News from the AI & ML world
DeeperML - #gpt-4o
|
Maria Deutscher@SiliconANGLE
//
OpenAI has officially rolled out native image generation capabilities within ChatGPT, powered by its GPT-4o model. This significant upgrade replaces the previous DALL-E integration, aiming for more consistent results, fewer content restrictions and improved accuracy in interpreting user prompts. The new feature is available to all ChatGPT users, including those on the free tier, with API access for developers planned in the near future.
The integration of image generation into GPT-4o allows users to create detailed and lifelike visuals through natural conversation, making it easier to communicate effectively through visuals. GPT-4o can accurately render text within images, supports complex prompts with up to 20 different objects, and can generate images based on uploaded references. Users can refine their results through natural conversation, with the AI maintaining context across multiple exchanges - making it easier to iteratively perfect an image through dialogue. Early testing shows the system produces more consistent images than DALL-E 3. Recommended read:
References :
Jesus Rodriguez@TheSequence
//
OpenAI has recently launched new audio features and tools aimed at enhancing the capabilities of AI agents. The releases include updated transcription and text-to-speech models, as well as tools for building AI agents. The audio models, named gpt-4o-transcribe and gpt-4o-mini-transcribe, promise better performance than the previous Whisper models, achieving lower word error rates across multiple languages and demonstrating improvements in challenging audio conditions like varying accents and background noise. These models are built on top of language models, making them potentially vulnerable to prompt injection attacks.
OpenAI also unveiled new tools for AI agent development, featuring a Responses API, built-in web search, file search, and computer use functionalities, alongside an open-source Agents SDK. Furthermore, they introduced o1 Pro, a new reasoning model, positioned for complex reasoning tasks, comes with a high cost, priced at $150 per million input tokens and $600 per million output tokens. The gpt-4o-mini-tts text-to-speech model introduces "steerability", allowing developers to control the tone and delivery of the model. Recommended read:
References :
Chris McKay@Maginative
//
OpenAI has recently unveiled new audio models based on GPT-4o, significantly enhancing its text-to-speech and speech-to-text capabilities. These new tools are intended to give AI agents a voice, enabling a range of applications, with demonstrations including the ability for an AI to read emails in character. The announcement includes the introduction of new transcription models, specifically gpt-4o-transcribe and gpt-4o-mini-transcribe, which are designed to outperform the existing Whisper model.
The text-to-speech and speech-to-text tools are based on GPT-4o. While these models show promise, some experts have noted potential vulnerabilities. Like other large language model (LLM)-driven multi-modal models, they appear susceptible to prompt-injection-adjacent issues, stemming from the mixing of instructions and data within the same token stream. OpenAI hinted it may take a similar path with video. Recommended read:
References :
Megan Crouse@techrepublic.com
//
OpenAI has unveiled a suite of advancements, including enhanced audio models and a significantly more expensive AI reasoning model called o1 Pro. The new audio models, including gpt-4o-transcribe and gpt-4o-mini-transcribe, offer improved transcription capabilities compared to Whisper, although they are susceptible to prompt injection attacks due to their foundation on language models. Users can access these models via the Realtime API, enabling real-time transcription from microphone input using a standalone Python script.
OpenAI's o1 Pro comes with a steep price tag of $150 per million input tokens and $600 per million output tokens. This makes it ten times more expensive than the standard o1 model and twice as costly as GPT-4.5. While OpenAI claims o1 Pro "thinks harder" and delivers superior responses for complex reasoning tasks, early benchmarks suggest only incremental improvements. Access to o1 Pro is currently limited to developers who have spent at least $5 on OpenAI's API services, targeting users building AI agents and automation tools. Recommended read:
References :
Emilia David@AI News | VentureBeat
//
OpenAI is enhancing GPT-4o with improved instruction following and problem-solving capabilities. The company has updated GPT-4o to better handle detailed instructions, especially when processing multi-task prompts, thus improving performance and intuition. This model can be accessed by subscribers through the API as "chatgpt-4o-latest" and in ChatGPT.
