On March 15th, 2023, the world bore witness to a monumental event in the history of artificial intelligence as OpenAI launched its much-awaited GPT-4 language model. The excitement and anticipation surrounding this release are well-deserved, as GPT-4 boasts several impressive features and capabilities.
OpenAI has launched a powerful new image- and text-understanding AI model, GPT-4, which the company refers to as “the latest milestone in its effort to scale up deep learning.”
According to a recent summary by a speaker, GPT-4 has the ability to understand both images and text, making it a versatile tool for numerous applications. OpenAI's paying users can access it through ChatGPT Plus, and developers can register for the waitlist to access the API at a reasonable price:
GPT-4 - 8K context window (about 13 pages of text) will cost $0.03 per 1K prompt tokens and $0.06 per 1K completion tokens.
GPT-4–32k with a 32K context window (about 52 pages of text) will cost $0.06 per 1K prompt tokens and $0.12 per 1K completion tokens.
Priority Access: Developers can get prioritized API access to GPT-4 for contributing model evaluations to OpenAI Evals that get merged, which will help them improve the model for everyone.
One of the most exciting aspects of GPT-4 is its ability to generate text and understand images. It can even analyze the content of images and provide instructions for the AI's subsequent interactions. This opens up a whole new world of possibilities for image-based applications and AI-powered systems.
However, GPT-4 can only accept images as inputs and generate captions, classifications, and analyses. This functionality is still in the works.
Despite its impressive capabilities, the speaker notes that GPT-4 is not without flaws, and there is still room for improvement. Nonetheless, they believe that the model has tremendous potential to improve people’s lives by empowering a wide range of applications.
In terms of handling text, GPT-4 can process over 25,000 words, enabling use cases like long-form content creation, extended conversations, and document search and analysis.
GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on OpenAI's internal evaluations.
Partner Organizations:
Duolingo: GPT-4 deepens the conversation on Duolingo.
Be My Eyes: Be My Eyes uses GPT-4 to transform visual accessibility.
Stripe: Stripe leverages GPT-4 to streamline the user experience and combat fraud.
Khan Academy: Khan Academy explores the potential for GPT-4 in a limited pilot program.
Infrastructure:
GPT-4 was trained on Microsoft Azure AI supercomputers. Azure's AI-optimized infrastructure allows OpenAI to deliver GPT-4 to users worldwide.
Limitations:
GPT-4 still has several known limitations, such as social biases, hallucinations, and adversarial prompts, as it also has data cut off till 2021 that OpenAI is working to address. OpenAI encourages and facilitates transparency, user education, and wider AI literacy as society adopts these models. OpenAI also aims to expand the avenues of input people have in shaping our models.
GPT-4 significantly reduces hallucinations relative to previous GPT-3.5 models (which have themselves been improving with continued iteration). GPT-4 scores 19 percentage points higher than our latest GPT-3.5 on our internal, adversarial-designed factuality evaluations.
Conclusion:
OpenAI's release of GPT-4 marks a groundbreaking achievement in the field of artificial intelligence, showcasing its unmatched capabilities in performing challenging professional and academic benchmarks. The model's remarkable performance in a wide range of natural language processing tasks, across multiple languages, without task-specific fine-tuning, has surpassed the performance of other existing large language models. Its predictable scaling has enabled accurate predictions of its loss and capabilities, while OpenAI remains committed to understanding and enhancing the safety and alignment of the model. Despite the new risks that come with the increased capabilities of GPT-4, it represents a significant step forward in the development of useful and secure AI systems, although further work remains to be done.
Note: The OpenAI research paper for GPT-4 is available for reading GPT-4 Technical Report.
Comentários