Circuit Verge

Tag: openai

  • OpenAI introduces the first AI model with the ability to engage in logical reasoning.

    OpenAI introduces the first AI model with the ability to engage in logical reasoning.

    OpenIA, famous for its ChatGPT, introduced a new Artificial Intelligence concept called o1, which focuses on logical reasoning tasks. o1 consists of multiple models, including o1-preview and o1-mini, which are accessible to ChatGPT or Team users.

    What sets 1 apart from ChatGPT?

    The O1 model, unlike ChatGPT, does not yet have the capability to search the internet or examine files. Its current version provides a constrained experience, with image analysis temporarily disabled for adjustments. Moreover, model usage is limited, with weekly caps set at 30 messages for o1-preview and 50 for o1-mini.

    Reprodução: YouTube
    Imagem: astrovariable/Burst

    The model offers an attractive pricing scheme for developers, with the o1-preview priced at $15 for every 1 million input tokens and $60 for every 1 million output tokens.

    OpenAI emphasizes the effectiveness of o1 in addressing mathematical and programming challenges, showcasing its capacity to consider options before answering queries. The model can integrate outcomes from different tasks, making it beneficial for tasks like sorting important emails in a legal inbox or creating marketing plans.

    Model 01 performs tasks of greater complexity.

    The first model was created using reinforcement learning methods, enabling it to think before giving a response through a series of internal reasoning processes. This development, as stated by Noam Brown from OpenAI, was achieved through a novel optimization algorithm and a dataset focused on reasoning, which consists of scientific literature tailored for model training.

    The 1 has shown significant progress, performing well in a qualification test for the International Mathematical Olympiad and scoring high in online programming competitions like Codeforces.

    O1 has some drawbacks, such as being slower in responding to specific inquiries, occasionally taking over 10 seconds. Moreover, OpenAI has chosen not to exhibit the full “thought processes” of the model, providing only condensed versions of these processes.

    OpenAI is expected to release the o1 to a larger audience and create updated versions capable of extended reasoning over hours, days, or weeks to enhance their abilities.

  • OpenAI recently revealed 4 new AI tools that are now accessible.

    OpenAI recently revealed 4 new AI tools that are now accessible.

    OpenAI revealed four additional functions of its developer-focused API at the DevDay conference in San Francisco, aiming to enhance accessibility and efficiency of AI-based products. The updates include improved customization, reduced costs, and advanced features in speech and computer vision technologies.

    Model Compression

    OpenAI’s new Model Distillation feature simplifies the optimization of smaller models like the GPT-4 mini by leveraging the output of larger models. The process, previously complex and multi-staged, can now be done more easily on the OpenAI platform. In a bid to promote its use, OpenAI is providing millions of free tokens until October’s end for developers to begin exploring distillation techniques.

    Prompt caching refers to the process of storing prompts for later use.

    OpenAI introduced Prompt Caching to help developers save money by enabling the reuse of prompts with lengthy prefixes without having to pay the full price each time. This feature is beneficial for applications utilizing AI models with precise and uniform directives. By identifying the same prefix within an hour, the cost of input is lowered by 50%, resulting in substantial savings for developers who frequently employ prompts.

    Vision Optimization

    Vision Fine-Tuning is a significant advancement that enables adjustments to be made to GPT-4 using both images and text, expanding its capabilities for visual search enhancements, object detection for autonomous vehicles, and medical image analysis. OpenAI has begun providing developers with complimentary tokens to evaluate the tool, with a charge per million tokens expected to be implemented starting in November.

    API that operates instantly

    The new Realtime API enables the development of applications that can engage in real-time conversations with users, processing audio instantly without transcription delays. This improves response quality by maintaining nuances like emphasis and accents, which is beneficial for quick interaction-based applications such as virtual assistants handling tasks like food orders or appointment bookings.

    Throughway