

Discover more from Last Week in AI
Last Week in AI #165: AI helps model volcanoes, Anthropic gets $580M for more explainable AI, AI algorithms that screen for child neglect, and more!
New AI algorithm models the insides of volcanoes, ex-OpenAI engineers to work on more explainable large language models, unclear how AI that screens for child negligence reports can be beneficial
Top News
Machine Learning Helps See into a Volcano’s Depths
Predicting volcanic eruptions and studying their impact continues to excite scientists. Recently, one group, while studying volcanic eruption patterns, improved our understanding of the “plumbing system” underneath a volcano using a new unsupervised machine learning algorithm. Unlike prior visualizations, this algorithm uses thermodynamic modeling to detect populations of crystal compositions (obtained from volcanic materials) in multidimensional space. The patterns of crystal compositions help scientists understand the pattern and magnitude of volcanic eruptions. This machine-learning algorithm was used to analyze a large (>2,000) compilation of mineral compositions from a highly active arc volcano in Villarrica, Chile.
Our Take: The idea of leveraging unsupervised machine learning in studying natural disasters is promising, particularly because much of the data pertaining to this research is raw and voluminous. Unsupervised learning-based techniques help in identifying patterns from unlabeled data. We hope this research can help improve current prediction systems for volcanic eruptions and also extend to volcanoes other than the arc volcano.
Anthropic’s quest for better, more explainable AI attracts $580M
Anthropic, founded under a year ago by Dario Amodei of OpenAI, aims to make AI more reliable and explainable. For a company that does not aim to turn a profit, Anthropic has raised a surprising amount of money: it raised a $124 million Series A when it was founded, and has now raised $580 million in a Series B. Anthropic has already published several papers and made progress along with its research directions of reverse engineering large language model behavior and language model steerability and robustness.
Our take: It's exciting to see a company with a mandate like Anthropic's receiving such funding--interpretability and robustness are vital to research areas as we enter a period dominated by large models whose inner workings are often mysterious. Yet, we should keep in mind that funds for this research may not keep coming forever. Industrial research labs may do incredibly important work, but if they do not turn a profit then their funding sources will be limited. It is possible that Anthropic and its ken could continue to find donors with deep pockets and stay afloat long enough to realize the fruits of their research agendas, but it is worth considering the event that they are not so lucky. AI safety researchers may be loath to commercialize their work themselves, but it may become a useful way to continue funding their work--the new alignment startup Conjecture seems to think so.
An algorithm that screens for child neglect raises concerns
Counties across the U.S. are adopting algorithmic tools to screen child negligence complaints and recommend whether or not an investigation should take place. These tools often give a risk score that suggests how likely a child will be placed in foster care in two years after investigation, and they make decisions based on features like “Medicaid, substance abuse, mental health, jail and probation records.” Developers of such tools argue that they help to ease the burden of social workers, who otherwise need to manually parse more than ten thousand complaints a year. Critics point to the racial disparities of the tools’ predictions - they flag a disproportionate number of Black children for mandatory neglect investigations. Such disparities largely disappeared from the dispatched investigations, where social workers have the final say. However, the lack of transparency on how these algorithms affect final outcomes as well as the fact that workers disagree with predicted risk scores one-third of the time make practical uses of such tools questionable. From attorney at ACLU:
If over a million dollars have been spent creating and maintaining this tool, only for call screeners to disagree with it, for racial disparities to stay essentially level, and for screen-ins to continue at unreasonably high rates, is that the best use of Allegheny County’s resources?
Our take: Child negligence investigations is a high-stakes domain that involves a lot of nuances that may require making decisions according to unobserved contexts and partial information, something AI algorithms are not good at. Obviously, in such domains, human decision-makers will be involved and have the final word, so it is difficult to see how AI algorithms can bring substantial improvements, especially in a cost-effective manner.
