Vidita Chandra | 16 May 2024 02:55 PM
The tech geniuses at OpenAI have conjured up the latest large language model (LLM) GPT-4o and it (sure) is causing quite the buzz. Billed as the company's fastest and most powerful AI model so far, GPT-4o is gratis, unlike OpenAI’s most advanced (yet) LLM GPT-4 that only deep-pocketed users could summon. So we've compiled a comprehensive roundup, servin' you deets on what it can do and how you can hop on the GPT-4o train!
First, the "o" in GPT-4o stands for "Omni" and it is a multimodal AI model – a steep leap from its predecessors. As per OpenAI, the model is designed to elevate human-computer relations, you can feed it combinations of text, audio, or visuals, and it'll provide you responses in the same format.
From the demos, GPT-4o appears to be an evolution of ChatGPT into a digital assistant capable of handling various tasks. Its abilities range from real-time translations to reading facial expressions and engaging in spoken dialogues, surpassing its contemporaries. GPT-4o can interact using both text and visuals, it can analyse screenshots, photos, documents, or charts and discuss them. According to OpenAI, this updated version of ChatGPT will also feature improved memory capabilities, enabling it to learn from previous conversations with users.
Large language models are the backbone of AI chatbots. Massive data sets are fed into the models so that they can learn and expand their capabilities. Departing from earlier approaches requiring multiple specialised models for different tasks, GPT-4o streamlines everything into one unified model trained across various data formats - text, visuals, and audio.
To explain this advancement, OpenAI's CTO Mira Murat highlights voice capabilities in previous models that cobbled together three separate models for transcription, language processing, and text-to-speech conversion. With GPT-4o, all those previously fragmented steps happen within its singular, natively multimodal architecture. Essentially, this unified design allows GPT-4o to process and comprehend inputs more holistically. For instance, it can simultaneously grasp tone, background noises, and emotional context from audio inputs.
In terms of features and capabilities, GPT-4o has speed and efficiency. It can respond to queries nearly as fast as humans in conversation, clocking response times between 232 to 320 milliseconds, as per OpenAI. This also marks a significant leap over prior models that lagged with response delays stretching to several seconds.
Another demo highlighted how GPT-4o tackled maths problems leveraging its visual processing capabilities. By analysing a basic maths problem shown visually, it could guide the user step-by-step in solving for the unknown variable X. When presented with code snippets highlighted on-screen, the GPT-4o could comprehend the code's functionality and provide suggestions to enhance it.
GPT-4o also comes with multilingual support, exhibiting significant strides in comprehending languages beyond English, widening AI's reach to global audiences. Plus, it features enhanced audio and vision understanding capabilities. During the live demo, the GPT-4o-powered ChatGPT solves a linear equation in real time as the user writes it on paper. It could also gauge the speaker's emotions via the camera and identify objects.
Instead of a full public release, GPT-4o's rollout is being done in stages. The text and image processing capabilities are already integrated into ChatGPT, with some functionalities accessible to free users. The audio and video capabilities, however, will be gradually introduced to developers and select partner organisations. This measured approach, as per OpenAI, will ensure each modality - voice, text-to-speech, vision - undergoes thorough vetting to meet rigorous safety standards before wider availability.