In the Power Rangers franchise, a group of teenage superheroes piloting robotic dinosaurs can merge their vehicles into a giant humanoid robot. ChatGPT has recently followed this path – yet not by becoming a skyscraper-sized robot that simply neverminds destroying a city while brawling with a giant rubber squid.
This month also provides a history of how LLMs have supported the schizophrenia diagnosis process and how DeepMind trains a neural network within the hallucinations of another neural network akin to Carthesian’s Brain in a Vat.
10.03.2023 DALL-E 3 Available in Bing Chat
DALL-E 3 is the latest image generation model designed by OpenAI, the company behind ChatGPT and several other groundbreaking AI technologies. Models designed and delivered by OpenAi are used by Microsoft tools. This month, Microsoft decided to provide all its users, not only the business or professional, with the latest image-generation model – even faster than OpenAI managed to include it in its own ChatGPT service.
Including the tool in ChatGPT or Bing Chat makes it way easier to use, enabling users to follow natural language and conversation patterns. According to Open AI, the system understands prompts better than before, as well as faster, and provides photorealistic quality.
More information can be found in this Microsoft Press Release.
10.05.2023 Microsoft introduces HoloAssist, an assistant-oriented dataset
In a recently published paper, “HoloAssist, an Egocentric Human Interaction Dataset for Interactive AI Assistants in Real World,” Microsoft tackled the idea of creating a dataset that combines verbal instructions with depictions of tasks done by humans. For example, a pair may be constructed of a simple instruction of “take the screwdriver” and a first-person view of a user grabbing the desired tool. This dataset may be able to help train real-life copilots that will suggest what to do in the real world using augmented reality headsets or similar tools.
More details about the technology can be found in this Microsoft press release.
10.03.2023 JPMorgan CEO: AI will be used everywhere
In a recent interview by Bloomberg, JPMorgan CEO Jamie Dimon emphasized his strong belief that AI will transform the way banks do business, impacting every single process, including hedging, error processing, and research. The CEO said that the new technology will augment humans and significantly boost their performance.
The interview can be found below:
10.09.2023 Large Language Models may help to diagnose schizophrenia
According to research performed by a team from University College London, Large Language Models (LLM) can be used to support the early schizophrenia diagnosis process. Contrary to somatic diseases, mental ones are rarely possible to diagnose using common means like blood tests, MRIs, or other comparable tools. Despite thousands of years of medical development, the main and best way to diagnose mental disease is still to talk with patients. Researchers from UCL have shown that a properly tuned LLM can spot subtle signs in the way a person speaks, signaling the early development of the disorder. More can be found on the UCL website.
10.10.2023 Increasing the Environmental Impact of AI
According to a research paper published in Joule, enriching every Google search with generative AI capabilities may increase power consumption tenfold. The research challenges the established approach of considering the AI training process to be the most power-hungry element of the cycle. The researcher has shown that the dawn of AI has resulted in a surge in demand for server chips manufactured by Nvidia, one of the globe’s leading chip providers. More can be found on Sciencedirect.com.
10.17.2023 DeepMind replaces simulators with Generative AI
The Brain in a Vat returns. Joint forces of researchers from DeepMind, UC Berkeley, MIT, and the University of Alberta have developed a generative AI neural network that works as an environment simulator for other agents or robots. These categories include, for example, autonomous vehicles and robotized domestic appliances.
So how does it work? UniSim is basically a live, on-demand video generator where the generated content is in response to actions taken by the learning system. For example, the image of a street with trees and pavement is generated when the car is driving through a city.
More about this technology can be found in this Arxiv paper.
10.17.2023 Baidu claims to turn people into “AI marketing teams” using ERNIE 4.0
Chinese tech giant Baidu has released a new version of its generative AI model, ERNIE. One of its new features includes Qingduo, a digital marketing and creation tool comparable to Canva or Adobe Cloud Stock. According to the company’s CEO, Robin Liu, this feature enables a single specialist to become an “AI Marketing team.”
ERNIE itself is a tool fully comparable with ChatGPT premium or Bing Chat, allowing users to not only have a Turingproof conversation with a machine but also to generate requested content, including not only text but also images or videos. More about the new feature can be found on VentureBeat.
10.18.2023 Stanford found out how transparent AI models really are
Stanford University’s Center for Research on Foundation Models (CRFM) has done extensive research on model transparency. The research is mind-boggling – no major model development meets the standards of model transparency, with Llama 2 from Meta gaining the highest mark (only 54%) followed by BLOOMZ from HuggingFace and GPT-4 from OpenAI.
The research is a clear indication that companies and organizations are handing an increasing number of tasks to virtual assistants and LLMs without assurance that there are no unexpected or unpredictable behaviors in the models themselves. More on the matter is available on the Stanford CRFM website.
10.23.2023 Nightshade can poison images to paralyze Generative AI models
Despite the general enthusiasm for generative AI models, there is strong opposition to the new solutions. It usually comes from people who are directly threatened by AI-powered solutions, mainly in creative fields. The situation is getting worse with legal doubts regarding the training material used by companies.
Nightshade is a tool delivered by a research team from the University of Chicago. It enables one to “poison” an image by altering pixel properties that are unseen by the human eye. Yet, with enough alterations, the image becomes ineligible for Artificial Intelligence models. Moreover, even a handful of the poisoned images scraped and digested by generative AI are enough to destabilize the model.
More about Nightshade can be found in this Arxiv paper.
10.23.2023 EU struggles to reach agreement on AI Act
The EU is still working on adopting the AI Act – a law designed to tackle the challenges of delivering AI-powered systems and the way these solutions will influence the lives of EU citizens. The law is sometimes compared to GDPR – the general framework that regulates the way personal data is processed by companies, and companies will pay for failing to comply with the new law.
The new regulation was to be adopted by the end of the year, yet there are still uncertainties between member nations, reducing the probability of reaching an agreement on time and postponing the new law to the year 2024. More about the situation can be found in this Reuters report.
10.25.2023 Amazon launches a generative AI-based advertisement product
The product enables merchants selling on Amazon’s platform to generate better ads. The user needs only to prompt with information about a desired background for a product image. One can add, for example, a kitchen as the background for a kitchen appliance or an exotic beach for a towel.
10.30.2023 ChatGPT goes MegaZord
In the Power Rangers franchise, a group of teenage superheroes piloting robotic dinosaurs can merge their vehicles into a giant humanoid robot. ChatGPT has recently followed this path – yet not by becoming a skyscraper-sized robot that simply neverminds destroying a city while brawling with a giant rubber squid. The service has enabled users to use all additional features, like analyzing files or generating images, in a single conversation. From now on, it can, for example, scan a PDF version of a book, synthesize the descriptions of main characters, and generate images based on those descriptions.
The features listed above are available for ChatGPT business and premium users.
10.30.2023 Midjourney, Stability AI, and DeviantArt win court battle with artists over copyright violation
A battle over the rights to use images to train image-generating models ended with artists’ claims being dismissed. A judge of the Northern District of California ruled that the artists’ claims were invalid due to the fact that they had never filed a copyright with the U.S. Copyright Office.
More on the story can be found on VentureBeat.