

Discover more from Last Week in AI
Last Week in AI #182: AI for text-to-video, Tesla's new Dojo AI supercomputer, state of AI policies, and more!
The future of AI text-to-video generation, Tesla builds in-house supercomputer for self-driving AI, latest in U.S. AI policies
Top News
After Text-to-Image, Now it’s Text-to-Video
Recently, AI-powered text-to-image technology has gone from research novelty to mainstream sensation with the releases of DALLE-2, Midjourney, and Stable Diffusion. Now, researchers are tackling the next challenge, text-to-video. This is a much more difficult problem because of the high compute cost and lack of sufficient datasets. However, there are some signs that text-to-video might be coming earlier than many may think. Stable Diffusion is already being integrated into runwayml’s AI-assisted video editor for tasks like background replacement, Deepmind has new results on generating realistic albeit very static and blurry video clips, and Microsoft has a model that can animate nature scenes, like moving clouds or rain. The main bottleneck appears to be limited datasets, with the latest text-to-video dataset only containing 41k videos. Many are developing new datasets to improve text-to-video in the near future.
Our take. The realization of AI-generated videos is not a question of if but when. The basic ingredients are already there, and the rapid growth of the generative models research community in recent years will only accelerate this progress. What’s more interesting is also the possibility of multi-modal generative models - imagine a model that can generate videos, dialog/speech, sound effects, and music, all at the same time. We’re not that far (probably only a few years) from such models going mainstream, and only time will tell the impact of such technologies on art, media, and society at large.
Tesla wants to take machine learning silicon to the Dojo
If you watched Tesla's AI Day, you'll know that the electric vehicle company has joined the myriad companies working on AI chips. More recently, Tesla revealed a number of details about their supercomputing architecture at Hot Chips 34. The system, unlike many others, is "built from an entirely custom architecture that spans the compute, networking, and input/output (I/O) silicon to instruction set architecture (ISA), power delivery, packaging, and cooling." In the quest to run specialized ML algorithms at scale, Tesla has started from scratch to avoid the limitations of traditional methods for scaling distributed workloads. The Dojo training tile can run 556 trillion floating point operations (at FP32 resolution) per second in a 15kW liquid-cooled package. Tesla has mentioned it plans to make its chip available to researchers, but has no timeline for this yet.
Our Take: I'm excited to see further innovation in developing AI chips, and Tesla certainly has an important use case. I think we'll need to see and invest in a variety of new computing architectures to figure out what will push AI forward, and it's good to see that Tesla has taken a "back to basics" approach here.
The State of State AI Policy
In the last few years, AI applications have become more common in the everyday lives of the general public, and technologies such as facial recognition, large language models, and autonomous vehicles are already being monetized. Their increased use is met with growing concerns about the potential harms of AI applications. While many states have taken a strong stance against the use of facial recognition in law enforcement, states like Illinois have approved the regulated use of video analysis as an aid in hiring applicants in interviews. An encouraging policy was passed in Mississippi, where lawmakers proposed to establish a school curriculum for artificial intelligence and machine learning, ensuring the next generation learns about the ethics of AI from a young age.
Our Take: As AI researchers, we find ourselves in a tricky position; on the one hand, we realize the importance of freedom in research and that the frequency as well as likelihood of scientific breakthroughs is typically inversely proportional to the constraints and laws bounding them. On the other hand, we are aware of the brittleness of AI models as well as their potential for social harm. This is a highly complex issue necessitating deep, critical analyses from many stakeholders. It’s interesting to see how the different U.S. localities are experimenting with different AI regulations and priorities. Noting their impact in the near future could help us better understand how to best regulate AI.
Other News
Research
Google's RawNeRF research brings us closer to yelling 'enhance' at images - "Earlier this year, Google Research presented its work on RawNeRF with “NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images” at the CVPR (Computer Vision and Pattern Recognition) 2022 conference. RawNeRF, which uses RAW images, does a very good job of removing noise."
Highly-Efficient New Neuromorphic Chip for AI on the Edge - "The NeuRRAM chip is the first compute-in-memory chip to demonstrate a wide range of AI applications while using just a small percentage of the energy consumed by other platforms while maintaining equivalent accuracy."
