OpenAI宣布推出了其最新的深度學習模型GPT-4，這是其長期致力於擴展深度學習領域的最新里程碑。GPT-4是一個大型多模態模型，可以接受圖像和文本輸入，並生成文本輸出。雖然在許多現實場景下，GPT-4的表現還不如人類，但在各種專業和學術基準測試中，它展現出與人類相當的性能。例如，GPT-4在模擬的律師考試中取得了排名前10％的成績，相比之下，GPT-3.5的成績則在倒數10％左右。此外，OpenAI還開源了其框架OpenAI Evals，以便任何人報告模型的缺陷，以幫助指導進一步的改進。

在過去的兩年中，OpenAI重新構建了整個深度學習，並與Azure共同設計了一個為他們的工作的超級計算機。一年前，他們訓練了GPT-3.5作為第一次的“測試運行”，找到和修復了一些錯誤和提高了理論基礎。因此，GPT-4的訓練運行是空前穩定的，成為他們首個訓練性能能夠預測的大型模型。他們的目標是進一步預測並提前準備未來的能力，而且安全性至關重要。對於一般的對話而言，GPT-3.5和GPT-4之間的差異不大，但當任務的複雜性達到足夠的閾值時，GPT-4的可靠性、創造力和處理更多指令的能力就會表現出來。

OpenAI has announced the release of its latest deep learning model, GPT-4, which is a milestone in their long-term efforts to expand the field of deep learning. GPT-4 is a large, multimodal model that can accept both image and text input and generate text output. While its performance is still not on par with humans in many real-world scenarios, it has demonstrated human-like performance on various professional and academic benchmark tests. For example, GPT-4 achieved a top 10% ranking on a simulated lawyer exam, compared to GPT-3.5’s ranking in the bottom 10%.

In the past two years, OpenAI has rebuilt the entire deep learning infrastructure and worked with Azure to design a supercomputer for their work. A year ago, they trained GPT-3.5 as a “test run” and found and fixed some errors and improved the theoretical foundation. As a result, the training run for GPT-4 was unprecedentedly stable, making it their first large-scale model with performance that can be predicted. Their goal is to further predict and prepare for future capabilities, and security is paramount.

For general conversations, there is not much difference between GPT-3.5 and GPT-4, but when the complexity of tasks reaches a sufficient threshold, GPT-4’s reliability, creativity, and ability to handle more instructions will come into play. OpenAI has tested it on various benchmark tests, including simulated human-designed exams.