A massive deal between Reddit and Google has been overshadowed by the ground-breaking SORA model, which transforms prompts into short movies.
Also, this edition will cover how Apple is entering the AI race, as well as a new, more sophisticated model from Mistral. Last but not least, as Gemini goes open source, Google is eating its cake and having it.
02.08.2024 – Deepmind introduces more reasonable LLMs
And “reasonable” is derived from “reasoning.” It’s a new prompting framework designed by Google DeepMind and the University of Southern California. The framework is named “SELF-DISCOVER” and promises substantial enhancements in reasoning tasks, delivering up to a 32% performance increase compared to traditional methods. The framework enables the LLM to pick a reasoning method it considers fitting, including step-by-step thinking processes and more complex critical thinking.
The full paper on this subject can be found on Arxiv.
02.07.2024 Apple introduces MGIE
MGIE or MLLM-Guided Image Editing is a tool that enables one to modify and edit an image by prompting for the desired change and letting the tool do the magic. The model can be applied to more complex tasks, like modifying backgrounds or manipulating light levels in images.
More on this can be found in The Verge.
02.13.2024 Stability AI introduces Stable Cascade
Stable Cascade is a new generation text-to-image model following in the footsteps of Stable Diffusion. The model is expected to be more efficient and better performing than its predecessor, as Stable Cascade uses a slightly different architecture that is composed of three stages. One of the greatest improvements should be visible in the area of in-image text generation.
More can be found in Stability AI’s press release.
02.15.2023 OpenAI introduces SORA
SORA is a text-to-video model, introduced by OpenAI, that transforms text prompts into short videos. The model is not currently available publicly, however, due to a need to secure the model and safeguard it against misuse. OpenAI is currently working with a red team to leverage existing safety methods and build new ones.
Sora has generated an enormous buzz around the internet with its ability to create lifelike short videos that tend to omit the known limitations of other AI models, like confusing the number of fingers or teeth.
More about the model can be found in OpenAI’s press release.
02.27.2024 OpenAI claims the Times used “deceptive prompts” to fill its claim
According to OpenAI representatives, Times exploited a bug that is currently being fixed. The sued company stated that “people are not using AI products this way” and that Times was actually abusing the system. According to OpenAI, the media company was feeding the model articles and later asking it to produce verbatim passages. Times, on the other hand, considers this technique to be a good way to show that there was a copyright infringement.
The full story can be found in TheVerge.
02.26.2024 Mistral partners with Microsoft
The Paris-based AI company that delivered the renowned open-source model, Mistral, has unveiled its partnership with Microsoft. The partnership will channel $16 million into making the model available in the Azure environment. The partnership followed the unveiling of the latest Mistral model that aims to challenge GPT4’s dominance.
Read more in Venturebeat.
02.23.2024 Google apologizes for ahistorical AI images
Google has officially apologized for its Gemini tool producing historically inaccurate images. Company officials said that some images were “inaccurate, or even offensive.” According to the company statement, the goal of the tool was to produce images of diverse people when using open-ended prompts, yet this approach has led to a loss in historical context.
More can be found in VentureBeat’s story.
02.21.2024 Deepmind unveils open source model, Gemma
Two models, one with 2 billion parameters and the second with 7 billion parameters. The models are created using the same approach, research and technology that was earlier used to create the Gemini model that currently powers all of Google’s AI. Apart from models, the company provides users with tools to fine-tune the models to make them better fit for commercial uses.
More information can be found in the company’s statement.
02.21.2024 Adobe adds AI assistant to manage large PDF files
Adobe has added a generative AI-powered assistant to the Adobe Acrobat PDF reader to help make handling large documents easier and more efficient. The tech assistant can be asked about the contents of a document, to find particular information, or to summarize selected parts.
More about Adobe’s new AI assistant can be found in this press release.
02.22.2024 Reddit licenses its user-generated content to Google for AI training
A deal was struck, and just before, Reddits expected an IPO. According to Reuter’s sources, Google will pay $60 million per year for access to the data. Information about the deal was known before, yet Reuters was the first to name the buyer.
More can be found on Reuters.
