In COVID-related AI news, Andy and Dave discuss survey results from Algorithmia, which shows that IT directors at large companies are looking to spend more money on AI/ML projects due to the pandemic. In regular AI news, the bipartisan Future of Defense Task Force releases its 2020 report, which includes the suggestion of using the Manhattan Project as a model to develop AI technologies. The US and UK sign an agreement to work together on trustworthy AI. Facebook AI releases Dynabench as a way to dynamically benchmark the performance of machine learning algorithms. Amsterdam and Helsinki launch AI registers that explain how they use algorithms, in an effort to increase transparency. In research, the Allen Institute of AI, University of Washington, and University of North Carolina publish research on X-LXMERT (learning cross-modality encoder representations from transformers), which trains GPT-3 on both text and images, to then generate images from scratch by providing descriptions (e.g., a large clock tower in the middle of a town). Researchers at Swarthmore College and Los Alamos National Labs demonstrate the challenges that neural networks of various sizes have in learning Conway’s Game of Life. Maria Jeansson, Claudio Sanna, and Antoine Cully create a stunning visual infographic on the “automated futures” technologies. And Joshua Epstein, a longtime expert in agent-based modeling, provides the European Social Stimulation Association Award Keynote speech.  

AI with AI

Rebroadcast: The Robohattan Project

Andy and Dave discuss the latest in AI news, including a report that the Israel Defense Forces used a swarm of small drones in mid-May in Gaza to locate, identify, and attack Hamas militants, using Thor, a 9-kilgram quadrotor drone. A paper in the Journal of American Medical Association examines an early warning system for sepsis, and finds that it misses out on most instances (67%) of cases, and frequently issued false alarms (to which the developer contests the results). A new bill, the Consumer Safety Technology Act, directs the US Consumer Product Safety Commission to run a pilot program to use AI to help in safety inspections. A survey from FICO on The State of Responsible AI (2021) shows, among other things, a disinterest in the ethical and responsible use of AI among business leaders (with 65% of companies saying that can’t explain how specific AI model predictions are made, and 22% of companies have an AI ethics board to consider questions on AI ethics and fairness). In a similar vein, a survey from the Pew Research Center and Elon University’s Imagining the Internet Center found that 68% of respondents (from across 602 leaders in the AI field) believe that AI ethical principles will NOT be employed by most AI systems within the next decade; the survey includes a summary of the respondents’ worries and hopes, as well as some additional commentary. GitHub partners with OpenAI to launch CoPilot, a “Programming Partner” that uses contextual cues to suggest new code. Researchers from Stanford University, UC San Diego, and MIT research Physion, a visual and physical prediction benchmark to measure predictions about commonplace real world physical events (such as when objects: collide, drop, roll, domino, etc). CSET releases a report on Machine Learning and Cybersecurity: Hype and Reality, finding that it is unlikely that machine learning will fundamentally transform cyber defense. Bengio, Lecun, and Hinton join together to pen a white paper on the role of deep learning in AI, not surprisingly eschewing the need for symbolic systems. Aston Zhang and Zack C. Lipton, and Alex J Smola release the latest version of Dive into Deep Learning, now over 1000 pages, and living only as an online version.  Follow the link below to visit our website and explore the links mentioned in the episode. https://www.cna.org/CAAI/audio-video

GPT Is My CoPilot

Andy and Dave discuss the latest in AI news, including research from the San Diego School of Medicine, which used an AI algorithm to analyze terabytes of gene expression data in response to viral infections, identifying 20 genes that predict the severity of a patient’s response (across many different viruses). Deputy Secretary of Defense Kathleen Hicks announces a new AI and Data Acceleration initiative, which includes operational data teams and flyaway technical experts. China says it has AI fighter jet pilots that can beat human pilots in simulated dogfights. A study from Stanford estimates the density of CCTV cameras in large cities around the globe (by using computer vision algorithms on street view image data). NIST held a workshop on AI Measurement and Evaluation, with an interesting 22-page read-ahead document. Appen updates its State of AI and Machine Learning report, examining various business-related views and metrics on AI, showing a general maturing of the AI market. Researchers from Tubingen and Max Planck show that the behavioral difference between human and machine vision is narrowing, but still has room for improvement (particularly with out-of-distribution data). Researchers from Stanford, University of College London, and MIT develop a counterfactual simulation model to provide quantitative predictions on how people think about causation, possibly serving as a bridge between psychology and AI. Adam Wagner uses a reinforcement learning approach to search for examples that would disprove conjectures in graph theory and finds examples that disprove five such conjectures. Justin Solomon’s Numerical Algorithms provides the core methods for machine learning. And Budiansky publishes a look at the life of Kurt Gödel, in Journey to the Edge of Reason. Follow the link below to visit our website and explore the links mentioned in the episode. https://www.cna.org/CAAI/audio-video  

