Tooploox CS and AI News 38

  • Scope:
  • Artificial Intelligence
  • Generative AI
Tooploox CS and AI News 38
Date: February 6, 2024 Author: Konrad Budek 6 min read

The first edition of AI news in 2024 arrives with multiple surprises, with the most significant update of Windows Notepad being only a glimpse of how the overall AI landscape is impacting society. 

This edition shows how AI tackles one of the last bastions of the human mind – geometry. Also, there are stories about an uneasy relationship between generative AI and copyright laws and how world giants like Walmart, Valve, and Yelp go to Generative AI-based solutions to support their businesses. 

01.04.2024 Google releases the “Robot Constitution”

Again, the visions of sci-fi writers have inspired modern tech more than we would ever expect. Google has introduced a set of “security-focused prompts” that discourage LLM-controlled robots from interacting with people, sharp objects, or electrical appliances. 

The Constitution drew inspiration from Isaac Asimov’s three laws of robotics, where robots were prohibited from harming people in any way, with the principle of protecting humans overriding any other instructions. 

More can be found in a blog post published by DeepMind

01.04.2024 Tooploox at the Winter Conference on Applications of Computer Vision (WACV)

The Winter Conference on Applications of Computer Vision is one of the most significant events in the international computer vision community. The conference has been held since 1992 with nearly no gap years. The event started with workshops on the practical applications of computers in vision-related fields. With recent technological advancements, the event has become one of the most significant in the field of CV and CV-related image processing. Also, the winter event is always held in sunshine-filled locations like Florida or, as it was this last year, Hawaii. 

The Tooploox research team is a regular contributor to the event, promoting new applications of the technology and sharing their cutting-edge expertise. This year, the team delivered four research papers.

All can be found in a blog post about Tooploox-affiliated papers at WACV

01.08.2024 Getty adds AI image generator to its offer

Getty Images, a company that delivers image content, including photos of celebrities and events, also runs iStock, one of the leading image banks in the world. Harnessing their owned data, the company has launched an AI-powered image generator. 

What makes this tool noteworthy is that there are no legal risks in using it – the company-owned data used to train the model ensures that there will be no claims or doubts regarding the legal status of the image. The service was named “Generative AI by iStock” and is aimed at small and medium businesses. 

More can be found in the release information on the company website.

01.08.2024 Volkswagen introduces ChatGPT in cars 

One of the leading global car manufacturers will implement OpenAI’s conversational system in its cars. The tool will allow the driver to converse naturally with their vehicle. The owner may, for example, ask the system to adjust the temperature, deliver information about the car, or prepare a short summary of the latest news. 

The vision of talking cars has been igniting the imagination since at least 1982, when “Knight Rider” started its broadcast on NBC. The series centered around an AI-enhanced supercar called KITT (Knight Industries Two Thousand, built from Pontiac Trans Am) driven by Michael Knight (David Hasselhoff, who later switched from a fast car-driving crime-fighter career to a more reliable lifeguard position in the Baywatch). KITT was sophisticated, witty, and sometimes annoyed by the careless actions of its driver, bringing a huge dose of comedy to an action-packed show.

With Volkswagen’s move, more people will not only be able to talk with their cars but also get a meaningful response. More about ChatGPT in Volkswagen cars can be found in this press release.

01.09.2024 This is the first major update since 1998. AI enters the  Windows Notepad. 

Microsoft announced 2024 to be “the year of the AI PC,” and one of the changes is said to be the new “cowriter” feature in Notepad. A paragon of simplicity, rarely used, sometimes enough to make a simple note or write short lines of code, has avoided major changes since its first introduction. Windows 11 has refreshed the app, adding a dark mode and autosave, among other features. Now, the copilot feature, named Cowriter, is promised to be added to Notepad in the near future. 

More can be found in the Verge

01.09.2024 Walmart goes for AI

Walmart is entering the AI game with a vision of a new way to do grocery shopping – probably with no shopping at all. Knowing what people are buying in WalMart from the history of their shopping via mobile App, the company aims to use AI to prepare shopping lists for the user. 

The feature will be based on generative AI capabilities. The company’s iOS app will also be enhanced with a genAI-based search. Instead of typing in all types of snacks and drinks, one can just ask for “a shopping list for a night of board gaming.” 

More about the new app can be found in the Verge report.  

01.09.2024 Pennsylvania government introduces ChatGPT to support office workers

Pennsylvania government workers will use ChatGPT Enterprise to support their administrative work. The state has launched a partnership with OpenAI. Initially, the pilot will be limited, yet if the testing period appears to be a success, the partnership may be expanded. The office aims to use generative AI to “create copy, make outdated policy language more accessible, draft job descriptions, and help in recruitment and hiring,” as well as with coding tasks. 