OpenAI has announced its support for Anthropic’s Model Context Protocol (MCP), an open-source standard designed to streamline the integration between AI assistants and various data systems. With MCP, AI models can connect directly to systems where data lives, eliminating the need for custom integrations and allowing real-time access to business tools and repositories. OpenAI will integrate MCP support into its Agents SDK immediately, with the ChatGPT desktop app and Responses API following soon. This protocol aims to create a unified framework for AI applications to access and utilize external data sources. ChatGPT Team users can now add internal databases as references, allowing the platform to respond with improved contextual awareness. By connecting internal knowledge bases, ChatGPT Team could become more invaluable to users who ask the platform strategy questions or for analysis. This allows users to perform semantic searches of their data, link directly to internal sources in responses, and ensure ChatGPT understands internal company lingo. Recommended read:
References :
@singularityhub.com
//
OpenAI models, including the recently released GPT-4o, are facing scrutiny due to their vulnerability to "jailbreaks." Researchers have demonstrated that targeted attacks can bypass the safety measures implemented in these models, raising concerns about their potential misuse. These jailbreaks involve manipulating the models through techniques like "fine-tuning," where models are retrained to produce responses with malicious intent, effectively creating an "evil twin" capable of harmful tasks. This highlights the ongoing need for further development and robust safety measures within AI systems.
The discovery of these vulnerabilities poses significant risks for applications relying on the safe behavior of OpenAI's models. The concern is that, as AI capabilities advance, the potential for harm may outpace the ability to prevent it. This risk is particularly urgent as open-weight models, once released, cannot be recalled, underscoring the need to collectively define an acceptable risk threshold and take action before that threshold is crossed. A bad actor could disable safeguards and create the “evil twin” of a model: equally capable, but with no ethical or legal bounds. Recommended read:
References :
Ryan Daws@Developer Tech News
//
References:
Developer Tech News
, Windows Report
OpenAI has unveiled a new suite of APIs and tools aimed at streamlining the development of AI agents. This initiative addresses the challenges faced by software developers in building production-ready applications, with the goal of transforming how they create systems capable of autonomously handling complex, multi-step tasks. The new offerings are designed to empower developers and enterprises to build, deploy, and scale reliable, high-performing AI agents more easily.
The suite includes the Responses API, which combines the simplicity of the Chat Completions API with the tool-use capabilities of the Assistants API. This API supports built-in tools like web search, file search, and computer use, facilitating the creation of agents that can interact effectively with real-world systems. Additionally, OpenAI has introduced the Agents SDK, an orchestration framework that simplifies the design and scaling of agents, featuring built-in observability tools for performance logging, visualization, and analysis. These tools are expected to enhance productivity and innovation across various industries by enabling the creation of more efficient and capable AI-driven applications. Recommended read:
References :
Ellie Ramirez-Camara@Data Phoenix
//
OpenAI has recently unveiled a suite of new tools aimed at simplifying the development of AI agents. This release includes the Responses API, designed as a flexible foundation for building agents, along with built-in capabilities for web search, file search, and computer use. An open-source Agents SDK is also part of the package, intended to help developers orchestrate both single-agent and multi-agent workflows, providing the essential building blocks for creating reliable AI agents.
OpenAI has also introduced o1 Pro, its latest AI reasoning model, but it comes with a steep price tag. At $150 per million input tokens and $600 per million output tokens, o1 Pro is significantly more expensive than previous models, costing ten times the price of the standard o1 model and twice as much as GPT-4.5. OpenAI claims that o1 Pro "thinks harder" and provides "consistently better" responses, especially for complex tasks, but early benchmarks suggest the improvements may be incremental. Recommended read:
References :
Matthias Bastian@THE DECODER
//
References:
Simon Willison's Weblog
OpenAI has released another update to its GPT-4o model in ChatGPT, delivering enhanced instruction following capabilities, particularly for prompts with multiple requests. This improvement is a significant upgrade which has also allowed it to acheive second place on the LM Arena leaderboard, only being beaten by Gemini 2.5. The update also boasts improved capabilities in handling complex technical and coding problems, alongside enhanced intuition and creativity, with the added benefit of fewer emojis in its responses.
This update, referred to as chatgpt-4o-latest, is also now available in their API, and also gives access to the model used for ChatGPT. This version is priced higher at $5/million input and $15/million output compared to the regular GPT-4o, which is priced at $2.50/$10. OpenAI plans to bring these improvements to a dated model in the API in the coming weeks, and although they released the update on Twitter, users have complained that a more suitable place for this announcement would be the OpenAI Platform Changelog. Recommended read:
References :
|
BenchmarksBlogsResearch Tools |