Journey to the Cause of Reason

 Andy and Dave discuss the latest in AI news, including the European Commission’s proposal for the regulation of AI. A report in Nature Medicine examines the limitations of the evaluation process for medical devices using AI that the FDA approves. Researchers at MIT translate spider webs into sounds to explore how spiders might sense their world, and they using machine learning to classify sounds by spider activities. An NIH panel releases its preliminary ethics rules on making brain-like structures such as neural organoids and neural transplants, and finds little evidence that these structures experience humanlike consciousness or pain. And Andy and Dave spend some time with xenobioticists Sam Kriegman and Doug Blackiston, who discuss the motivations and findings behind their latest generation of xenobots, synthetic living machines that they have been designing and building in their labs. 

Rebroadcast: Xen and the Art of Motorcell Maintenance

Andy and Dave discuss the latest in AI news, including the launch of the National AI Research Resource Task Force, which will serve as a federal advisory committee and produce at least two reports to Congress (a roadmap and implementation plan) by November 2022. Google and Harvard University release a 1.4 PB reconstruction of a cubic millimeter of human brain tissue. Google reports a deep reinforcement-learning system that outperforms humans in designing floorplans for microchips, both in time and inefficiency. Researchers from the UK, Germany, and China fuse electronics to the Madagascar hissing cockroach to create an insect-computer hybrid for autonomous search and rescue. The Navy’s MQ-25 tanker drone refuels a manned aircraft for the first time. Researchers use large-scale experiments and machine learning to discover a greater hierarchy of theories of human decision-making. OpenAI introduces a Process for Adapting Language Models to Society (PALMS) as a way to try to mitigate bias in transformer models such as GPT-3. A concept paper from DeepMind examines why reward systems are enough to constitute a solution to artificial general intelligence. And Richard Sutton and Andrew Barto publish the second edition of Reinforcement Learning: An Introduction. Click here to visit our website and explore the links mentioned in the episode. 

Reward of the Coprophages

Andy and Dave discuss the latest in AI news, starting with the US Consumer Products Safety Commission report on AI and ML. The Deputy Secretary of Defense outlines Responsible AI Tenets, along with mandating the JAIC to start work on four activities for developing a responsible AI ecosystem. The Director of the US Chamber of Commerce’s Center for Global Regulatory Cooperation outlines concerns with the European Commission’s newly drafted rules on regulating AI. Amnesty International crowd-sources an effort to identify surveillance cameras that the New York City Police Department has in use, resulting in a map of over 15,000 camera locations. The Royal Navy uses AI for the first time at sea against live supersonic missiles. And the Ghost Fleet Overlord unmanned surface vessel program completes its second autonomous transit from the Gulf Coast, through the Panama Canal, and to the West Coast. Finally, CNA Russia Program team members Sam Bendett and Jeff Edmonds join Andy and Dave for a discussion on their latest report, which takes a comprehensive look at the ecosystem of AI in Russia, including its policies, resourcing, infrastructure, and activities. Click here to visit our website and explore the links mentioned in the episode.

