How OpenAI stress-tests its large language models OpenAI has lifted the lid (just a crack) on its safety-testing processes. It has put out two papers describing how it stress-tests its powerful large language models to try to identify potential harmful or otherwise unwanted behavior, an approach known as red-teaming.
The first paper describes how OpenAI directs an extensive network of human testers outside the company to vet the behavior of its models before they are released. The second presents a new way to automate parts of the testing process, using a large language model like GPT-4 to come up with novel ways to bypass its own guardrails. MIT Technology Review got an exclusive preview of the work.
—Will Douglas Heaven
Who should get a uterus transplant? Experts aren’t sure. Over 135 uterus transplants have been performed globally in the last decade, resulting in the births of over 50 healthy babies. The surgery has had profound consequences for these families—the recipients would not have been able to experience pregnancy any other way.
But legal and ethical questions continue to surround the procedure, which is still considered experimental. Who should be offered a uterus transplant? Could the procedure ever be offered to transgender women? And if so, who should pay for these surgeries? Read the full story.
—Jessica Hamzelou
This story is from The Checkup, our weekly newsletter about the latest in biotech and health. Sign up to receive it in your inbox every Thursday.
The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 OpenAI may launch a web browser
Which would be a full-frontal assault on Google (The Information $)
+ The Google browser break-up is an answer in search of a question. (FT $)
+ OpenAI accidentally deleted potential evidence in a training data lawsuit. (The Verge)
2 Border militias are ready to help with Trump’s deportation plans
Regardless of whether they’re asked to or not. (Wired $)
+ Trump’s administration plans to radically curb the powers of the federal agency that protects unions. (WP $)
3 Russia hit Ukraine with a new type of missile
Here’s what we know about it so far. (The Guardian)
4 Microsoft is about to turn 50
And it’s every bit as relevant and powerful as it’s ever been. (Wired $)
5 China has overtaken Germany in industrial robot adoption
South Korea, however, remains streets ahead of both of them. (Reuters $)
+ Three reasons robots are about to become way more useful. (MIT Technology Review)
6 The irresistible rise of cozy tech
Our devices, social media and now AI are encouraging us to keep looking inward. (New Yorker $)
+ Inside the cozy but creepy world of VR sleep rooms. (MIT Technology Review)
7 Churchgoers in a Swiss city have been spilling their secrets to AI Jesus 😇
And they’re mostly really enjoying it. Watch out, priests. (The Guardian)
8 A French startup wants to make fuel out of thin air
Then use it to fuel ships and airplanes. (IEEE Spectrum)
+ Everything you need to know about alternative jet fuels. (MIT Technology Review)
9 WhatsApp is going to start transcribing voice messages
This seems a good compromise to bridge people’s different communication preferences. (The Verge)
10 Want a new phone? You should consider second-hand
It’s better for the planet—and your wallet. (Vox)
Quote of the day
“Nope. 100% not true.”
—Jeff Bezos fires back at Elon Musk’s claim that he was telling everyone that Trump would lose pre-election in a rare post on X.
The big story
This chemist is reimagining the discovery of materials using AI and automation
DEREK SHAPTON
October 2021
Alán Aspuru-Guzik, a Mexico City–born, Toronto-based chemist, has devoted much of his life to contemplating worst-case scenarios. What if climate change proceeds as expected, or gets significantly worse? Could we quickly come up with the materials we’ll need to cheaply capture carbon, or make batteries from something other than costly lithium?
Materials discovery—the science of creating and developing useful new substances—often moves at a frustratingly slow pace. The typical trial-and-error approach takes an average of two decades, making it too expensive and risky for most companies to pursue.
Aspuru-Guzik’s objective—which he shares with a growing number of computer-savvy chemists—is to shrink that interval to a matter of months or years. And advances in AI, robotics, and computing are bringing new life to his vision. Read the full story.
Source : Technology Review
Author : News7
Publish date : 2024-11-22 13:10:00
Copyright for syndicated content belongs to the linked Source.