Sam Altman "gpt-4 now significantly smarter" | OpenAI Updates GPT-4 and Reveals Open Source Evals

OpenAI has introduced GPT-4 Turbo, which is more conversational and smarter, alongside an open-source library for evaluating language models to increase transparency. The release of GPT-4 Turbo has shown improvements in various tests and has quickly regained the top spot in the chatbot Arena rankings, despite ongoing debates about the relevance and reliability of such tests.

OpenAI has released GPT-4 Turbo to improve user experience in writing chatbot responses. Sam Altman mentioned that GPT-4 is now significantly smarter and uses more conversational language. They have also open-sourced a lightweight library for evaluating language models to provide transparency about accuracy numbers. This library will be used to publish results alongside the latest models, such as GPT-4 Turbo released on April 9th, 2024.

The new GPT-4 Turbo model has shown significant improvements in various well-known tests, except for the MML test which has some known issues. However, there is ongoing debate about the relevance and reliability of such tests in determining the overall quality of a model. Different individuals have their preferred methods of testing model performance, but there is no universal standard test that everyone trusts.

The chatbot Arena, where users from around the world blind test different models to determine the best one, recently saw CLA 3 Opus overtaking GPT-4 as the number one model. Nonetheless, the newly released GPT-4 Turbo has quickly reclaimed the top spot, indicating its superior performance in the arena rankings. The rankings are based on user preferences and votes, with GPT-4 Turbo currently leading by a narrow margin.

There is a lack of standardized methods for testing language models, leading to variations in the formulations used in recent tests. OpenAI’s open-sourced library emphasizes the zero-shot Chain of Thought setting, where models are tested without prior examples or prompts. This technique aims to provide a better reflection of a model’s real-world performance and how it will be used by ordinary users.

In other news, OpenAI has dismissed two AI safety researchers for allegedly leaking information, including one with ties to the effective altruism movement. The researchers were accused of leaking information, which included details about Chief scientist Ilia Sare. This incident has raised concerns about data security and confidentiality within the organization.