Arstecnica
On Tuesday, OpenAI announced GPT-4. GPT-4 is a large-scale multimodal model that can accept text and image inputs while returning text output that “shows human-level performance on a variety of professional and academic benchmarks,” according to OpenAI. Also on Tuesday, Microsoft announced that Bing Chat has been running on GPT-4 all along.
If it works as claimed, GPT-4 could represent the dawn of a new era in artificial intelligence. In its announcement, OpenAI wrote that it “passed the mock bar exam and scored in the top 10% of the test takers.” “In contrast, GPT-3.5 scores were in the bottom 10%.”
OpenAI plans to release GPT-4 text capabilities through ChatGPT and its commercial API, but initially there is a waiting list. GPT-4 is now available for ChatGPT Plus subscribers. The company is also testing the image capture capabilities of his GPT-4 with a single partner, Be My Eyes. Be My Eyes is a smartphone app that can recognize and describe scenes.
Benj Edwards / Ars Technica
GPT stands for “Generative pre-trained Transformer” and GPT-4 is part of a set of foundational language models dating back to the original GPT in 2018. Following the initial release, OpenAI announced his GPT-2 in 2019 and his GPT-3 in 2019. 2020. A further refinement, called GPT-3.5, is coming in 2022. In November OpenAI released ChatGPT. At the time, this was a fine-tuned conversation model based on GPT-3.5.
GPT-series AI models are trained to predict the next token (word fragment) in a sequence of tokens using large amounts of text, mostly pulled from the internet. During training, the neural network builds a statistical model representing the relationships between words and concepts. Over time, OpenAI has increased the size and complexity of each GPT model. As a result, it generally performed better than the model compared to how humans complete text in the same scenario, depending on the task.
In addition to the introductory website, OpenAI has also released a technical paper describing the capabilities of GPT-4 and a system model card detailing its limitations.
microsoft ace of holes
Orrich Lawson | Getty Images
The simultaneous announcement of GPT-4 by Microsoft means that OpenAI has been working on GPT-4 since at least November 2022, when Microsoft first tested Bing Chat in India.
“We’re happy to see the new Bing running on GPT-4 customized for search,” Microsoft wrote in a blog post. “If you’ve used the new Bing in preview at any point in the last six weeks, you’ve seen the power of OpenAI’s latest model early on. As OpenAI updates GPT-4 and beyond, , Bing will benefit from the following: These improvements will give users the most comprehensive Copilot capabilities available.”
The Bing Chat timeline matches an anonymous tip Ars Technica heard last fall. Internally, OpenAI said he was ready for GPT-4 but was hesitant to release until better guardrails were implemented. While the nature of Bing Chat’s attribution has been a matter of debate, GPT-4 guardrails now come in the form of more attribution training. Using a technique called Reinforcement Learning from Human Feedback (RLHF), OpenAI uses human feedback from his GPT-4 results to train a neural network and to learn whether OpenAI is sensitive or potentially harmful. Refused to discuss topics that they thought were.
OpenAI writes on its website: Refusing to go outside the guardrail. “
This is part of our breaking news and will be updated as new details emerge.
