OpenAI The global rollout of a new language model that can be used through ChatGPT and replaces GPT-3.5 in the chatbot, the only one it had when it was launched in November 2022 and to which were later added GPT-4 and GPT-4o. GPT-4o Minithe new model, is multimodal like its older brother and can interpret both text and images, as well as generate them using DALL-E. It is a lighter and cheaper model for developers and for OpenAI itself and has as one of its objectives facilitate their access to this technology so that they can use it in their applications through the API.
GPT-4o Mini is the company’s first model to use a technique called ‘hierarchy of instructions’ which causes an AI to prioritize certain instructions over others. According to OpenAI, this makes it difficult to perform prompt injection attacks or other techniques that subvert the guidelines of the system.
GPT-4o Mini is gratuitous for users of the standard version of ChatGPT and is now being rolled out to these and subscribers of ChatGPT Plus and Teamwhile those of ChatGPT Enterprise Coming next week, GPT-3.5 will no longer be an option for ChatGPT users, but will still be available to developers via the API.
We’re continuing to make advanced AI accessible to all with the launch of GPT-4o mini, now available in the API and rolling out in ChatGPT today. https://t.co/sTxtOfUapJ
— OpenAI (@OpenAI) July 18, 2024
OpenAI has noted that it is a model intended for simple tasks that can be executed at a lower cost than the previous ones, but that offers superior results to GPT-3.5. In the Massive Multitasking Language Comprehension TestMMLU for its acronym in English, GPT-4o Mini achieved a score of 82 percent. In this test, GPT-3.5 Took out a 70%, GPT-4o a 88.7% and, according to Google, Gemini Ultra has achieved the highest score in history with a 90%MMLU is a benchmark exam for assessing AI capabilities consisting of about 16,000 multiple-choice questions across 57 academic subjects, but it is not a third-party verified test.
OpenAI is not the first company to release a smaller version of an existing language model. It is a common practice in the AI industry by companies such as Meta, Google and AnthropicThese language models, also cheaper, are usually aimed at API users an acronym for Application Programming Interface, which pay a fixed price per input and output token for use the models in your own applications. In this case, offering GPT-4o Mini for free as part of ChatGPT It will also save OpenAI money.
OpenAI API product manager Olivier Godement told Bloomberg that “in our mission to enable the cutting edge, to build the most powerful and useful applications, we of course want to continue to do the frontier models, to push the boundaries here. But we also want to have the best small models available.‘.
Large language models, LLMs, are usually smaller have fewer parameters than larger ones. This means that an LLM has a smaller neural network, which limits an AI model’s ability to understand context. Larger models are more powerful, but Fewer parameters also means fewer calculations required to run the model. and this is what allows, for example, Gemini Nano and not Gemini Pro or Gemini Ultra to run locally on a Pixel 8.
OpenAI’s goal with GPT-4o Mini is save costs and provide something lightweight and inexpensive for developers to build applications and tools that cannot be afforded with a larger, more expensive model like GPT-4. Its use is a 60% cheaper than the GPT-3.5 Turbo, the latest iteration of that model, and will support it in the future interpret and generate audio and video.