More than 1,000 humans fail to beat AI contender in top crossword battle
Plus: Deepfake satellite images and Google fails to cite relevant research in its own large language model paper
In brief An AI system has bested nearly 1,300 human competitors in the annual American Crossword Puzzle Tournament to achieve the top score.
The computer, named Dr Fill, is the brainchild of computer scientist Matt Ginsberg, who designed its software to automatically fill out crosswords using a mixture of “good old-fashioned AI” and more modern machine-learning techniques, according to Slate.
It was able to solve multiple word conundrums fast with fewer errors than its opponents. Dr Fill, however, was not eligible for the $3,000 cash prize, which instead went to the best human player, a man named Tyler Hinman, who presumably isn't feeling somewhat redundant.
Ginsberg’s machine contained a computer running a 64-core CPU and two GPUs, and was trained on tons of text scraped from Wikipedia to learn words, and a database of crossword clues and their answers to parse the competition questions. You can watch it in action below.
Google defends large language models like the ones used by Google
In a new paper, researchers from Google and University California, Berkeley have outlined various ways to slash the environmental impact of the large amounts of energy consumed during the training of text-generation models like the ones used by Google.
Large language models are a particularly controversial area for The Chocolate Factory. The co-leads of its AI Ethics research group, Timnit Gebru and Margaret Mitchell, were ousted this year over a paper that detailed the power usage and financial costs of these models as well as concerns over their inscrutable nature.
Now, Google has published a counter-study. Large language models don’t have that big of a carbon footprint if they are trained using resources from data centers running efficiently in countries using renewable energy, the internet giant argued. You can read the whole thing here.
The paper coauthored by Mitchell, Gebru, computational linguistics professor Emily M. Bender, and others was shot down by Google for supposedly not including enough references to relevant research. What’s unfortunate here is that Google's latest paper failed to mention or reference Gebru and Bender’s paper in their study. One of the researchers later confirmed they were going to add a hat-tip to the pair in an updated version of their study.
Oops, thanks for pointing out. We had some text discussing Bender et al. in an earlier draft, but it looks like an editing pass to rework some text from a few weeks ago accidentally dropped that. We'll upload a new version soon with that discussion restored.— Quoc Le (@quocleix) April 22, 2021
Beware of deepfake satellite imagery
Academics are warning of the potential dangers of fake AI-generated satellite images.
A team of geographers led by the University of Washington in the US demonstrated how machine-learning algorithms could be trained to spit out fake geospatial images. The outputs could be used to disrupt applications relying on satellite imagery, such as Google Earth or even military software.
“This isn’t just Photoshopping things. It’s making data look uncannily realistic,” said Bo Zhao, assistant professor of geography at the UW and lead author of the study published in the journal Cartography and Geographic Information Science, this week. “The techniques are already there. We’re just trying to expose the possibility of using the same techniques, and of the need to develop a coping strategy for it.”
Zhao showed examples of how real images from cities could be manipulated by pasting on fake buildings to create made-up towns or adding false fires to mimic natural disasters. While it’ll take a lot more than deepfakes to attack real software systems, the researchers are raising awareness of it now in the hopes they can be one step ahead of the threat.
iGiant to create new jobs in AI
Apple pledged to invest $430bn in the US to employ 20,000 new staff focusing on emerging technologies, like AI to new chips, over the next five years. Apple plans to spend $1bn to launch a new campus in North Carolina too, with around 3,000 employees working on advanced research and development.
“At this moment of recovery and rebuilding, Apple is doubling down on our commitment to US innovation and manufacturing with a generational investment reaching communities across all 50 states,” Apple’s CEO. Tim Cook, announced this week.
“We’re creating jobs in cutting-edge fields — from 5G to silicon engineering to artificial intelligence — investing in the next generation of innovative new businesses, and in all our work, building toward a greener and more equitable future.”
SiFive, customer tape out AI chip on 14nm Samsung node
An AI accelerator system-on-chip developed in collaboration between SiFive and a mystery partner is set to be manufactured by chip Samsung.
Not much is known about the chip, except that it’s based on a 14nm FinFET design and contains 64-bit SiFive RISC-V cores as well as PCIe Gen 4 connectivity, two AI acceleration engines – one designed by the unnamed customer and one by Microsoft – and quad-channel LPDDR4 memory support.
SiFive didn’t reveal who the chip was for or when it would be sent off for mass production, just that it had taped out.
“Working in partnership with Samsung Foundry has accelerated SiFive’s ability to deliver our highly-efficient and configurable approach for SoC design and implementation,” Yunsup Lee, CTO of SiFive, said in a statement.
“We’re excited to continue to co-innovate with Samsung Foundry as we launch our latest SiFive Intelligence products to accelerate the development of next-generation AI SoCs with Samsung’s advanced process technology.” ®