Top News

Learning to Play Minecraft with Video PreTraining (VPT)

OpenAI announced a recent research project in which they trained an AI agent to play Minecraft. Specifically, the agent is capable of completing a challenging task that no previous AI systems achieved - making diamond tools. OpenAI achieved this by training its AI agent on very large datasets, including 70,000 hours of unlabeled Minecraft videos + 2000 hours of labeled videos produced by contractors who were paid to play the game. While the 70k unlabeled data gave the agent the ability to make wooden sticks, the 2k labeled videos gave the ability to make stone pickaxes, and finally, additional Reinforcement Learning led the agent to learn how to make a diamond pickaxe, which takes 24k+ gameplay actions in 20 minutes.

Our take: Players familiar with Minecraft know that making diamond tools is difficult as it requires sequencing many different subtasks, so this is an impressive achievement, especially given that the agent acts directly on pixel information and outputs keyboard/mouse controls, like a human player. The method, which OpenAI calls Video Pretraining (VPT), gives new evidence that scaling up data is a promising approach to unlocking new AI capabilities. It’s important to note that the agent is still very limited in what it can do and can’t seem to handle easy “edge” cases (see Twitter thread below). Still, it will be exciting to see what such approaches can do couple of years in the future, or even just one year from now.

Karolis Ramanauskas @Karolis_Ram

I’ve had early access to the new @OpenAI @Minecraft model for some time and had a blast playing around with it. The agent is way better than anything we’ve had before in RL. Here are some highlights … and some derps … and some of me trying to break it. 1/n

OpenAI @OpenAI

We trained a neural network to competently play Minecraft by pre-training on a large unlabeled video dataset of human Minecraft play and a small amount of labeled contractor data. https://t.co/a2pyBqvLvg https://t.co/XbqtwQSTwU

Instagram is testing an AI tool that verifies your age by scanning your face

If users attempt to edit their date of birth from under 18 to 18 or over on Instagram, they will be required to verify their age. This can be done using one of the three methods: 1) uploading a government ID, 2) asking a mutual friend, or 3) uploading a video selfie to Yoti, a company that uses computer vision algorithms to estimate age from face videos. According to the blog post by Instagram, Yoti and Meta both delete the video selfie once the algorithm verification is complete. Moreover, they claim that Yoti’s algorithm can only identify the age and not the identity of the user.

Our Take: Age verification on the platform is important for social media both for protecting the mental health of minors and for staying compliant with various regulations. However, using AI to estimate age based on facial features is not a foolproof solution yet. For instance, as stated in the article, the algorithm would predict a person to have a younger age if they didn’t have glasses. Moreover, users have reported that they are inaccurate when used on female faces with dark skin. So, it is appropriate to say that these algorithms aren’t mature enough to be used. There’s also the trust issue around data privacy. Many users already do not trust Instagram with their data, and using another company for age verification might make the trust issue worse.

Amazon launches CodeWhisperer, a GitHub Copilot-like AI pair programming tool

One of the developer-facing capabilities of large language models is the ability to assist people in writing code. GitHub Copilot, trained on billions of lines of code, used OpenAI's Codex to deliver this ability. Now, Amazon has a rival. Amazon CodeWhisperer is an AI pair programming tool "that can autocomplete entire functions based on only a comment or a few keystrokes."

Our Take: It's not a surprise that there's competition on the AI pair programmer front. Amazon has even introduced some interesting features to CodeWhisperer, such as continuously examining a user's code and taking their style into account. I think one of the most important features is that CodeWhisperer will include the license of functions it generates that are close to those in its training data. I think it's great that multiple companies are stepping onto the scene with these sorts of tools, which I hope will help ensure (a) that no one company is overly dominant and (b) encourage better practices in the development of those tools.

Other News

Research

Pathways Autoregressive Text-to-Image Model - "As we discuss at greater length in the paper, text-to-image models introduce many opportunities and risks, with potential impact on bias and safety, visual communication, disinformation, and creativity and art."

AI Detects Autism Speech Patterns Across Different Languages - "Machine learning algorithms help researchers identify speech patterns in children on the autism spectrum that are consistent between different languages."

UCF Researchers Develop Technology for AI that Mimics the Human Eye - "University of Central Florida researchers have developed a device for artificial intelligence that mimics the retina of the eye. The development could lead to advanced AI that can instantly recognize what it sees, like automatic descriptions of pictures taken by a camera or phone."

Robot can find keys in a bag just by listening as it rummages around - "A robotic arm with an attached microphone learned to locate objects in a bag based on the sound they make, including a ring of keys and a bag of crisps."

