OpenAI has introduced GPT-4o, their latest flagship AI model, alongside other updates at the ‘Spring Update’ event, emphasizing their commitment to advancing AI while ensuring accessibility for all.
GPT-4o: Faster, Smarter, Multimodal
GPT-4o, dubbed “o for omni,” offers GPT-4-level intelligence but with remarkable speed improvements and expanded capabilities in text, voice, and video processing.
Notably, it excels in understanding and discussing images, enabling tasks like translating menus or explaining sports rules in real-time. Moreover, it supports over 50 languages across various functions, enhancing global accessibility.
- Model Evaluations: GPT-4o achieves top-tier performance in text, reasoning, and coding intelligence. It sets new standards in multilingual, audio, and vision capabilities.
- Language Tokenization: 20 representative languages benefit from the new tokenizer’s compression across various language families.
The model’s versatility allows seamless interaction with inputs and outputs in text, audio, and image formats. With response times comparable to human conversation, GPT-4o delivers high performance across languages and modalities, especially in video and audio comprehension.
OpenAI’s CEO, Sam Altman, praises the new voice (and video) mode as the best computer interface he’s experienced. He finds it remarkably real, reminiscent of AI from movies.
Altman appreciates its speed, intelligence, enjoyment, naturalness, and supportiveness, contrasting it with previous interfaces. He foresees an exciting future with features like personalization and task execution, where computers empower users like never before.
Safety and Limits
OpenAI emphasizes safety in GPT-4o, employing techniques like filtered training data and post-training refinements across its modalities.
Evaluation through their Preparedness Framework ensures medium risk levels in critical areas such as cybersecurity and model autonomy.
Extensive external assessments aid in identifying and mitigating risks, especially in novel areas like audio outputs, which will undergo gradual release with safety measures in place.
Enhanced Accessibility for Free Users
In line with their mission, OpenAI extends advanced AI tools to more users, including ChatGPT Free subscribers.
Features like GPT-4 level intelligence, web responses, data analysis, photo discussions, file uploads, access to GPTs, and Memory are now available. However, usage limits apply, with ChatGPT switching to GPT-3.5 upon reaching the threshold.
Additional Updates
- Desktop App: The new macOS app streamlines user workflow, enabling instant access to ChatGPT and seamless screenshot discussions.
- Voice Conversations: Users can engage in voice chats directly from the desktop app, with plans for audio and video enhancements in the future.
- macOS App Rollout: The macOS app debuts for Plus users, with a Windows version slated for release later.
- Improved Interface: ChatGPT sports a revamped, user-friendly interface for a more engaging experience.
Availability
GPT-4o’s capabilities are gradually rolling out, with text and image features already integrated into ChatGPT. Available in the free tier and offering higher message limits for Plus users, GPT-4o aims to enhance accessibility.
- Developers can access GPT-4o in the API, enjoying faster speeds and increased rate limits at half the price.
- Support for audio and video capabilities will soon extend to select API partners.
Speaking about GPT-4o and updates, Sam Altman, CEO of OpenAI, stated:
Our mission is to provide highly capable AI tools to people for free (or at an excellent price). I’m very proud that we’ve made the world’s best model available for free in ChatGPT, without any ads or similar distractions.
While we are a business and will monetize various aspects, this will enable us to offer exceptional AI services for free to (hopefully) billions of people.
Finally, a huge thanks to the team that dedicated so much effort to making this possible!