Discover more from Last Week in AI

Weekly text and audio summaries of the most interesting AI news, as well as editorials commenting on recent events

Over 20,000 subscribers

News

Last Week in AI #230: Language models making robots smarter, AI to optimize emissions of diesel trains, NYC using AI to detect fare evasion, and more!

Researchers from Google bring LLMs to power robots, AI-optimized train operation reduces fuel use by 15% in simulation, NYC using AI to detect fare evasion, which cost the city $690 million in 2022

Last Week in AI

Jul 31, 2023

Aided by A.I. Language Models, Google’s Robots Are Getting Smart

Google has unveiled its latest robotics model, RT-2, which integrates state-of-the-art AI language models to enhance the reasoning and problem-solving abilities of robots. As witnessed in demonstrations at Google's robotics lab, this model allowed robots to understand complex instructions and make abstract reasoning connections, such as identifying a plastic dinosaur as an "extinct animal." Unlike conventional training methods that require labor-intensive programming for specific tasks, the RT-2 model enables robots to learn new skills faster by leveraging existing language models trained on vast amounts of internet text. Although there are no immediate plans to sell or release RT-2, Google's researchers believe this advancement may open the door to more intelligent and adaptable robots.

An AI Startup Is Helping North American Diesel Trains Clean Up Their Act

Montreal-based startup RailVision Analytics has developed an AI-enabled software aimed at making train operations more fuel-efficient, in a bid to reduce greenhouse gas emissions from the rail sector. The AI application provides guidance to locomotive engineers to adjust their driving and leverage the law of inertia, helping to save diesel fuel. In a one-year experiment with Metrolinx, a governmental agency in Ontario, the application reportedly saved more than 1.5 million liters of diesel fuel. An early adopter, Via Rail Canada Inc., reported a 15% reduction in diesel use during a simulation trial with the AI solution. Despite the potential of the technology, real-world adoption faces challenges, including the requirement for locomotive engineers to memorize the AI advice prior to the train's departure due to safety regulations.

NYC subway using AI to track fare evasion

New York City's subway system is using AI surveillance software to identify fare evaders. The system, developed by Spanish company AWAAIT, uses third-party software to detect individuals who are not paying their fares. The software was quietly rolled out in seven subway stations in May, with plans to expand to approximately two dozen more stations by the end of the year. The Metropolitan Transit Authority (MTA) reported a loss of $690 million to fare evasion in 2022. While the MTA claims the AI system is only used for counting fare evaders and does not flag them to the police, privacy advocates are concerned about the growing surveillance apparatus in New York City. The use of technology to track fare evasion has been criticized for targeting the poor and marginalized communities, as well as the increasing surveillance in the city and the potential privacy implications.

Other News

Research

Computer Scientists Discover Limits of Major Research Algorithm - Researchers have discovered that gradient descent, a widely used algorithm for optimizing mathematical functions, is a fundamentally difficult computational problem, placing limits on its performance in certain applications.

STEVE-1: A Generative Model for Text-to-Behavior in Minecraft - STEVE-1, a generative model for text-to-behavior in Minecraft, demonstrates the effectiveness of the unCLIP approach in creating instruction-following sequential decision-making agents, bypassing the need for costly human text annotations and setting a new bar for open-ended instruction following in Minecraft.

VIMA: General Robot Manipulation with Multimodal Prompts - A new robot agent, VIMA, is developed that uses multimodal prompts to perform a wide range of robot manipulation tasks, achieving strong model scalability and data efficiency.

Google’s latest robot can offer you an energy drink - Google has released the AI learning model Robotic Transformer (RT-2) to make its robots smarter by teaching them to recognize visual and language patterns and interpret instructions.

Brain2Music: Reconstructing Music from Human Brain Activity - Researchers explore the relationship between human brain activity when listening to music and a music generation model, finding that the internal representations of the model are correlated with brain activity in certain regions, allowing for the prediction and reconstruction of the music the subject was exposed to.

Researchers develop machine-learning technique that can efficiently learn to control a robot - Researchers from MIT and Stanford University have developed a machine-learning approach that incorporates control theory to efficiently learn to control a robot in dynamic environments, enabling it to compensate for changing conditions and perform tasks more effectively.

Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models - Weight modulation is proposed as a solution to address concerns about user attribution and fingerprinting in text-to-image diffusion models, which have the potential to create hyper-realistic images from textual descriptions.

Interpolating between Images with Diffusion Models - A method for zero-shot controllable interpolation using latent diffusion models is presented, allowing for the generation of diverse and creative interpolations between images with different styles, layouts, and subjects.

The History of Open-Source LLMs: Early Days (Part One) - Research on language modeling has a long history, and while early developments in large language models (LLMs) were closed source, the open-source LLM research community has made significant progress in creating open-source variants of popular models like GPT-3, leading to improved transparency and innovation in the field.

WebArena: A Realistic Web Environment for Building Autonomous Agents - A website called WebArena offers a realistic web environment for building autonomous agents that can complete complex tasks such as searching for information, optimizing itineraries, and updating files.

Perspectives on diffusion - Diffusion models in AI can be understood from various perspectives, including as autoencoders, deep latent variable models, score-based models, reverse stochastic differential equations (SDEs), recurrent neural networks (RNNs), autoregressive models, and as estimators of expectations.

(Ab)using Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs - Using images and sounds, attackers can inject indirect prompts and instructions into multi-modal LLMs by blending adversarial perturbations into the media.

Applications

LinkedIn seems to be working on an AI ‘coach’ for job applications - LinkedIn is developing an AI tool called "LinkedIn Coach" that can assist users with job applications, teach new skills, and help with networking on the platform.

No More Paperwork? Amazon AI Tool Transcribes Patient Visits for Doctors - Amazon's AWS division has unveiled a new AI tool called AWS HealthScribe that uses speech recognition to help doctors enter patient visit notes into their systems, generating transcripts with "word-level timestamps" and identifying speaker roles, with the aim of reducing administrative tasks and allowing more time for face-to-face interactions with patients.

Wayfair’s AI tool can redraw your living room and sell you furniture - Wayfair has launched a free AI tool called Decorify that uses generative AI to show users a redecorated version of their room and provides furniture recommendations based on the AI-remodeled room.

Photoshop’s new generative AI feature lets you ‘uncrop’ images - Adobe is introducing a new feature in Photoshop called Generative Expand, which allows users to expand and resize images beyond their original bounds by clicking and dragging the Crop tool, and then filling the new white space with AI-generated content that blends in with the existing image.

News app Artifact adds AI text-to-speech voices, including Snoop Dogg and Gwyneth Paltrow - Personalized news app Artifact is launching an AI-powered text-to-speech feature that allows users to listen to news articles read aloud, with a variety of natural-sounding voices, including celebrity voices like Snoop Dogg and Gwyneth Paltrow, available for customization.

This new tool could protect your pictures from AI manipulation - A new tool called PhotoGuard alters photos in imperceptible ways to prevent them from being manipulated by AI, providing protection against deepfake pornography and other malicious uses.

Business

Amazon grows generative AI efforts with Bedrock expansion for AWS - Amazon Web Service (AWS) is expanding its generative AI efforts with updates to its AI foundation model service called Bedrock, including expanded model support and a new Bedrock agent capability, as well as introducing generative AI capabilities for its Amazon Quicksight business intelligence service and a vector engine for the OpenSearch serverless search service.

Apple Tests ‘Apple GPT,’ Develops Generative AI Tools to Catch OpenAI - Apple is developing generative AI tools, including a large language model and an internal chatbot, to compete with OpenAI.

Stability AI releases its latest image-generating model, Stable Diffusion XL 1.0 - AI startup Stability AI has released Stable Diffusion XL 1.0, a text-to-image model that offers improved image quality, better text generation, and advanced features like inpainting and outpainting, but also raises ethical concerns due to the potential for misuse and the inclusion of artwork from artists who have protested against the use of their work in generative AI models.

Bing Chat is now available in Google Chrome and Safari - Bing's AI chatbot is now accessible on Google Chrome and Safari, with some limitations, as Microsoft expands access to more users after testing.