No Time to AI

Information about the AI Event Series mentioned in this episode: https://twitter.com/CNA_org/status/1400808135544213505?s=20 To RSVP contact Larry Lewis at LewisL@cna.org. Andy and Dave discuss the latest in AI news, including a report on Libya from the UN Security Council’s Panel of Experts, which notes the March 2020 use of the “fully autonomous” Kargu-2 to engage retreating forces; it’s unclear whether any person died in the conflict, and many other important details are missing from the incident. The Biden Administration releases its FY22 DoD Budget, which increases the RDT&E request, including $874M in AI research. NIST proposes an evaluation model for user trust in AI and seeks feedback; the model includes definitions for terms such as reliability and explainability. EleutherAI has provided an open-source version of GPT-3, called GPT-Neo, which uses an 825GB data “Pile” to train, and comes in 1.3B and 2.7B parameter versions. CSET takes a hands-on look at how transformer models such as GPT-3 can aid disinformation, with their findings published in Truth, Lies, and Automation: How Language Models Could Change Disinformation. IBM introduces a project aimed to teach AI to code, with CodeNet, a large dataset containing 500 million lines of code across 55 legacy and active programming languages. In a separate effort, researchers at Berkeley, Chicago, and Cornell publish results on using transformer models as “code generators,” creating a benchmark (the Automated Programming Progress Standard)  to measure progress; they find that GPT-Neo could pass approximately 15% of introductory problems, with GPT-3’s 175B parameter model performing much worse (presumably due to the inability to fine-tune the larger model). The CNA Russia Studies Program leases an extensive report on AI and Autonomy in Russia, capping off their biweekly newsletters on the topic. Arthur Holland Michel publishes Known Unknowns: Data Issues and Military Autonomous Systems, which clearly identifies the known issues in autonomous systems that cause problems. The short story of the week comes from Asimov in 1956, with “Someday.” And the Naval Institute Press publishes a collection of essays in AI at War: How big data, AI, and machine learning are changing naval warfare. Finally, Diana Gehlhaus from Georgetown’s Center for Security and Emerging Technology (CSET),  joins Andy and Dave to preview an upcoming event, “Requirements for Leveraging AI.” Interview with Diana Gehlhaus: 33:32 Click here to visit our website and explore the links mentioned in the episode.  

Someday My ‘Nets Will Code

Information about the AI Event Series mentioned in this episode: https://twitter.com/CNA_org/status/1400808135544213505?s=20 To RSVP contact Larry Lewis at LewisL@cna.org. Andy and Dave discuss the latest in AI news, including the first flight of a drone equipped with the Air Force’s Skyborg autonomy core system. The UK Office for AI publishes a new set of guidance on automated decision-making in government, with Ethics, Transparency and Accountability Framework for Automated Decision-Making. The International Red Cross calls for new international rules on how governments use autonomous weapons. Senators introduce two AI bills to improve the US’s AI readiness, with the AI Capabilities and Transparency Act and the AI for the Military Act. Defense Secretary Lloyd Austin lays out his vision for the Department of Defense in his first major speech, stressing the importance of emerging technology and rapid increases in computing power. A report from the Allen Institute for AI shows that China is closing in on the US in AI research, expecting to become the leader in the top 1% of most-cited papers in 2023. In research, Ziming Liu and Max Tegmark introduce AI Poincaré, an algorithm that auto-discovers conserved quantities using trajectory data from unknown dynamics systems. Researchers enable a paralyzed man to “text with his thoughts,” reaching 16 words per minute. The Stimson Center publishes A New Agenda for US Drone Policy and the Use of Lethal Force. The Onlife Manifesto: Being Human in a Hyperconnected Era, first published in 2015, is available for open access. And Cade Metz publishes Genius Makers, with stories of the pioneers behind AI. Click here to visit our website and explore the links mentioned in the episode.

Just the Tip of the Skyborg

Andy and Dave welcome Arthur Holland Michel to the podcast for a discussion on predictability and understandability in military AI. Arthur is an Associate Researcher at the United Nations Institute for Disarmament Research, a Senior Fellow at the Carnegie Council for Ethics in International Affairs, and author of the book Eyes in the Sky: the Secret Rise of Gorgon Stare and How It Will Watch Us All. Arthur recently published The Black Box, Unlocked: Predictability and Understandability in Military AI, and the three discuss the inherent challenges of artificial intelligence and the challenges of creating definitions to enable meaningful global discussion on AI.  

