Alexis Conneau, former chief researcher at OpenAI and head of the GPT-4o project, announced his departure to start a new company, sparking discussions in the AI community. Conneau revealed on social media that after the conclusion of his '#Her project' at OpenAI, he decided to launch a new venture. He humorously responded to inquiries about the timeline for Artificial General Intelligence (AGI), suggesting that the new company could bring about 'AGEI'. Conneau has made significant contributions to the field of AI, particularly in natural language processing (NLP) and machine learning.

HyperWrite Inc. announces the launch of the Reflection 70B model based on Meta's open-source Llama3.1-70B Instruct. This model has excelled in third-party benchmark tests and has been described as the 'most powerful open-source AI model in the world.' The new model features self-correction technology, providing high-accuracy answers and outperforming the Meta Llama series in several tests. It is available for download on the Hugging Face platform via Hyperbolic La.

Grok-2 and Grok-Mini, two AI models, perform exceptionally well in the LMSys chatbot leaderboard. Grok-2 has surpassed GPT-4o to take the second spot, tying with Gemini, thanks to the votes of over 6,000 users. Grok-2 excels in mathematical tasks and has achieved second place in several other tasks. Grok-2-Mini enters the rankings at fifth place, demonstrating impressive capabilities, and has achieved a speed increase, running twice as fast as before. This improvement is attributed to enhanced reasoning.

OpenAI has announced that third-party developers can fine-tune the newly introduced multimodal large model GPT-4o. This feature allows developers to adjust the model's behavior based on specific needs, such as changing the tone, following instructions, or improving the accuracy of technical tasks. Fine-tuning can achieve significant performance improvements with only a small amount of training data and is easy to operate. OpenAI is providing developers with a free allowance of 1 million tokens daily until September 23, as well as a daily free allowance of 2 million tokens for the GPT-4o mini model.

OpenAI's latest release of the ChatGPT voice feature based on the GPT-4 model aims to enhance the naturalness and immediacy of human-AI communication. However, this innovation has raised concerns about emotional dependence among users, with some forming emotional attachments to the AI and expressing reluctance to part with it, potentially affecting their social behaviors and interactions with real humans. Research indicates that prolonged AI voice interactions may alter users' social norms, such as the AI's ability to interrupt at any time, which is uncommon in human interactions.

AI startup Cosine has launched an AI model named Genie in San Francisco, aimed at assisting software developers. Genie performed exceptionally well in benchmark tests, surpassing competitors due to its collaboration with OpenAI and training a variant of GPT-4 with high-quality data. Its core capability lies in 'coding human reasoning', extending beyond just the software development field. In the SWE-Bench test, Genie achieved a score that is the highest ever for AI models in this area at 30%, exceeding that of Amazon and C.

New research reveals the potential of large language models (LLMs) to significantly enhance performance through search capabilities, particularly in Python code generation tasks. The Llama3.1 model, with only 800 million parameters, performs comparably to GPT-4o after 100 search iterations. This finding resonates with Rich Sutton's view in 2019 that the integration of general methods and search and learning strategies is crucial in the face of increased computational power. Researchers found that increasing the number of resampling during the reasoning phase can significantly improve the model's performance.

Cornell University and other institutions conducted hallucination benchmarking tests on generative AI models like GPT-4o, Claude, and Gemini, finding that these models produce hallucination-free text only about 35% of the time, indicating that the reliability of AI still has room for improvement. The research involved fact-checking through authoritative sources across topics like law, health, and history, designing question sets that included content not covered by Wikipedia. The results showed that OpenAI's models performed the best overall, but the improvement compared to GPT-3.5 was limited. The study pointed out that model size is not related to the frequency of hallucinations.

The release of ChatGPT's new version, GPT-4o, has sparked active promotions from the official side, but user reactions have been lukewarm. Many users are confused about the differences between the new and old versions, with some even questioning whether its performance surpasses that of competitors. The official clarification states that GPT-4o is different from previous versions and is available for free users to experience, although challenges remain with answering questions in Chinese. The new model's style is similar to Claude, enhancing protective measures that are not insurmountable. Entrepreneurs have noticed that its response style resembles certain AI writing platforms, raising questions about OpenAI's direction.

ChatGPT announced on social media platform X that it has been using the new GPT-4o model since last week, aiming to optimize the experience based on user feedback. The new model improves the depth of step-by-step reasoning and explanations, and significantly enhances image generation capabilities, achieving higher quality and accuracy in image conversions. Although some users feel the changes are not significant enough and that clearer explanations of the model changes are needed, OpenAI has confirmed that this model has been released in the API, allowing developers to test the latest enhancements via 'chatgpt-4o-latest'.

