Top News

AI suggested 40,000 new possible chemical weapons in just six hours

A recent paper highlights the potential for the misuse of AI models used for drug discovery. Two types of ML models are typically used for drug discovery - a generative model that suggests candidate molecules and a toxicity model that predicts how toxic the molecules are. Usually, the generative model is used to suggest candidates that have low toxicity scores. Researchers in this paper demonstrated that by “flipping the sign” and leading the generative model to follow higher toxicity scores, they were able to obtain 40k lethal molecule candidates in less than 6 hours, some of which are predicted to be more toxic than VX, the most potent nerve agent ever developed.

Our take: The researchers note that there are likely false positives in the predicted molecules, and that actually synthesizing them would be very difficult since their ingredients are similar to common chemical warfare agents, which are carefully monitored and regulated. Still, this research highlights the dual use of powerful and accessible AI models (someone who knows what they’re doing can probably reproduce the paper just from open-source tools and datasets), and further discussion and care are needed across many AI application communities to ensure ethical use.

Deepfake Zelenskyy surrender video is the 'first intentionally used' in Ukraine war

Mikael Thalen @MikaelThalen

A deepfake of Ukrainian President Volodymyr Zelensky calling on his soldiers to lay down their weapons was reportedly uploaded to a hacked Ukrainian news website today, per @Shayan86

A Deepfake video of Ukrainian President Volodymyr Zelenskyy encouraging Ukranians to "lay down arms" was posted online last week and received a large number of views. Zelenkskyy quickly released a video on his own Instagram page debunking this misinforomation. Seemingly in response, a deepfake video of Russian President Vladimir Putin telling Russian soldiers to "go home while you're alive". The head of security policy at Meta confirmed that the company has removed the deepfake video of President Zelenskyy.

Our take: The potential of deepfakes to be used for disinformation has been discussed for years, and this appears to be the first major instance of that happening. Other forms of disinformation on social media are far more prevalent and problematic, but this example does demonstrate how deepfakes could become a potent tool for disinformation in the future, especially as the algorithms for it continue to improve and become easier to use.

Meta AI’s Sparse All-MLP Model Doubles Training Efficiency Compared to Transformers

Transformers have become the dominant model paradigm for language over the past few years, and are quickly making their way into vision as well. But recent papers have proposed MLP architectures that boast competitive performance and offer efficiency benefits. In their paper Efficient Language Modeling with Sparse all-MLP, a team from Meta AI and SUNY Buffalo propose sMLP, a sparsely-activated all-MLP architecture that is up to 2x more efficient than recent methods. The team also introduces novel modules and routing strategies that increase sMLP's model capacity while keeping compute constant. The study claims sMLP outperforms SOTA sparse transformer-based MoE models in generalization and achieves better accuracy in zero-shot in-context learning.

Our take: It is exciting to see some pushback against transformers in certain ML domains. While expressive and powerful, transformers are incredibly expensive architectures to train. I think we need to start exploring alternative strategies to simply scaling architectures based on transformer architectures to massive sizes, and sMLP is one development along that path. I hope we will see more researchers propose and investigate architectures that can achieve the performance and expressiveness of transformers while saving on compute.

Other News

Research

What do psychedelic drugs do to our brains? AI could help us find out - "The words people used to describe their trip experiences could lead to better drugs to treat mental illness."

Lessons Learned on Language Model Safety and Misuse - "The deployment of powerful AI systems has enriched our understanding of safety and misuse far more than would have been possible through research alone."

When it comes to AI, can we ditch the datasets? - "A machine-learning model for image classification that’s trained using synthetic data can rival one trained on the real thing, a study shows."

Researchers From Italy Use Machine Learning To Distinguish Different Stages And Severity Of Parkinson’s Disease By Voice - "Parkinson’s disease (PD) is a neurological condition that causes tremors, stiffness, and difficulty walking, balancing, and coordinating. "

Researchers from Northwestern University Come Up With More Efficient AI Training With a Systolic Neural CPU - "Northwestern University researchers present a new design that combines a traditional CPU with a systolic convolutional neural network (CNN) accelerator on a single core, resulting in a highly programmable and versatile unified design."