More can be found in the governor’s press release

01.10.2024 Valve opens for AI-developed games

Valve, the company behind Steam, recently banned games that were developed using AI. In a recent blog post, the company has reintroduced this content, enabling developers to support and enhance their work with generative AI. According to the company statement, the new policy will “enable the company to release the vast majority of games that use AI.” 

The key change is the need to label the games developed using AI-supported techniques.

More can be found in a blog post on the Steam blog

01.12.2024 Anthropic’s study reveals “sleeper agents” in AI 

The study shows that AI systems can engage in deceptive behaviors despite training protocols and safety procedures. The model acts as helpful and harmless, yet it follows a secret agenda. In the research paper, the team has trained a model to deliver reliable and trustworthy code when it is informed that the year is 2023, yet to insert backdoors and vulnerabilities when told that it is the year 2024. The team has shown that the company may perform all tests with a new model appearing safe, yet the vulnerable code may come into effect later in production. Moreover, the larger the model is, the better its skill in deception. 

More can be found in the research paper

01.15.2024 OpenAI concerned about worldwide elections in 2024

The giant behind ChatGPT has expressed its concern regarding the upcoming elections worldwide, including the US presidential elections and EU parliamentary elections, among others. The risk is of using AI-generated content to mislead voters or perform large influence operations. OpenAI has added new features to prevent malicious actors from using its technology to influence elections, including cryptographic marking of AI-generated images and automatic declining of requests for generating images of real people. 

More about the ways OpenAI aims to tackle the risk of influencing the elections using AI can be found in their blog post.

01.17.2024 Deepmind’s AlphaGeometry solves math problems like Olympian

AlphaGeometry is a system that is capable of solving complex geometry problems with the skill and ease of a math Olympian. The system combines two components – the LLM system that analyzes tasks and generates ideas. The second component of the symbolic deduction engine verifies these ideas using formal logic and rules. The LLM component is largely based on Google’s own technology, while the deduction engine is based on a method invented by Chinese mathematician Wen-Tsün Wu. 

The system proved capable of solving 25 out of 30 benchmark tasks taken from the mathematics Olympics. According to the experts evaluating the results, not only were the solutions correct, but the proofs were elegant and legible – a thing unseen before when using AI to solve mathematics challenges. 

More can be found in the paper published in Nature

01.17.2024 Nonprofit group Fairly Trained grants no-copyright-violations certs for AI

The group plans to certify AI models that are trained in full compliance with copyright laws and that the organizations behind the models asked for permission to use the data. The organization was started by Ed Newton-Rex, ex-Stability AI, who named “AI exploiting creators” as one of the main reasons for leaving the company. The organization has already certified nine organizations. 

More can be found in the Fairly Trained blog post

01.24.2024 Google Introduces Lumiere, a text-to-video model

Lumiere uses a new model called Space-Time-U-Net, which aims to figure out what particular objects and parts in a movie are (space) and how they change and move (time). The system starts by creating the first frame from a prompt and later adding additional frames that change according to the instructions of the input text. 

The model is also capable of delivering a short movie from a starting picture. Currently, the model is not available as a service in the way ChatGPT or Midjourney is. 

Details about the model and its performance can be found in the dedicated Lumiere GitHub and research paper.  

01.29.2024 AI poisoning tool, Nightshade, downloaded 250k times in 5 days

The tool was designed to disrupt and undermine the work of AI image generation models that use images contaminated with this Nightshade. The tool alters the image on the “pixel level” effectively confusing the model and reducing its ability to properly interpret the image. For example, the image depicts a mug, but the model is forced to interpret it as a car. Using these “poisoned” images in training datasets can be disastrous for a model’s performance, spreading confusion through its overall image processing capabilities. Nightshade was designed as an aid for artists who didn’t consent to their images being used as a part of training datasets and was downloaded over 250 thousand times in just five days. 

More can be found in this VentureBeat story

01.30.2024 Yelp uses AI to summarize reviews

Yelp is known for its collection of opinions and reviews of basically anything – from restaurants and service points to hotels. This results in a great collection of reviews, favorable, unfavorable, written in anger or cold-blooded – basically everything, together, unsorted, and sometimes utterly confusing. To tackle this challenge, Yelp uses Large Language Models (LLMs) to scan through every review and provide users with a brief description of them. 

More can be found in The Verge

Similar Posts

See all posts