In brief AI proto-boffins can now rent up to 16 GPUs, the largest amount on any single virtual machine currently available, via Google Cloud.
"We are excited to announce the general availability of A2 VMs based on the NVIDIA Ampere A100 Tensor Core GPUs in Compute Engine," the company said.
Devs can use the A2 VMs in various configurations, and scale across 1, 2, 4, 8, and 16 GPUs.
With 16 GPUs, the cluster has an interconnect speed of 9.6TB/s using Nvidia's NVlink, and up to 3TB of local SSD memory. Having more GPUs on a single machine will benefit those training particularly large AI models.
These instances are available to customers in the us-central1, asia-southeast1, and europe-west4 zones at the moment, and cost from $0.87 per hour per GPU on Google Cloud's preemptible service, depending on local demand.
You can see the full pricing list here.
The latest autonomous car startup merger
Cruise, the self-driving startup backed by General Motors, has snapped up one of its smaller rivals, Voyage, for an undisclosed fee.
Voyage, led by Oliver Cameron, originally set its sights on building a fleet of autonomous vehicles that were trained on a set of predetermined routes to shuttle senior citizens to and fro. Now it has joined GM-backed Cruise, with a loftier goal of developing self-driving taxis that can drive just about anywhere, come rain or shine.
As the industry progresses, the bigger startups are swallowing up the smaller ones in the race to build the world's first L5 autonomy robocar. "At Voyage, we've always believed that to lead this industry requires initial commercialization on either calm roadway or complex roadway, with no middle-ground to be found," Cameron said in a statement.
He believes Cruise is the best bet and is obviously more than happy to join. "Voyage's experience and development of Commander (our self-driving AI), Shield (our collision mitigation system), and Telessist (our novel remote assistance solution) will only supercharge Cruise's goal of superhuman driving performance."
How to get around issues of consent in computer vision databases? Blur out the faces
The creators of the ImageNet dataset commonly used to train computer vision algorithms have released an updated version, where the faces in the photographs have been purposely blurred out.
ImageNet came under fire for hosting inappropriate images of nude children, porn actresses, and more. Millions of images were scraped from the web and labelled without much consideration for people's privacy. In an attempt to solve this problem, researchers from Princeton and Stanford University have scrubbed the dataset and blurred any faces that appear in it.
The team used Amazon's Rekognition to identify and obscure the faces then paid Mechanical Turk workers to go through the images and confirm whether each one had been blurred correctly.
Although that may solve the privacy issue, it could introduce all sorts of practical problems if a model trained on the dataset is deployed in the real world. "One important problem to consider is what happens when you deploy a model that was trained on a face-blurred data set," Olga Russakovsky, co-author of the research and an assistant professor at Princeton University, told Wired.
Ultimate catfish using AI computer vision FaceApp
Netizens were shocked when they discovered a popular social media account featuring numerous snaps of a young woman riding a motorcycle around Japan was actually run by a 50-year-old man disguising his face with an AI image filter on FaceApp.
The Twitter account @azusagakuyuki has attracted over 20,000 followers; the user appears to be a pretty female biker with luscious orange hair. But internet sleuths discovered one strange photo, which shows her reflection looks nothing like she does in her selfies.
The mystery caught the attention of Japanese journalists, who eventually tracked her down only to realize she was not a young woman after all, but a middle-aged man duping the internet using FaceApp. He admitted that he wanted to look like a "beautiful woman" on social media so his photos would be more popular, and said no one wants to see an "uncle", according to Singaporean news site Mothership.
FaceApp is powered by generative adversarial networks, which are often used to paste fake faces onto real bodies. The technology has progressed so much that things like deepfakes are now getting scarily convincing. ®