HPC-AI’s FastFold Shortens AlphaFold Training Time from 11 Days to 67 Hours - "It significantly reduces the time and economic costs of AlphaFold model training and inference by applying large model training techniques such as parallelism strategies and communication optimization."

Idiap Research Institute Proposes HyperMixer: A Competitive MLP-based Green AI Alternative to Transformers - "a team from Switzerland’s Idiap Research Institute proposes a novel multi-layer perceptron (MLP) model, HyperMixer, as an energy-efficient alternative to transformers that retains similar inductive biases."

Applications

Building games and apps entirely through natural language using OpenAI’s code-davinci model - "TL;DR: OpenAI has a new code generating model that’s improved in a number of ways and can handle nearly two times as text (4,000 tokens.) I built several small games and applications without touching a single line of code."

Machine Learning Reimagines the Building Blocks of Computing - "Traditional algorithms power complicated computational tools like machine learning. A new approach, called algorithms with predictions, uses the power of machine learning to improve algorithms."

Chrome Extension Can Detect Fake Profile Pictures with 99.29% Accuracy - "V7 Labs has created a new artificial intelligence-based (AI) software that works as a Google Chrome extension that is capable of detecting artificially generated profile pictures — like the ones above — with a claimed 99.28% accuracy."

Microsoft’s AI tech helps blind kids to make friends, build confidence - "Tech giant Microsoft has developed a novel Artificial Intelligence-based technology that helps blind children and young people better understand their immediate social environment, interact with peers and gain confidence more easily."

You.com launches an AI-powered writing tool powered by OpenAI - "You.com’s new tool is powered by the same technology behind OpenAI’s GPT-3, an AI language system that can generate human-like poetry, emails, recipes, short stories, movie scripts, and more. "

Meet ElliQ, the robot who wants to keep Grandma company - "A new AI device optimized for empathy will either solve the problem of senior isolation or fling us into a dystopic abyss"

Business

Miso Robotics partners with Chipotle for tortilla chip-making robots - "After years of frying and flipping burgers at White Castle locations, Miso Robotics’ autonomous kitchen assistant has a new gig: making tortilla chips for Chipotle."

Bear Robotics Announces $81M Series B as Restaurants Struggle to Meet Dining Demand - "the company stated that these robots have made more than 28 million deliveries"

Dive deep into Snorkel.AI’s pitch technique that raised $135m - "Snorkel.AI’s co-founder, and CEO, Alex Ratner, honed his storytelling chops at a Stanford lab, and that ability, he said on a recent TechCrunch Live episode, was key to helping him raise $135.3 million over four rounds since the company was founded in 2019."

AI critical to business success in post-pandemic world, study finds - "Over three-quarters (77%) of knowledge workers said that all three could benefit their company in the next 36 months"

‘No-Code’ Brings the Power of A.I. to the Masses - "A growing number of new products allow anyone to apply artificial intelligence without having to write a line of computer code. Proponents believe the “no-code” movement will change the world."

AI Startup Aibee Becomes Unicorn With Fresh Funds From Xiaomi - "The startup is also receiving new funds from China Merchants Capital Management Co., according to a press release posted on the company’s official WeChat account. The press release didn’t give details of the size of their investment. "

Spending on Artificial Intelligence Solutions Will Double in the United States by 2025, According to a New IDC Spending Guide - "– Spending on artificial intelligence in the United States will grow to $120 billion by 2025, representing a compound annual growth rate (CAGR) of 26.0% over the 2021-2025 forecast period. "

Concerns

The AI Placed You at the Crime Scene, but You Weren’t There - "Like a lot of tech solutions to complex problems, facial recognition algorithms aren't perfect. But when the technology is used to identify suspects in criminal cases, those flaws in the system can have catastrophic, life-changing consequences."