New hybrid machine learning forecasts lake ecosystem responses to climate change - "Sugihara, Boston University's Ethan Deyle, and three international colleagues spent five years searching for a better way to forecast and manage Lake Geneva's ecological response to the threat of phosphorus pollution, to which the effects of climate change must now be added"

Single MRI scan promises to diagnose early- and late-stage Alzheimer’s - "Using machine learning, researchers have developed an algorithm that can accurately diagnose Alzheimer’s disease from a single MRI brain scan."

Machine learning could minimize quantum tunnelling in transistors – Physics World - "Two researchers in China have shown how unwanted quantum tunnelling in field-effect transistors (FETs) could be suppressed by controlling the lattice orientations of materials used in the devices."

A WaveNet Rival? Stanford U Study Models Raw Audio Waveforms Over Contexts of 500k Samples - "The effective modelling of long-term dependencies enables conditioning new model outputs on previous inputs and is critical when dealing with longer text, audio or video contexts."

Meta releases largest open source AI language model to date - "Meta is now releasing the larger OPT-66B language model, following the OPT 125 million to 30 billion parameter models (see news below). According to Meta, it is the largest open-source AI language model accessible without restrictions to date."

By Exploring Virtual Worlds, AI Learns in New Ways - "Intelligent beings learn by interacting with the world. Artificial intelligence researchers have adopted a similar strategy to teach their virtual agents new skills."

Applications

General Motors is using AI to speed up the vehicle inspection process - "General Motors is bringing artificial intelligence into the vehicle inspection process."

Cosmo Let AI Design Its Latest Magazine Cover - "The publication tapped a cutting-edge image generator."

IRS expands AI-powered bots to set up payment plans with taxpayers over the phone - "The Internal Revenue Service is handling more of its call volume through automation, which gives its call-center employees more time to address more complex requests from taxpayers."

Graph Machine Learning at Airbnb - "Many real-world machine learning problems can be framed as graph problems. On online platforms, users often share assets (e.g. photos) and interact with each other (e.g. messages, bookings, reviews). These connections between users naturally form edges that can be used to create a graph."

How Microsoft and Google use AI red teams to “stress test” their systems - "It was a snowy day in February, and Amanda Minnich was attacking an AI system. With one block of code—and no other details—she needed to hack into one of the most complex machine learning systems operated by a Microsoft partner."

The King’s Swedish: AI Rewrites the Book in Scandinavia - "If the King of Sweden wants help drafting his annual Christmas speech this year, he could ask the same AI model that’s available to his 10 million subjects."

AI’s Benefit to Solar and Wind Power Isn’t Hot Air - "Solar and wind power are flourishing, but the world’s transition to renewable electricity remains too slow to meet climate goals. Harnessing the wind and sun on a global scale is easier said than done, for many reasons."

Earthquake tech could limit deaths. Afghanistan shows it’s not easy. - "Researchers are building algorithmic tools to better predict earthquakes and give people more time to flee. But structural challenges remain."

Bunnings, Kmart and The Good Guys using facial recognition technology to crack down on theft, Choice says - "Bunnings, Kmart and The Good Guys have been using facial recognition technology in a bid to crack down on theft in-store, research by Choice says."

Business

Cruise’s driverless autonomous cars start giving rides to paying passengers - "The era of commercial autonomous robotaxi service is here — Cruise officially became the first company to offer faired rides to the general public in a major city as of late Wednesday."

Microsoft Plans to Eliminate Face Analysis Tools in Push for ‘Responsible A.I.’ - "The technology giant will stop offering automated tools that predict a person’s gender, age and emotional state and will restrict the use of its facial recognition tool."

Copilot, GitHub’s AI-powered programming assistant, is now generally available - "Last June, Microsoft-owned GitHub and OpenAI launched Copilot, a service that provides suggestions for whole lines of code inside development environments like Microsoft Visual Studio."

IBM Reveals New AI and Cloud Powered Fan Experiences for Wimbledon 2022 - "Win Factors' brings enhanced explainability to 'Match Insights': Building on the existing Match Insights feature of the Wimbledon app and Wimbledon."

IKEA rolls out an AI-powered interactive design experience for shoppers - "With the new feature, U.S. customers can design and visualize their own living spaces with digitalized furniture on their smartphones instead of traveling to the brick-and-mortar store"

10 years of Amazon robotics: how robots help sort packages, move product, and improve safety - "It’s been a decade of discovery at Amazon. In 2012, we made a big investment to improve productivity in the supply chain by acquiring a robotics company called Kiva."