Rebroadcast: A.I. in the Sky

Andy and Dave discuss the latest in AI news, including a new AI website from the White House at AI.gov, which provides a variety of resources on recent reports, news, key US agencies, and other information. The U.S. Navy destroys a surface vessel using a swarm of drones (in combination with other weapons) for the first time. The NYPD announces the retirement of its Boston Dynamics robot dog (Digidog) due to negative public reaction at its use. The French Defence Ministry releases a report on the Integration of Autonomy into Lethal Weapon Systems. A paper in Digital Medicine examines the use of decision-aids in clinical settings. Matt Ginsberg (along with the Berkeley NLP Group) develops Dr. Fill, an algorithm that won this year’s American Crossword Puzzle Tournament, with three total errors. And the University of Glasgow publishes research on using return echoes over time to render a 3D image of an environment. Researchers use MRI and machine learning to identify brain activation configurations for 12 different cognitive tasks. Facebook AI Research, Inria, and Sorbonne publish research on emerging properties of self-supervised vision transformers, which includes the ability to segment objects with no supervision or segmentation-targeted objectives. Florian Jaton publishes The Constitution of Algorithms: Ground-Truthing, Programming, Formulation, which examines how algorithms come to be. Melanie Mitchell publishes a paper on Why AI Is Harder Than We Think. And UneeQ creates a Digital Einstein for people to interact with. Click here to visit our website and explore the links mentioned in the episode.  

Doggone

Andy's out this week, but Dave recently had a chance to do a series of interviews on a paper that he wrote, Superhumans, Implications of genetic engineering and human-centered bioengineering. So this week's podcast will feature a rebroadcast of the interview that Dave had on Titillating Sports. A big thanks to Rick Tittle and Darren Peck from the Sports Byline USA Network for conducting the interview and for allowing us to share it. Rick and Dave discuss the latest and greatest in genetic engineering and human-centered technology and talk about some of the near-term and far-term implications. Report: https://www.cna.org/CNA_files/PDF/Superhumans-Implications-of-Genetic-Engineering-and-Human-Centered-Bioengineering.pdf Titillating Sports Podcast: https://podcasts.apple.com/us/podcast/titillating-sports-with-rick-tittle/id1451555608

Superhumans

Andy and Dave discuss the latest AI news and research, including a blog post from the Federal Trade Commission that businesses can and will be held accountable for the fairness of their algorithms. A bipartisan coalition of U.S. Senators has introduced the “Fourth Amendment Is Not for Sale Act,” which would ban law enforcement and intelligence agencies from buying data on people in the U.S. and about Americans abroad, if that data was obtained from a user’s account or device, through deception, hacking or other violations of privacy policies or terms of service. Bob Work releases his seven Principles for the Combat Employment of Weapon Systems with Autonomous Functionalities; these principles go into much greater detail about employment and provide a useful way to discuss issues surrounding autonomous weapons. The Congressional Research Service provides a short, but dense overview on Lethal Autonomous Weapon Systems. The Ozcan Research Group and UCLA publish research that identifies handwritten numbers by using an optical network made from 3D printed wafers that diffract polarized light. Project CETI aims to decode whale language using decades of recorded whale sounds. Researchers from the Centre for Neuroscience and the Indian Institute of Science explore whether the similarities and differences in how deep networks “see” compared to humans, by examining 13 specific perceptual effects, such as mirror confusion. Researchers from Stanford and UCSD examine how children’s drawing and recognition of visual concepts change over time. On a similar topic, other research explores the relationship between episodic memory and generalization, finding that the relationship changes as children develop. The book of the week is an open access paper from Stanford, which examines and provides tools for vector embedding of large sets of data, to include minimizing distortion. Ben Vickers and K. Allado-McDowell publish the Atlas of Anomalous AI, with reference to the Mnemosyne Atlas. Andy and Dave accidentally change the pronunciation of “neh-meh-zeen” and completely destroy the joke of this week’s podcast title. And take a look at the “fun” site of the week, which puts an eye on webcams, with the EyeCam, the webcam that looks like and mimics the movements of the human eye. Listeners Survey: https://bit.ly/3bqyiHk Click here to visit our website and explore the links mentioned in the episode. 

