๐ค Gemini Ultra's Launch in the Crowded AI Arena
Dear curious minds,
Having various choices is beneficial, and the chatbot landscape is seemingly crowded with numerous contenders such as OpenAI, Anthropic, xAI, Mistral and Meta, to mention a few. Yet, when it comes to peak performance, GPT-4 from OpenAI has been the benchmark for some time. Google aims to disrupt this status quo with the introduction of Gemini Ultra, which has now been launched. It won't be long before we know whether it can match or even exceed the capabilities of GPT-4.
In this week's issue, I bring to you the following topics:
Build and Share GPTs for Free: Introducing Assistants on HuggingChat!
Tracing Art Origins: OpenAI's Image Watermarks
Gemini Ultra Launch: The Next Big Leap for AI?
If nothing sparks your interest, feel free to move on, otherwise, let us dive in!
๐ ๐ฅ Build and Share GPTs for Free: Introducing Assistants on HuggingChat!
HuggingFace has introduced a new feature called Assistants for their HuggingChat platform, a free alternative to ChatGPT, designed to rival custom GPT models.
New Assistant option in HuggingChat. Users can create an Assistant by filling out a simple form that requires choosing a model, naming the Assistant, and setting up a system prompt. The creation process is designed to be straightforward, and all Assistants are public by default, displaying the author's name.
So far the Assistants are rather basic as the main option to only option to adapt them is the system prompt. The highest voted feature on the announcement page is the implementation of RAG (Retrieval Augmented Generation), which is used to work with documents as knowledge base for your chatbot, and web search.
The open-source release of it, shows that HuggingChat assistants are not using OpenGPTs from LangChain but their own implementation
In the default option, HuggingChat shares your chats with the authors of the used models. However, you can opt out of this in the settings. More about this in the privacy statement.
My take: I am excited to see a free alternative for custom GPTs from OpenAI. However, the lack of advanced features, especially RAG, limits its potential as an alternative to OpenAI's custom GPTs. I am looking forward to future releases which change this.
๐ผ๏ธโ
Tracing Art Origins: OpenAI's Image Watermarks
OpenAI shared in a blog article that they now embed metadata in images generated with DALLยทE 3.
C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard designed to embed metadata in media content, enabling verification of its origin and history. This standard is being adopted across various industries, including AI, camera manufacturing, and journalism, to enhance trust and authenticity in digital media.
OpenAI's implementation of C2PA in DALLยทE 3 involves embedding specific metadata in images generated through ChatGPT and its API. This metadata allows users to verify if an image was produced by the DALLยทE 3 model, using tools like Content Credentials Verify. However, it's important to note that metadata can be removed, intentionally or unintentionally, which can affect the verification process.
The impact on file size and quality is minimal, ensuring that the addition of C2PA metadata does not affect the latency or the quality of the generated images.
Examples of embedded metadata show how images generated via the API and ChatGPT contain distinct signatures, indicating their creation through DALLยทE 3 and ChatGPT, respectively.
An image created with DALLยทE 3 in ChatGPT and checked with the content credentials page. Opening and saving an image, even without any modification, in the image editor GIMP removed the credentials according to the verification site content credentials referred to in the article from OpenAI.
The DALLยทE 3 image opened and saved in GIMP looses all OpenAI credentials.
My take: While OpenAI's implementation of C2PA metadata in DALLยทE 3 images is a noteworthy effort to boost media authenticity, I'm skeptical about its real-world effectiveness, particularly if the watermark can be easily removed simply by saving the image in a standard image editor. This loophole could potentially undermine many usage scenarios, as the ease of removal shows that it only proves that an image is generated with DALLยทE 3, but not the opposite.
๐๐ฎ Gemini Ultra Launch: The Next Big Leap for AI?
I covered the Gemini launch and a critical look at the launch material in earlier issues of this newsletter in December 2023.
Bard is now renamed to Gemini.
The update 2024.02.08 brings a new paid tier named Gemini Advanced, which is available in more than 150 countries. This includes, despite rumors telling differently, countries in the EU. I could subscribe from Germany.
You can now choose between Gemini (Gemini Pro Model) and Gemini Advanced (Gemini Ultra Model). [source] Google added a new Google One plan named โAI premiumโ which costs โฌ21.99 per month and can be used two months for free. The plan gives users access to Gemini Advanced, which currently uses the Gemini Ultra 1.0. Furthermore, it is stated that subscribers will soon be able to use Gemini in their tools like Gmail, Docs and more.
The new AI Premium plan in Google One lets you use the Gemini Ultra model. [source] There is also an official Android app which was released on the 5th of February. The URL has still Bard in its name, but the page it links to is already updated and titled with Google Gemini. You can choose to use Gemini instead of the known Google Assistant. However, the changelog of the latest Gemini release states that the app will for now not be available in most EU countries.
The previous update from the first of February added the capability to create images inside of Bard, or Gemini, how it is named today. This feature uses the text-to-image diffusion model Imagen 2. It is not directly usable from Germany, but activating a VPN pointing to the US makes it accessible. It is stated that there are invisible watermarks embedded at the pixel level.
Generating images in Gemini does currently not work without a VPN from Germany. There seems to be no difference in the quality of generated images with Gemini and Gemini Advanced. After generating the images above, my next request was to generate a title for suggestion for this article, but Gemini Advanced failed totally at it. I have to test more, but this is already a terrible first impression. ChatGPT delivered as expected.
Gemini failed and ChatGPT succeeded in create a title suggestion for this article. Even so, the first impression of Nick Dobos, which I have followed for quite some time on ๐, is underwhelming.
If you want to understand how Google handles your input and outputs, take a closer look at the Gemini Apps Privacy Hub. You can opt out of usage of your conversation as future training material, but, as with ChatGPT, you will not be able to access previous chats if you do so.
My take: I do not want to judge too early, but so far, I am not impressed by the new release from Google. My current perception of the AI space is perfectly visualized in the following image:
Establishments in the AI space visualized in a 2D plot setting their announcements and what they deliver in relation. Shared by Florian S on ๐. [source] This publication is free. If you would like to support me, please recommend this newsletter to anyone you think would enjoy it.
Disclaimer:ย This newsletter is written with the aid of AI. I use AI as an assistant to generate and optimize the text. However, the amount of AI used varies depending on the topic and the content. I always curate and edit the text myself to ensure quality and accuracy. The opinions and views expressed in this newsletter are my own and do not necessarily reflect those of the sources or the AI models.
