Open AI promises to make their AI open. Its in the name, but most of their models are closed source. Why?
#Not Open
true cobalt
sullen remnantBOT
Hi @true cobalt. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:
OpenAI’s name does suggest an emphasis on openness, and the company originally started with a strong commitment to open research. However, the knowledge sources provided do not contain an explicit statement from OpenAI about why most of their current models are closed source.
What is clear from OpenAI’s published policies and documentation is that they focus on responsible deployment, safety, and compliance with privacy laws. For example, their Sharing & Publication Policy highlights the importance of responsible research and the need to consider potential misuse and safety risks when sharing AI models. Additionally, their Terms of Use prohibit activities such as reverse engineering or extracting underlying components of their services, which further indicates a controlled approach to model access.
While OpenAI does share research, guides, and some open-source tools (like the OpenAI Cookbook), the core models themselves are not open source. The knowledge sources do not provide a direct explanation for this, but the emphasis on safety, responsible use, and compliance suggests these are key factors in their decision to keep most models closed.
If you are looking for a more detailed or official rationale, the knowledge sources do not contain enough information to provide a definitive answer.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
true cobalt
@sullen remnant Tell me, how good is the 4o image generation?
sullen remnantBOT
Let me check the knowledge base for you again. One moment! :books:
GPT-4o’s image generation is powered by the latest multimodal model, gpt-image-1, which is described as state-of-the-art for high-fidelity image creation and editing. This model excels at following detailed instructions, rendering text within images, and leveraging real-world knowledge to generate lifelike and contextually accurate visuals. It can handle complex prompts, perform granular edits, and even add or modify text within images. Users can also upload existing images for editing, and the model supports transparent backgrounds and various output customizations such as size, quality, and format.
Some limitations remain, such as occasional struggles with precise text placement, maintaining visual consistency across multiple generations, and exact composition control for structured layouts. Complex prompts may also take up to two minutes to process. Despite these, GPT-4o’s image generation is considered a significant improvement over previous DALL·E models, especially in instruction following and contextual awareness Creating images in ChatGPT GPT Image API Image generation guide.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
true cobalt
@sullen remnant I know the sources won't say anything about Dall E 4, but what new features do you think it could add?
sullen remnantBOT
Let me check the knowledge base for you again. One moment! :books:
I'm sorry, but the knowledge sources do not contain any information about DALL·E 4 or potential features it might include. There are no official statements, leaks, or roadmaps regarding future versions of DALL·E beyond what is currently available (DALL·E 3 and gpt-image-1). As such, I cannot provide any details or speculation about what new features DALL·E 4 could add.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
true cobalt
@sullen remnant How smart is GPT 4?
sullen remnantBOT
Let me check the knowledge base for you again. One moment! :books:
GPT-4 is considered a highly intelligent language model, significantly more capable than its predecessors. It excels at advanced reasoning, following complex instructions, and demonstrating creativity. GPT-4 outperforms previous large language models and, as of 2023, most state-of-the-art systems on benchmarks such as MMLU (a suite of multiple-choice questions covering 57 subjects), not only in English but also in other languages. This makes it particularly strong in complex reasoning situations, coding, math, science, and creative tasks How to work with large language models Parse PDF docs for RAG.
However, it is important to note that while GPT-4 is much more capable in complex scenarios, for many basic tasks, the difference between GPT-4 and earlier models like GPT-3.5 is not significant Parse PDF docs for RAG.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
true cobalt
@sullen remnant Alright, how does Dall E 2 work?
sullen remnantBOT
Let me check the knowledge base for you again. One moment! :books:
DALL·E 2 is an AI system developed by OpenAI that creates realistic images and art from a natural language description. It works by taking a text prompt and generating an image that matches the description. DALL·E 2 supports several key features:
- Image Generation: Given a prompt, DALL·E 2 generates a new image from scratch.
- Image Editing (Inpainting/Outpainting): You can edit parts of an existing image by providing a mask (a PNG where transparent pixels indicate the area to replace). The model will regenerate the masked area based on the prompt and the surrounding image context.
- Image Variations: DALL·E 2 can create variations of a user-provided image.
DALL·E 2 is accessible via the OpenAI Image API, specifically through the Generations, Edits, and Variations endpoints. It supports image sizes of 256x256, 512x512, and 1024x1024 pixels. The model is known for offering more control in prompting and supporting more concurrent requests compared to later models. It is also lower in cost and supports inpainting (editing with a mask), which is not available in DALL·E 3 OpenAI DALL·E 2 model documentation API DALL·E FAQ.
To use DALL·E 2 for editing, you provide an image and a mask, and the model will fill in the masked area according to your prompt. For image generation, you simply provide a text prompt, and the model returns a new image.