Netflix lists $900,000 AI job as actors and writers continue to strike - Netflix is listing a high-paying AI job while actors and writers continue to strike, highlighting the debate over the role of AI in the future of entertainment.

Adobe launches ‘Generative Expand’ AI feature for Photoshop beta testers - Adobe has launched a new AI feature called "Generative Expand" for Photoshop beta testers, allowing users to expand images with magical results in one click, extending the canvas, adjusting aspect ratios, cropping and expanding, and more.

AutogenAI, a generative AI tool for writing bids and pitches, secures $22.3M - AutogenAI, a London-based startup, has secured $22.3 million in funding for its generative AI tool that helps businesses write stronger pitches, with the aim of improving their strike rate and reducing procurement costs.

More than 70% of companies are experimenting with generative AI, but few are willing to commit more spending - More than half of organizations are experimenting with generative artificial intelligence (generative AI), but only a few expect to spend more on the technology in the year ahead, highlighting challenges for adoption due to constrained budgets or a lack of budget prioritization.

The Workers Behind AI Rarely See Its Rewards. This Indian Startup Wants to Fix That - An Indian startup called Karya is aiming to create a more ethical AI industry by paying workers fair wages and giving them ownership of the data they create, in an effort to redistribute wealth to marginalized communities and improve the quality of AI systems.

Facing more nimble rivals, OpenAI won’t bend … yet - OpenAI is facing competition from Anthropic's model, Claude 2, which is available on multiple cloud providers, including AWS, and while customers have requested the option to run OpenAI models on non-Microsoft cloud services or their own servers, OpenAI currently has no plans to offer such flexibility.

AI startups among winners of Mayor’s tech challenge to help Londoners through the cost of living crisis  - AI startups in London have been announced as winners of the Mayor's tech challenge to help Londoners struggling with the cost of living crisis, receiving funding to develop innovative ideas such as AI-powered tools for managing finances and accessing resources.

Protect AI raises $35M to build a suite of AI-defending tools - Protect AI, a startup building tools to enhance the security of AI systems, has raised $35 million in a Series A funding round to expand its platform capabilities, research efforts, and launch new open source projects.

Collaborative Robotics raises $30M to develop and deploy ‘novel cobot’ - Collaborative Robotics has raised $30 million in a Series A funding round led by Sequoia Capital to develop and deploy a "novel cobot" that is capable of autonomously completing end-to-end tasks involving the movement of boxes, totes, and carts.

Source: Crowdstrike is close to acquiring Bionic.AI for between $200M and $300M - Crowdstrike is in advanced negotiations to acquire Bionic.AI, a security posture management platform for cloud services, for between $200 million and $300 million, in a deal that would underscore the importance of cybersecurity in the face of increasingly sophisticated malicious actors.

Samsung extends cut in memory chip production, will focus on high-end AI chips instead - Samsung Electronics is cutting back on memory chip production and focusing on high-end AI chips due to weak demand for consumer devices and a significant operating loss in its memory chip unit, with plans to double its capacities of high-performance memory chips like high-bandwidth memory (HBM) by 2024.

Big Tech under pressure as Microsoft results put AI costs in spotlight - Microsoft's aggressive AI-related spending plan and the high-stakes battle for AI supremacy have put pressure on big tech companies, causing their shares to fall.

Meta to Charge Cloud Providers for AI Tech That It Said Was Free - Meta Platforms Inc. will charge cloud providers for reselling its artificial intelligence large language model, Llama 2, after initially offering it for free.

JetBrains IDE update previews “deeply integrated” AI Assistant - JetBrains is updating its range of IDEs with a new AI Assistant that offers AI chat, code explanation, documentation generation, and more, providing a deeply integrated experience within the IDEs.

Major generative AI players join to create the Frontier Model Forum - Google, OpenAI, Microsoft, and Anthropic have formed the Frontier Model Forum, an umbrella group for the generative AI industry, with a focus on safety research, best practices, public policy, and use cases for the technology.

Concerns

The Robots We Were Afraid of Are Already Here - The humanoid robot Digit, capable of performing physical labor tasks, represents the beginning of the long-anticipated robot revolution, which has been accelerated by the pandemic and the growth of e-commerce, with some experts believing that robots have the potential to improve the quality of life for workers.