Other News
Research
Jurassic-X: Crossing the Neuro-Symbolic Chasm with the MRKL System - "We present the next evolution in natural language processing – the Modular Reasoning, Knowledge and Language (MRKL, pronounced “miracle”) system – and Jurassic-X, AI21 Labs’ implementation of it."
Deep Science: AI simulates economies and predicts which startups receive funding - "This week in AI, scientists conducted a fascinating experiment to predict how “market-driven” platforms like food delivery and ride-hailing businesses affect the overall economy when they’re optimized for different objectives, like maximizing revenue."
Using Machine Learning to Identify Far-Right Extremism - "This article discusses topics of extremism and radicalisation. Due to this; text, themes, and content relating to far-right extremism are present in this repository. Please continue with care."
An easier way to teach robots new skills - "Researchers have developed a technique that enables a robot to learn a new pick-and-place task with only a handful of human demonstrations. This could allow a human to reprogram a robot to grasp never-before-seen objects, presented in random poses, in about 15 minutes."
AI predicts lung cancer tumor growth after radiation : NHS study - "In a study newly published in the journal eBioMedicine, the researchers describe how they built an artificial intelligence model that can predict lung cancer patient’s individual risk of recurrence within two years of radiation therapy."
Using AI to detect cancer from patient data securely - "A new way of using artificial intelligence to predict cancer from patient data without putting personal information at risk has been developed by a team including University of Leeds medical scientists."
From blurry to bright: AI tech helps researchers peer into the brains of mice - "Johns Hopkins biomedical engineers have developed an artificial intelligence (AI) training strategy to capture images of mouse brain cells in action."
Which Animal Viruses Could Infect People? Computers Are Racing to Find Out. - "Machine learning is known for its ability to spot fraudulent credit charges or recognize faces. Now researchers are siccing the technology on viruses. "
Facebook Parent Meta Is Studying the Human Brain to Improve AI - "The social media giant is hoping to change that with a new long-term research effort that compares how the human brain learns language to how AI processes information."
Lensless' imaging through advanced machine learning for next generation image sensing solutions - "Recent advances in computing technology can simplify the lens system by substituting computing for some parts of the optical system."
Engineers use artificial intelligence to capture the complexity of breaking waves - "Their model’s predictions should help researchers improve ocean climate simulations and hone the design of offshore structures."
Schlumberger expands AI innovation network with opening of artificial intelligence center in Europe - "Schlumberger announced it has expanded its global INNOVATION FACTORI network with the inauguration of a new center in Oslo, Norway."
Microrobot collectives display versatile movement patterns - "Collective behavior and swarm patterns are found everywhere in nature. Robots can also be programmed to act in swarms. Researchers have developed collectives of microrobots, which they can move in every formation they wish."
Applications
Meta is using AI to create low-carbon concrete for its data centres - "Meta has used AI to develop concrete that emits 40 per cent less carbon, but researchers say that other concrete mixtures can already achieve that"
Everyone Has Opinions, Even AI - "Coders and marketers show that machine learning can geek out on product reviews, too."
AI can now kill those annoying cookie pop-ups - "The notifications have been put on notice"
Machine learning, harnessed to extreme computing, aids fusion energy development - "MIT research scientists Pablo Rodriguez-Fernandez and Nathan Howard have just completed one of the most demanding calculations in fusion science — predicting the temperature and density profiles of a magnetically confined plasma via first-principles simulation of plasma turbulence."
AI can predict missed appointments. How can hospitals use that data for better care? - "Missed appointments are a common problem at health systems. And they’re a particularly attractive target for machine learning researchers, who can use patient datasets to get a handle on what’s causing patients to miss out on needed care."
Business
Start-up Pony.ai says it's the first self-driving company to get a taxi license in China - "Self-driving start-up Pony.ai announced Sunday it received a taxi license, the first of its kind in China. The license allows Pony.ai to operate 100 self-driving cars as traditional taxis in the Nansha district of the southern city of Guangzhou, the company said."