Artificial intelligence model can detect Parkinson’s from breathing patterns - "An MIT-developed device with the appearance of a Wi-Fi router uses a neural network to discern the presence and severity of one of the fastest-growing neurological diseases in the world."
AI-based method for dating archeological remains - "By analyzing DNA with the help of artificial intelligence (AI), an international research team led by Lund University in Sweden has developed a method that can accurately date up to ten-thousand year-old human remains."
Study Says Machine Learning MRI Model May Help Predict Recurrence Risk of Hepatocellular Carcinoma - "An emerging machine learning model, which incorporates post-contrast magnetic resonance imaging (MRI), reportedly enhances the prediction of recurrence among patients treated for early-stage hepatocellular carcinoma (HCC)."
AI-designed camera only records objects of interest while being blind to others - "A new research paper published in eLight demonstrated a new paradigm to achieve privacy-preserving imaging by building a fundamentally new type of imager designed by AI."
Protons contain intrinsic charm quarks, machine-learning analysis suggests – Physics World - "The Large Hadron Collider: evidence for intrinsic charm quarks in protons has been found in LHC data."
Deep learning pioneer Geoffrey Hinton receives prestigious Royal Medal from the Royal Society - "The University of Toronto’s Geoffrey Hinton has been honoured with the Royal Society’s prestigious Royal Medal for his pioneering work in deep learning – a field of artificial intelligence that mimics the way humans acquire certain types of knowledge."
Applications
The Bionic-Hand Arms Race - "The prosthetics industry is too focused on high-tech limbs that are complicated, costly, and often impractical"
CSIRO scientists develop machine learning tool to predict patient deterioration - "An early deterioration alert can be set two to eight hours before an actual event."
Google has opened up the waitlist to talk to its experimental AI chatbot - "Earlier this year, Google unveiled AI Test Kitchen — an Android app that lets users talk to one of its most advanced AI chatbots, LaMDA 2. Today, the company is opening up registrations for early access."
Advancing conservation with AI-based facial recognition of turtles - "Finding solutions to improve turtle reidentification and supporting machine learning projects across Africa. Protecting the ecosystems around us is critical to safeguarding the future of our planet and all its living citizens."
Machine-Learning Tools Predict Post-Op Complications, Surgery Duration - "Researchers from Washington University in St. Louis have developed machine-learning tools that can predict post-operative complications and surgery duration using perioperative data."
Machine Learning Could Revolutionize Mineral Exploration - "Using a global data set of zircon trace elements, new research demonstrates the power of machine learning algorithms to accurately identify and locate porphyry copper deposits."
Business
Writer’s GPT-powered CoWrite handles content ‘drudgery’ and leaves creativity to humans - "Writer is an AI-powered tool for checking and guiding content creators in organizations where voice and branding are essential. Its new feature CoWrite does that writing itself — but don’t worry, this isn’t quite the content apocalypse we’ve been worried about."
Virtual Rapper FN Meka Powered By Artificial Intelligence Signs to Major Label - "Hip-hop is now in the world of Artificial Intelligence. A virtual rapper named FN Meka, which is powered by AI, has signed with a major label."
SCALE AI announces a major financing round in artificial intelligence: $50 million to support 12 new projects - "With an additional 12 projects representing more than $50 million in investments announced today by SCALE AI, artificial intelligence (AI) is continuing on its path of rapid adoption by Canadian industry leaders."
Sanas, the buzzy Silicon Valley startup that wants to make the world sound whiter - "Silicon Valley startup Sanas has a lofty goal: to make call center workers sound white and American, no matter the country they're from. And that's just the beginning of their grand plan."
Half of AI Models Never Make It To Production: Gartner - "The survey found an average of 54% of AI projects are making the leap from pilot to production, up from 53% as reported in the Gartner 2019 AI in Organizations survey. "
World's First AI Writing Assistant Powered by GPT-3 Launches in the App Store. - "ParagraphAI Inc., a leading developer of advanced AI-powered productivity tools, today announced the release of their natural language processing ("NLP") AI Writing Assistant."
DearDoc Announces Partnership with WebMD and Vitals.com to bring AI Chat to Practitioner Listings - "DearDoc, the cloud-based growth platform for practice owners, is announcing a partnership with WebMD and Vitals.com to enable direct patient-to-practitioner communication through DearDoc’s AI Chatbot technology."