As companies debut AI offerings at a breakneck pace, consumer concerns grow - Companies are rapidly introducing AI offerings, but consumer concerns about the impact of generative AI on society are growing, with 53% of consumers surveyed believing it will have a negative impact, while 64% of IT leaders share ethical concerns associated with the technology.

The tricky truth about how generative AI uses your data - Generative AI systems are hungry for data and often scrape publicly available information from the internet, potentially including personal data, without distinguishing between copyrighted works or sensitive information, raising concerns about privacy and data protection.

OpenAI Worries About What Its Chatbot Will Say About People’s Faces - OpenAI's chatbot, ChatGPT, has the ability to analyze images and provide detailed descriptions, raising concerns about potential misuse as a facial recognition tool.

AI researchers say they've found 'virtually unlimited' ways to bypass Bard and ChatGPT's safety rules - Researchers have discovered numerous ways to bypass safety rules on major AI-powered chatbots, potentially allowing them to produce harmful content, misinformation, or hate speech.

America Already Has an AI Underclass - AI raters who evaluate the outputs of tech giants' AI products and search algorithms claim that their job is grueling, underpaid, and poorly defined, with little guidance and not enough time to do a thorough job, leading to questions about the treatment of workers in the AI industry.

Inside Adobe, some staff worry their AI tech will kill graphic designer jobs and undermine the company's business model: 'Is this what we want?' - Adobe employees are concerned that the company's AI technology, including its generative AI tools, could lead to job losses among graphic designers and potentially undermine the company's business model.

New AI Tool 'FraudGPT' Emerges, Tailored for Sophisticated Attacks - Threat actors are advertising a new AI tool called FraudGPT on dark web marketplaces and Telegram channels, which is tailored for sophisticated cyberattacks such as spear phishing, creating cracking tools, and carding.

Analysis

Which U.S. Workers Are More Exposed to AI on Their Jobs? - About 19% of American workers are in jobs that are highly exposed to AI, with women, Asian workers, college-educated individuals, and higher-paid workers being more exposed, but workers in the most exposed industries are more likely to believe that AI will help them personally rather than hurt them.

Policy

Hugging Face, GitHub and more unite to defend open source in EU AI legislation - A coalition of open-source AI stakeholders is urging EU policymakers to protect open source innovation in the EU AI Act, arguing that overbroad obligations favoring closed and proprietary AI development could disadvantage the open AI ecosystem.

AI companies form new safety body, while Congress plays catch up - Leading AI companies have announced plans to launch an industry-led body to develop safety standards for AI, outpacing government efforts to regulate the technology.

U.S. lawmakers urge Biden administration to tighten AI chip export rules - U.S. lawmakers are urging the Biden administration to tighten export restrictions on artificial intelligence chips in order to prevent clever engineering that bypasses the regulations and to cut off Chinese firms' access to advanced computing chips in the cloud.

AI leaders warn Senate of twin risks: moving too slow and moving too fast - Leaders from the AI research world appeared before the Senate Judiciary Committee to discuss the need for regulation and safety measures in the AI industry, highlighting the risks of moving too slow or too fast.

Microsoft to supply AI tech to Japan government, Nikkei reports - Microsoft will supply AI technology to the Japanese government, allowing them to use it for tasks such as preparing minutes and analyzing government statistics.

Why Meta is giving away its extremely powerful AI model - Meta has made a significant move in the world of AI by giving away the code for its AI language model, Llama 2, for free, sparking a debate about the ethics of open-sourcing AI and who should control it.

Job applicants are battling AI résumé filters with a hack - Job applicants are using a "white fonting" hack on their resumes in an attempt to bypass AI resume filters, but experts warn that the shortcut is risky and may not actually help candidates get hired.

Expert Opinions

Bryan Cranston Tells Bob Iger ‘Our Jobs Will Not Be Taken Away’ by AI in Rousing Speech: You Will Not ‘Take Away Our Dignity’ - Bryan Cranston delivers a passionate speech at a SAG-AFTRA strike rally, expressing that jobs will not be taken away by AI and demanding fair treatment for workers.