 |
| Chat GPT-4 is Here |
The most recent step in OpenAI's endeavor to scale up deep learning is GPT-4, which we developed. GPT-4 is a sizable multimodal model that accepts image and text inputs and emits text outputs. While less effective than humans in many real-world situations, GPT-4 performs at a human level on various academic and professional benchmarks. For instance, it successfully completed a mock bar exam with a score in the top 10% of test takers, as opposed to GPT-3.5, whose score was in the bottom 10%. Our adversarial testing program and ChatGPT lessons were used to iteratively align GPT-4 over the course of six months, yielding our best-ever results (albeit still far from ideal) in factuality, steerability, and refusal to stray outside of set boundaries.
We completely rebuilt our deep learning stack over the last two years and co-designed a supercomputer for our workload with Azure from the bottom up. We trained GPT-3.5 as the system's initial "test run" a year ago. We identified several problems, rectified them, and strengthened our theoretical underpinnings. Our GPT-4 training run was consequently (for us at least!) unprecedentedly steady, making it our first large model whose training performance we were able to precisely forecast in advance. We intend to improve our methods as we keep our attention on dependable scaling in order to better foresee and plan for future capabilities, which is essential for safety.
Via ChatGPT and the API, we are making a text input for GPT-4 available (with a waitlist). It's been a while since I last posted. Also, we're making OpenAI Evals, our platform for automatically assessing AI model performance, publicly available so that anybody may report issues with our models and contribute to their ongoing development.
Capabilities
The distinction between GPT-3.5 and GPT-4 can be difficult to make in informal speech. When the difficulty of the task reaches a certain level, GPT-4 distinguishes itself from GPT-3.5 by being more dependable, inventive, and capable of handling far more complex instructions.
We tested on a range of benchmarks, including recreating tests that were initially created for humans, to understand the differences between the two models. For the purpose of this study, the term "responsible party" refers to a person who has the authority to act on behalf of another person. We didn't receive any special preparation for these tests. Although only a small portion of the exam difficulties was encountered by the model during training, we think the results are still accurate. For more information, see our technical report.
This blog is referring to the blog post on OpenAI's website where they talk about GPT-4.
VISIT THE FOLLOWING LINK TO LEARN MORE.
Comments
Post a Comment