Cerebras Systems Sets Record for Largest AI Models Ever Trained on a Single Device - "Single CS-2 system trains multi-billion parameter NLP models including GPT-3XL 1.3B, GPT-J 6B, GPT-3 13B and GPT-NeoX 20B; provides simple setup, faster training and enables switching between models with a few keystrokes"

Exclusive: Grindr partners with Spectrum Labs on AI moderation - "Grindr is partnering with Spectrum Labs, tapping the startup's AI-based system to help filter postings on the LGBTQ dating service."

McDonald’s Expands Drive-Thru Voice Assistant Pilot After IBM’s McD Tech Acquisition - "McDonald’s has widened its drive-thru voice assistant pilot program to 24 stores, according to a note from BTIG analyst Peter Saleh."

Einride gets NHTSA approval to operate its autonomous electric trucks on US roads - "Autonomous EV freight company Einride has received approval from the National Highway Traffic Safety Administration (NHTSA) to operate its Autonomous Electric Transport (AET) trucks on public roads in the US."

Concerns

Senatskanzlei Berlin @RegBerlin

Es gab keine Anhaltspunkte dafür, dass die Videokonferenz nicht mit einer echten Person geführt wird. Allem Anschein nach handelt es sich um Deep Fake. Das LKA Berlin, Staatsschutz, ist eingeschaltet. https://t.co/SWgvn9JRNR

Senatskanzlei Berlin @RegBerlin

Zum angekündigten Gespräch mit Bürgermeister Klitschko: Der Verlauf des Gesprächs und die Themensetzung haben auf Berliner Seite ein Misstrauen hervorgerufen. Das Gespräch wurde vorzeitig abgebrochen. (1/3)

Vitali Klitschko fake tricks Berlin mayor in video call - "A video call between Berlin Mayor Franziska Giffey and a man pretending to be Kyiv Mayor Vitali Klitschko broke off as suspicion grew he was an imposter, the Berlin Senate Chancellery said on Friday."

“Sentience” is the wrong discussion to have on AI right now - "This article is part of Demystifying AI, a series of posts that (try to) disambiguate the jargon and myths surrounding AI. After reading a dozen different takes on the topic, I have to say that the media has become (a bit) disillusioned with the hype surrounding current AI technology."

Blake Lemoine Says Google's LaMDA AI Faces 'Bigotry' - "The question of whether a computer program, or a robot, might become sentient has been debated for decades. In science fiction, we see it all the time. The artificial intelligence establishment overwhelmingly considers this prospect something that might happen in the far future, if at all."

The dangerous fallacy of sentient AI - "Ignore the Silicon Valley messiahs. They only expose how little most of us know about the technology and its ethics."

China’s Surveillance State Is Growing. These Documents Reveal How. - "A New York Times analysis of over 100,000 government bidding documents found that China’s ambition to collect digital and biological data from its citizens is more expansive and invasive than previously known."

Neural networks built from biased Internet data teach robots to enact toxic stereotypes - "A robot operating with a popular Internet-based artificial intelligence system consistently gravitates to men over women, white people over people of color, and jumps to conclusions about peoples' jobs after a glance at their face."

The Power and Pitfalls of AI for US Intelligence - "Artificial intelligence use is booming, but it's not the secret weapon you might imagine."

Why LaMDA Is Nothing Like a Person - "Language models are no more sentient than your reflection in the mirror"

Alexa has a new voice — your dead relative’s - "Alexa will be able to mimic any person's voice based on short audio clips. But experts say the technology could bring security and ethical problems."

Policy

U.S. court will soon rule if AI can legally be an ‘inventor’ - "Can artificial intelligence (AI) be legally listed as an inventor? After all, if AI can legally invent products, the number of patents on drug-discovery tools will shoot up fast. The issue is currently before a United States court. The U.S."

Pentagon Unveils Plan to Make 'Responsible Military AI' More Than Just a Buzzword - "The Pentagon’s marching forward with AI weapons of war… responsibly."

Expert Opinions

Lessons learned from a PhD in Machine Learning - "This is Part 1 in the series. You can find Part 2 here and Part 3 here. At the time of this writing, I have recently finished my PhD in machine learning at ETH Zürich."

Fun

Wordalle is a guess-the-prompt game that combines Wordle and DALL-E - "I was lazy and sloppy so it took me four guesses to suss out the prompt for these images generated with text-to-image software. The first word was easy, and I'm sure you can guess what it is. The last word is also easy, but that's where I stumbled. Wordalle looks promising. I'm going to keep…"

Copyright © 2022 Skynet Today, All rights reserved.

Last Week in AI

Last Week in AI #173: AI learns to do tasks…