AI and disinformation in the Russia-Ukraine war - "Recently recommendation algorithms have led to the promotion of disinformation on social media about the Russia-Ukraine war. Particularly on TikTok, disinformation and misinformation are rife."

There’s More to AI Bias Than Biased Data, NIST Report Highlights - "Bias in AI systems is often seen as a technical problem, but the NIST report acknowledges that a great deal of AI bias stems from human biases and systemic, institutional biases as well. "

Ransomware Hackers Used AI Images, Microsoft Flaw in Campaign - "A group of ransomware hackers used a variety of techniques to try breaching hundreds of companies last year, exploiting a vulnerability in Microsoft Corp.’s Windows and using artificial intelligence technology to create fake LinkedIn profiles, Alphabet Inc.’s Google found. "

Bias in Automated Speaker Recognition - "AI enabled voice biometrics are a hidden and prevalent form of authentication. This paper examines sources of bias in the development and evaluation practices of voice-based identification systems."

Analysis

The 2022 AI Index: Industrialization of AI and Mounting Ethical Concerns - ""2021 was the year that AI went from an emerging technology to a mature technology—we're no longer dealing with a speculative part of scientific research, but instead something that has real-world impact, both positive and negative," says Jack Clark, co-chair of the AI Index."

California, Texas Continue to Dominate Artificial Intelligence (A.I.) Jobs - "When it comes to plentiful jobs in artificial intelligence (A.I.) and machine learning, the largest states continue to dominate, according to CompTIA’s latest tech jobs report. California and Texas enjoyed thousands of open A.I. jobs in February, with notable month-over-month increases."

Deep learning isn’t living up to the hype, but still shows promise - "We’ve been overhyping deep learning for too long. It’s time to start embracing it as a complement to, not replacement for, human ingenuity."

Artificial intelligence is everywhere now. This report shows how we got here. - "Artificial intelligence is getting cheaper, better at the tasks we assign it, and more widespread—but concerns over bias, ethics, and regulatory oversight still remain."

How Zillow's homebuying scheme lost $881 million - "Zillow made news last week as it reported a loss of $881 million on its house-buying business last year⁠. That’s especially remarkable because house prices rose dramatically for much of the year."

Why changing computing trends across different ML eras matter - "the history of compute in machine learning can be categorised into three eras – the Pre Deep Learning Era, the Deep Learning Era and the Large-Scale Era."

Here's how Americans view facial recognition and driverless cars - "In recent years, the proliferation of artificial intelligence has given the world technology like Siri, Netflix recommendations and chat customer support. But a new survey shows Americans are still torn about how it may continue to impact society."

Policy

Powered by artificial intelligence, ‘autonomous’ border towers test Democrats’ support for surveillance technology - "U.S. Customs and Border Protection has deployed about 175 surveillance towers along the southern border so far. But it is making some Democrats uneasy."

Pentagon names Navy’s Palmieri as deputy for new digital and AI office - "Her deep knowledge and relationships within DoD will ensure the CDAO navigates and cuts through the bureaucracy to get things done"

Congress wants new $200 million program to strengthen AI at combatant commands - " U.S. lawmakers this month pitched a new program to increase adoption of artificial intelligence across combatant commands, one of the latest efforts to boost a technology that officials say provides an edge over world competitors like Russia and China."

Fun

Magical Utopia - We've got a wizard living in the shed, who occasionally comes out to do magic and eat cheese sandwiches.

AI-generated utopias - "AI isn't known for being able to solve the big problems, but what about the VERY large problems, such as possible futures to strive for? I decided to find out if I could get GPT-3 to come up with new ideas for utopias."

The 6 Best Fiction Books for the AI Fanatic - "You.com’s new tool is powered by the same technology behind OpenAI’s GPT-3, an AI language system that can generate human-like poetry, emails, recipes, short stories, movie scripts, and more. "

Copyright © 2022 Skynet Today, All rights reserved.

Last Week in AI

Last Week in AI #159: AI discovers lethal…