Virtual Being Developer Hour One Raises $20M - "As part of my PhD research at Bristol University I’ve developed Pinpoint, a machine learning binary classification framework, used to identify violent far-right extremism"
Investors flock to fund an AI cornerstone: Feature stores - "Enter the feature store, a centralized repository for organizing, storing and serving the features that AI systems rely on."
This 1,000-core RISC-V processor is generating buzz in the AI space - "According to a press release from the computer software company, its new ET-SoC-1 AI Inference Accelerator is undergoing initial evaluations by a number of firms ahead of its release."
Inside Meta's AI optimization platform for engineers across the company - "To address these needs at Meta, we’ve built an end-to-end AI platform called Looper, with easy-to-use APIs for optimization, personalization, and feedback collection."
Armed with new $150 million raise, Oregon robotics startup sets its sights on Pittsburgh - "The company is also looking to expand in Pittsburgh, where its CEO Damion Shelton is based, as it looks to move from one side of the city to the other in pursuit of more space."
Adept aims to build AI that can automate any software process - "[W]e’re training a neural network to use every software tool in the world, building on the vast amount of existing capabilities that people have already created."
How an engineer at Google's DeepMind AI lab pulled off a career pivot from marketing into machine learning without a Ph.D. - "As more industries find innovative ways to apply artificial intelligence to their goods and services, companies want to staff up with experts in machine learning — fast."
VisionNav Robotics closes $80M for AGVs - "VisionNav Robotics, a Shenzhen, China-based company that develops logistics robots, raised more than $80 million in Series C+ funding. VisionNav is committed to providing automated material handling systems with visual artificial intelligence navigation technology."
Concerns
IBM research details AI skills gap across Europe - "According to IBM research, the tech sector across the UK, Germany and Spain is struggling to find employees with adequate AI skills or experience The AI skills gap across Europe threatens to stifle digital innovation and hold back economic growth"
The risks of attacks that involve poisoning training data for machine learning models - "The attacks examined by the researchers separate the loss distributions of members and non-members, making them more distinguishable."
Analysis
Q&A: Inside DARPA’s Subterranean Challenge - "What SubT means for the future of autonomous robots"
Policy
Lyft exec Craig Martell tapped as Pentagon’s AI chief - "Lyft exec Craig Martell will take a major role in the Pentagon’s digital world."
U.S. Senate Democrats urge Buttigieg to develop autonomous vehicle rules - "A group of 12 U.S. Senate Democrats on Wednesday urged Transportation Secretary Pete Buttigieg to develop a comprehensive federal framework for autonomous vehicles, according to a letter seen by Reuters."
Tough questions, few answers: The FDA wrestles with its approach to AI in medicine - "Earlier this week, the agency put one of its regulatory officers in front of a room full of scientists clamoring for answers on how the agency intends to regulate artificial intelligence in medicine."
Pentagon’s flagship AI effort, Project Maven, moves to NGA - " Once the Pentagon’s top-priority program to speed the use of artificial intelligence across the military, Project Maven is now being transferred to the National Geospatial-Intelligence Agency, according to senior Intelligence Community officials."
DoD Announces Dr. Craig Martell as Chief Digital and Artificial Intelligence Officer - "As the CDAO, Dr. Martell will serve as the department’s senior official responsible for accelerating the adoption of data, analytics, digital solutions, and AI functions to generate decision advantage from the boardroom to the battlefield."
Energy Department launches council to coordinate AI activities - "The Department of Energy established the Artificial Intelligence Advancement Council earlier this month to coordinate funding and development of algorithms and hold agencies accountable for how they are used."
Fun
I keep getting terrible advice from AI assistants. Help me rate the worst and the best responses :) - "I decided to prepare a compilation and rank the answers. Can you help me with deciding how bad or good are some of the answers here?"
This Woman Used AI To Show What "Harry Potter" Characters Would Look Like In Real Life Based On Their Book Descriptions, And My Jaw Is On The Floor - "This AI version of Draco gives me full-body chills."
Copyright © 2022 Skynet Today, All rights reserved.