OpenAI officially announces a major open - source release. Altman reveals that 1 million new users have been added in the past hour. Sam Altman personally posts: It's the first open - source release since GPT - 2. Countdown to the developers' celebration. Key points: OpenAI announces that it will release an open - source large - scale model (open - we...

OpenAI Officially Announces a Major Open-Source Release. Altman Reveals 1 Million New Users in the Past Hour

Sam Altman personally posted: The first open-source release since GPT-2. Countdown to the developers' carnival. Key points: OpenAI announces that it will release an open-source large model (open-weight), the first new model to be open-sourced since GPT-2. It focuses on strong reasoning ability. Sam Altman bluntly said, "We want to make it very, very excellent." Developers can participate in the feedback in advance. Offline events in San Francisco, Europe, and Asia-Pacific are about to start. Security is the core: Special tests such as adversarial fine-tuning have been launched. No high-risk models will be released. Altman certified "the craziest moment of dissemination in history." And this time, OpenAI chooses to return to its original intention of open-source - this is the first time OpenAI has open-sourced a language model since GPT-2 in 2019. Why open-source now? Altman is "frank": He has long wanted to do it, but it has never been a priority. Now the time is ripe, and the team thinks "this matter has become important," and we must @DeepSeek here. Two characteristics of the new model: • Open-source (open-weight) • Enhanced reasoning ability. Although the specific parameters have not been disclosed, the official used two "verys" to emphasize the goal: "We want to make it a very, very excellent model." OpenAI has launched a global warm-up: The first offline event in San Francisco. European and Asia-Pacific sessions will follow. The pre-registration channel is open (link at the end of the article). The purpose is to collect feedback and let developers experience the prototype in advance. The official clearly stated: We want to see how large companies and governments independently deploy the model. Full security measures. From pre-training to release, OpenAI has implemented triple protection: Adopt the self-developed "Preparedness Framework" evaluation framework. Focus on studying the unique risks of open-source models (such as adversarial fine-tuning). Internal testing + double insurance with third-party experts. Altman said firmly: "We will never release a model that may cause catastrophic risks." Summary in plain language: This open-source release by OpenAI is both an olive branch to developers and a reassurance to the industry - "Open-source and security can coexist." As for how powerful the model is? Share your expectations in the comment section! Attachment: Official feedback link https://openai.com/open-model-feedback全文：山姆·奥特曼 (Sam Altman)OpenAI26 months ago, when ChatGPT was just released, it was one of the craziest phenomenon-level dissemination events I have witnessed. At that time, we attracted one million users in just five days. And in the past hour, we have added another one million users. To make a long story short: We are very happy to release a powerful new open-source weight language model with reasoning ability in the next few months. We hope to communicate with developers on how to make the most of this model: https://openai.com/open-model-feedback We are very much looking forward to making it a very, very excellent model! We are planning to release the first open-source weight language model since GPT-2. We have been thinking about this for a long time, but there were other more urgent matters before. Now we feel it is time (or think it is important) to do this. Before the release, we will evaluate the model according to our Security Preparedness Framework, just like any other model. And considering that we know the model may be modified after release, we will do additional work. We still have some decisions that have not been finalized yet, so we will hold developer events to collect feedback and let everyone try the early prototype later. The event will start in San Francisco (SF) in a few weeks, and then it will be held in Europe and Asia-Pacific (APAC). If you are interested in participating, please register through the link above. We are looking forward to seeing what developers will create and how large enterprises and government agencies that prefer to independently deploy and run the model will use it. Security is the core concern of our development of this open-source weight model, throughout the entire process from pre-training to release. Although open-source models bring unique challenges, we will follow our Security Preparedness Framework and will not release a model that we think may pose a catastrophic risk. We are particularly focused on studying adversarial fine-tuning and other risks unique to open-source models. Like the release of all models, before the public release, we are conducting extensive security tests, which include internal testing and testing in cooperation with trusted third-party experts. ...

PC version: https://www.cnbeta.com.tw/articles/soft/1489634.htm
Mobile version: https://m.cnbeta.com.tw/view/1489634.htm

via cnBeta.COM中文业界资讯站 - Telegram Channel