FDA clears Point Robotics' hand-held spinal surgery system - "In contrast to other up-and-coming robotic surgery systems, one newly cleared by the FDA acts less like an autonomous assistant and more like a surgeon’s extra pair of hands."
Levi’s Jumps Inventory Hurdles With Machine Learning E-commerce Tech - "Goodbye inventory roadblocks. On-shelf availability just got a little bit smarter with Levi Strauss & Co.’s machine learning-powered innovation, reducing obstacles related to stock-outs. "
AI-powered videoconferencing platform Headroom raises $9M - "Headroom, a startup developing AI-powered software to make meetings ostensibly more efficient, today announced that it raised $9 million in funding led by Equal Opportunity Ventures with participation from Gradient Ventures, LDV Capital, AME Cloud Ventures, and Morado Ventures."
Telexistence Announces Mass Production of AI Robot to be Installed in Japan's 300 FamilyMart Stores - "Telexistence Inc. today announced the mass production of its originally developed artificial intelligence robot, TX SCARA, to be installed in 300 FamilyMart stores, Japan's top-tier convenience store chain in major metropolitan areas, starting later this month, validating its AI-based "robot-as-a-service" solution for grocery retailers."
Concerns
You just hired a deepfake. Get ready for the rise of imposter employees. - "Data-tampering represents a different type of threat, and one that could be potentially even more serious for certain organizations, depending on the situation."
Why is Meta’s new AI chatbot so bad? - "It’s a conspiracy theorist obsessed with Genghis Khan, and that’s not even the weird part."
This AI Tool Is Being Used To Make Freaky, Machine-Generated Porn - "People are creating the DALL-E of porn with something called ‘stable diffusion’—and the company behind the tool wants no part of it."
Deepfakes for all: Uncensored AI art model prompts ethics questions - "A new open source AI image generator capable of producing realistic pictures from any text prompt has seen stunningly swift uptake in its first week."
We Need to Talk About How Good A.I. Is Getting - "We’re in a golden age of progress in artificial intelligence. It’s time to start taking its potential and risks seriously."
AI Creating 'Art' Is An Ethical And Copyright Nightmare - "If a machine makes art, is it even art? And what does this mean for actual artists?"
Analysis
Military Drone Swarms and the Options to Combat Them - "This article provides a broad overview on the current state of drones for commercial and military use, the impact drone swarms can play in the military environment, and the options available to combat swarms."
Policy
Human drivers in UK won’t be liable for accidents when vehicle is self-driving - "Manufacturers operating self-driving vehicles in the UK will be liable for a vehicle’s actions when in autonomous mode, according to the country’s new roadmap to achieve a widespread rollout of self-driving vehicles by 2025."
FDA Begins Third Phase of AI Imported Seafood Pilot Program - "U.S. Food and Drug Administration (FDA) launched the third phase of the Artificial Intelligence (AI) Imported Seafood Pilot program. The pilot uses AI and machine learning (ML) to strengthen the screening process for foods entering the U.S. market."
The outgoing White House AI director explains the policy challenges ahead - "Algorithmic bias, a small AI workforce, and a lack of AI education may hold the US back from leading the world in the technology, says Lynne Parker."
Expert Opinions
AI And The Limits Of Language - "An artificial intelligence system trained on words and sentences alone will never approximate human understanding."
Explainers
Building a backend from scratch using only OpenAI Codex - "In the last few weeks, I and the rest of our team have been playing around a lot with OpenAI's Codex Model, as we've incorporated it into our latest feature (the suggester), and we've been generally impressed by it's performance."
Fun
Modern Day Henry VIII Recreated With AI, And He Doesn't Look Happy About It - "An Artificial Intelligence (AI)-generated image has given us a glimpse of what King Henry VIII of England would look like in the modern world, and he does not look comfortable with the circumstances. AI-generated imagery is getting spookily good."
AI creates ‘scariest thing on Earth’ — and it will haunt your dreams - "We’re using technology for evil instead of good. The popular site Craiyon AI (formerly DALL-E mini) created the vivid nightmare fuel by pulling together images based on keywords that are typed into a prompt."
Here's how the best AI art generators compare - "We already know that AI art generators are not all created equal."
Copyright © 2022 Skynet Today, All rights reserved.