses knowledge of the world to make inferences or generate hypotheses about the causes of incoming information.
Computational neuroscientists are building artificial neural networks that learn to make predictions about incoming information. The studies show that these AI models have abilities that seem to mimic real brains, with some supporting the hypothesis that brains evolved as prediction machines to satisfy energy constraints.
New AI test diagnoses glaucoma in just 10 seconds
IEEE Spectrum
A team of engineers and ophthalmologists have developed a machine learning model which can diagnose glaucoma in just 10 seconds. “Existing AI glaucoma tests require the patient to be perfectly still for up to 10 minutes. Our tech does the job in 10 seconds, without compromising on accuracy.”
Algorithms mimic the process of biological evolution to learn efficiently
Tech Xplore
Researchers have developed new evolutionary algorithms that mimic the process of natural selection. By subjecting the evolutionary algorithms to three learning scenarios, the researchers found that the algorithms were able to discover mechanisms of synaptic plasticity and thereby successfully solve a new task. “A promising approach to gain deep insights into biological learning principles and accelerate progress towards powerful artificial learning machines,” say the researchers who hope this may pave the way to truly intelligent, adaptive machines.
OpenAI’s API now available with no waitlist
OpenAI
OpenAI has removed the waitlist of GPT-3, their large language model. GPT-3 was first made available through private beta in 2020 and has since been used in more than 300 different apps by “tens of thousands” of developers — producing 4.5 billion words per day.
If you’d like to receive this newsletter in your inbox, you can subscribe here