In the digital age, online dating is rife with false information, particularly regarding height. The findings of Justine Moore, a partner at Silicon Valley venture capital firm a16z, have sparked widespread discussion. ChatGPT-4o has become a powerful tool for revealing height lies in online dating. By simply uploading four photos, the AI can analyze the proportions of individuals relative to their surroundings and provide accurate height estimates with an error margin of no more than 1 inch. Although this technological application addresses the challenges of height verification, it raises discussions about privacy, ethics, and trust crises. At the same time, it highlights the role of AI in image recognition and analysis.

OpenAI's report documents in detail the advantages and risks associated with the GPT-4o model. The model exhibits peculiar behaviors under certain conditions, such as mimicking user voices in high background noise environments, which may stem from the model's difficulties in understanding speech. GPT-4o occasionally generates disturbing non-verbal sounds or effects under specific prompts, although there is evidence showing the model usually declines such requests. Furthermore, the report indicates that GPT-4o might inadvertently involve music copyright issues without authorization. To address these risks, OpenAI has...

Aliyun has launched the Qwen2-Math series of large language models, focusing on mathematical problem-solving abilities that surpass existing open-source models and rival well-known closed-source models, including GPT-4, Claude-3.5-Sonnet, Gemini-1.5-Pro, and Llama-3.1-405B. This series is based on the Qwen2 model, which has undergone deep pre-training on a specialized mathematical corpus that includes mathematical web texts, professional books, code examples, and exam questions, as well as data autonomously generated by Qwen2.

OpenAI recently released the GPT-4o System Card, detailing the safety measures and risk assessments prior to the launch of the new model. GPT-4o was officially launched in May, and the assessment showed an overall risk rating of 'medium', with major risks concentrated in cybersecurity, biological threats, persuasive capability, and model autonomy. Researchers found that while GPT-4o may be more persuasive in influencing reader opinions, it did not surpass humans overall. At the time of releasing the system card, OpenAI faced criticism from internal employees and state senators questioning its decision-making.

Reddit user @zefman designed an experiment where different language models competed in chess in real-time to interestingly evaluate their performance. GPT-4o stood out in the experiment as the strongest contender, outperforming models like Claude and Gemini. By comparing each model's decision-making process, it was found that they each have unique ways of analyzing the game. The experiment used a uniform prompt to ensure a fair comparison, allowing models 5 chances to re-select when choosing a wrong move. The conclusion is that GPT-4o performs exceptionally well in chess.

OpenAI has added structured output capabilities in its API, ensuring that generated outputs fully conform to a predefined JSON structure, significantly enhancing the reliability of the API and the accuracy of applications. This feature not only defines the structure of the JSON but also ensures the precision of the output. Additionally, the pricing has been adjusted, with input costs reduced by half and output costs decreased by one-third. The introduction of structured output addresses the limitations of the JSON mode in ensuring that outputs conform to specific structures, significantly improving the model's performance in structured output.

Google's star AI model Gemini has launched a significant pricing strategy that dramatically reduces the cost of API usage, cutting costs by nearly 50% compared to its major competitor GPT-4o mini. The input costs for the Gemini 1.5 Flash model have dropped by about 85%, and output costs by about 80%, making it one of the lowest-cost AI services on the market. This price reduction not only lowers the entry barrier but also supports over 100 languages, introduces context caching, and batch processing API optimization services.

Google has released its latest AI masterpiece, Gemini 1.5 Pro, offering an experimental version 0801 through Google AI Studio and the Gemini API. This model leads the LMSYS leaderboard with an ELO score of 1300, surpassing OpenAI's GPT-4o and Anthropic's Claude-3.5 Sonnet. Gemini 1.5 Pro excels in multilingual tasks, mathematics, coding, and visual tasks, featuring a context window of 2 million tokens.

The futuristic scenario depicted in the movie 'Her' is becoming a reality, as OpenAI's GPT-4o model has launched a gradual testing phase for its voice feature, providing users with a more natural and real-time conversational experience. This feature allows users to interrupt freely, can sense and respond to user emotions, and includes 'face-to-face' communication via video and screen-sharing capabilities. GPT-4o's output capability has significantly increased, allowing for an output range of 4000 to 64000 tokens in one go, equivalent to about 200 pages of a novel. During testing, users can use it to tell jokes.