Mnemosyne That Before

Andy and Dave discuss the latest in AI news, including the European Commission’s proposal for the regulation of AI. A report in Nature Medicine examines the limitations of the evaluation process for medical devices using AI that the FDA approves. Researchers at MIT translate spider webs into sounds to explore how spiders might sense their world, and they using machine learning to classify sounds by spider activities.  An NIH panel releases its preliminary ethics rules on making brain-like structures such as neural organoids and neural transplants, and finds little evidence that these structures experience humanlike consciousness or pain. And Andy and Dave spend some time with xenobioticists Sam Kriegman and Doug Blackiston, who discuss the motivations and findings behind their latest generation of xenobots, synthetic living machines that they have been designing and building in their labs. Listeners Survey: https://bit.ly/3bqyiHk Click here to visit our website and explore the links mentioned in the episode. 

Xen and the Art of Motorcell Maintenance

Andy and Dave discuss the latest in AI news, including the National Intelligence Council’s 7th Edition Global Trends 2040 Report, which sprinkles the importance of AI and ML throughout future trends. A BuzzFeed report claims that the NYPD has misled the public about its use of the facial recognition tool, Clearview AI, having run over 5100 searches with the tool.  European Activist Groups ask the European Commission to ban facial recognition completely, with calls to protect “fundamental rights” in Europe. A report in Digital Medicine examines the diagnostic accuracy of deep learning in medical imaging studies, and calls for an immediate need to develop AI guidelines. Neuralink demonstrates the latest with its brain-computer interface device with a demonstration that shows a monkey playing Pong with his brain. And the Director of the JAIC, Lt Gen Groen, and the co-chair of the NSCAI, Bob Work, spoke for about an hour on the use and direction of AI in the Department of Defense. In research, Andrew Jones examines how different parameters scale with board games, identifying the scaling of scaling laws. Research for AIST, Tokyo Institute of Technology, and Tokyo Denki University demonstrate that they can pre-train a CNN using no natural images, but instead using digital images created using fractals. In the paper of the week, Ben Goertzel provides his general theory of general intelligence. And the fun site of the week features the 1996 game, “Creatures,” with a look into the AI that made them come alive. Listeners Survey: https://bit.ly/3bqyiHk Click here to visit our website and explore the links mentioned in the episode. 

Donkey Pong

Andy and Dave discuss the latest in AI news, including the resignation of Samy Bengio from Google Brain, which fired ethicists Gebru in December and Mitchell in February. The Joint AI Center releases its request for proposals on Data Readiness for AI Development (DRAID). DARPA prepares for the quantum age with a program for Quantum Computer Benchmarking. And a separate DARPA program seeks to enable fully homomorphic encryption with its Data Protection in Virtual Environments (DPRIVE) program. A poll from Hyland on digital distrust shows that Americans think that over the next decade, AI has the most potential to cause harm. Amazon introduces the next level of “biometric consent” required for its delivery drivers, which includes an always-on camera observing the driver and gathering other data; drivers will lose their jobs if they do not consent to the monitoring. And Josh Bongard of the University of Vermont and Michael Levin of Tufts University along with other researchers from Wyss and Harvard join together to form the Institute for Computationally Designed Organisms (ICDO), which will focus on “AI-driven designs of new life forms.” In research, Bongard publishes the latest iteration of its mobile living machines, with Xenobots II, using frog cells to create life forms capable of motion, memory, and manipulation of the world around them. Researchers from the universities of Copenhagen, York, and Shanghai use neural cellular automata to grow 3D objects and functional machines within the Minecraft world. And OpenAI Robotics demonstrates the ability for a robotic arm to solve manipulation tasks, including tasks with previously unseen goals and objects, with asymmetric self-play. And the Book / Fun Site of the Week comes from the Special Interest Group on Harry Q. Bovik (SIGBOVIK), which presents “April Fools” research, descriptions of truly absurd, but fascinating, research. Listeners Survey: https://bit.ly/3bqyiHk Click here to visit our website and explore the links mentioned in the episode. 

Xenomania

AI with AI explores the latest breakthroughs in artificial intelligence and autonomy, and discusses the technological and military implications. Join Andy Ilachinski and David Broyles as they explain the latest developments in this rapidly evolving field. The views expressed here are those of the commentators and do not necessarily reflect the views of CNA or any of its sponsors.

hubhopper

AI with AI

hubhopper