Yes, AI Models Can Get Worse over Time

Trouble viewing? View in your browser. View all Scientific American publications.

    

August 08, 2023

Hello Readers, I'm covering for Sophie this week. Three computer scientists testing the AI GPT-4 noticed a decline in its abilities between March and June, with no explanation from OpenAI, the company that developed GPT. Of course, the continued development of this technology could be at the heart of its problems; making changes to the code or adding training data with one outcome in mind carries the potential for ripple effects elsewhere, say experts. Ready more in this week's top story. Enjoy!

Andrea Gawrylewski, Chief Newsletter Editor

Artificial Intelligence Yes, AI Models Can Get Worse over Time More training and more data can have unintended consequences for machine-learning models such as GPT-4 By Lauren Leffer

Weather Why Homes Often Feel Warmer Than the Thermostat Suggests The reason some homes feel hotter than their air temperature has to do with radiant heat. Here's what to do about it By Jonathan Bean,The Conversation US

Economics Social Media Is Rewriting the Banking Playbook Our financial systems need to be overhauled to avoid the kind of social-media-fueled bank run that killed Silicon Valley Bank By Alex Lipton,Alex Pentland

Artificial Intelligence We Need Smart Intellectual Property Laws for Artificial Intelligence "One-size-fits-all" regulation will sideline medical and research benefits promised by the advent of artificial intelligence By James Love

Artificial Intelligence Tech Companies' New Favorite Solution for the AI Content Crisis Isn't Enough From college plagiarism to cybercrime scams, generative AI is eroding trust in online content. Digital watermarking is no quick fix for the problem By Lauren Leffer

Climate Change A Sun Shield over Earth? Catch an Asteroid, and It Might Work A resurfaced idea for solar geoengineering imagines a sunlight-blocking space shield tethered to an asteroid By Chelsea Harvey,E&E News

Defense Meet the Woman Who Supervised the Computations That Proved an Atomic Bomb Would Work Naomi Livesay worked on computations that formed the mathematical basis for implosion simulations. Despite her crucial role on the project, she is rarely mentioned as more than a footnote—until now By Katie Hafner,The Lost Women of Science Initiative

Astronomy Can You Decode an Alien Message? An artist, a programmer and a scientist have created a simulation of extraterrestrial communication to test Earthlings' ability to understand it By Shi En Kim

Renewable Energy Largest U.S. Offshore Wind Project Could Produce Power This Year Construction has begun on the Vineyard Wind project off the coast of Massachusetts. It could provide enough clean energy to power 400,000 homes per year By Benjamin Storrow,E&E News

QUOTE OF THE DAY "You'll begin to trust a certain kind of [GPT] behavior, and then the behavior changes without you knowing. From there your whole application that you built on top starts misbehaving." Vishal Misra, computer science professor at Columbia University

FROM THE ARCHIVE AI-Generated Data Can Poison Future AI Models As AI-generated content fills the Internet, it's corrupting the training data for models to come. What happens when AI eats itself? By Rahul Rao | July 2023

LATEST ISSUES

Questions?   Comments?

Download the Scientific American App

To view this email as a web page, go here. You received this email because you opted-in to receive email from Scientific American. To ensure delivery please add newsletters@scientificamerican.com to your address book. Unsubscribe     Manage Email Preferences     Privacy Policy     Contact Us

Scientific American

1 New York Plaza, FDR Dr, Floor 46, New york, NY 10004

Unsubscribe - Unsubscribe Preferences