OpenAI’s New Models Aren’t Really Open: What to Know About Open-Weights AI

Despite the name of the company, Openai has not abandoned an open version of its AI models since GPT-2 in 2019.

If the open weights are a new piece of jargon, don’t worry. In the simplest possible terms, open weights are a category of AI models that feed products such as chatbots, image and video generators. But they are philosophically different from the technology underlying the AI tools that you could use now. Chatgpt, Gemini and Copilot are all fueled by closed models, which means that we do not have real information on the operation of these black box machines. Open weight models give us a look at the mechanical assistant behind the curtain, so to speak.

You don’t need to be a developer or an automatic learning expert to understand how these open models work or even to manage them yourself. Here is everything you need to know about open weights and open source AI models.

What is an open weight AI model?

All AI models have weights, which are characteristics or elements. The models are formed to give certain connections more weight or value.

An open weight model does exactly what its name implies – the weights are accessible to the public, as defined by the Federal Trade Commission. Developers can see these weights and how they are used in creating AI models.

“No doubt, the most precious thing in large [language] The models are actually the weight. You can do a lot if you have the weight, which is somewhat different from traditional software, “Omar Khattab, assistant professor of IT and researcher in his computer laboratory and artificial intelligence (CSAIL), told CNET.

For example, a chatbot is designed to be really good to predict the next logical word in a sentence. It is formed to train words in its outings which frequently appear next to each other in its training data, probably in a logical order. Words that appear to each other more frequently can receive more weight than words that do not often appear next to each other.

These weights are only numbers, but open weight models are also delivered with a card.

“By open weight [models]You get the weights, which are these figures, and you get how to map these weights in the structure of the neural network, so the layers of the neural network, so that they can actually execute it, “said Khattab. The architecture of the model shows how a company structures its models, which is” incredibly precious “.

Open weight models are mainly intended for developers, who can integrate the model into existing projects, such as helping building AI agents. For the “committed amateur”, as Khattab says, you can use the specifications to execute the model locally on your laptop, which could help to mitigate the confidentiality problems which come, for example, with the use of AI via the mobile application of a company. Researchers will also have an overview more and more how AI work internally.

How do the new open models from Openai accumulate?

The new open models are available in two sizes, 120 billion parameters (128 experts and a 128K context window) and 20 billion parameters (32 experts but the same 128K context window). Experts refer to the number of sub-neural networks of a model, and context windows describe the amount of information that a model can process and include in its responses. More important numbers for both indicate that a model is capable of more sophisticated responses and has more fireplace.

In terms of performance, Openai reports that the 120B model “has reached almost a” model of reasoning, O4-Mini on basic reasoning references while operating on a single GPU of 80 GigaCtets. The 20B model of open weight occurred similarly to O3 -Mini and operated on a device of 16 gigabytes – which means that this smaller open weight model could be fairly well executed on laptops and certain smartphones. (As all the models of AI run locally, your speed will depend on the firepower of your device.)

The models will be available under the Apache 2.0 license, an open source license type. You can consult the most in-depth specifications of the model and paper card on safety training, get advice from the Délénai developer guidelines and see the weights for yourself on Huggingface and Github.

Are open weights the same as open-source AI?

Open weight models are linked to open source AI, but they are not exactly the same. Open Source as a concept refers to software that has no owners, whose source code is accessible to the public and can be used by most open source licenses. The open source initiative, a non -profit organization that recommends open source software, defines open source AI as “a system made available in terms that grant users the freedom to use, study, modify and share [it]. “”

An AI of open weight is not the same as an open source model. One way of thinking about the difference between the two is like cooking, Suba Vasudevan, chief of the farm in Mozilla.org and main vice-president of Mozilla Corporation, told Cnet.

“An open weight model is someone who gives you this cake baked in the oven and says:” Oh, it is made of flour, sugar and eggs. “These are the weights.

For open models, the types of things that are not disclosed are the data on which the model has been formed and the code used to train it. Training data is a point of discord between AI companies and humans creating content; AI companies are voracious for content generated by high quality humans to refine and improve their models. Some companies collect this data through license agreements, but some publishers and creators have filed prosecution alleging that IA companies illegally acquire their content protected by copyright. (Disclosure: Ziff Davis, CNET’s parent company in April, filed a complaint against Openai, alleging that it has violated Ziff Davis Copyrights in the training and exploitation of its AI systems.)

Training data, whatever its origin, is one of the most precious things of an AI company. But it will probably not be included in any version of open weight model.

“I think that because of a pure scale [of data]Because of the responsibility, because a large part of it is under license and you are not supposed to share it, I think it is probably fair to suppose that no business, at least for profit, will publish this soon, perhaps never, “said Khattab.

The really open source AI comes with information more accessible to the public, said Vasudevan. Open weight models may be more difficult to recycle or inspect the bias without additional information. “You are still a little in the dark about how it was built or what data have shaped it,” said Vasudevan.

Look at this: How you talk to Chatgpt counts. Here is why

04:12

Why should I care about Open-Source AI?

These differences between open source, open weights and closed models will probably not affect your average experience using an AI chatbot. But these are important qualities that developers must consider when selecting and using certain AI models, and they are more broadly important for us to understand the technology that infiltrates our online life.

There is no guarantee that the internal operation of its open models will reflect what is inside its closed models, Khattab said. But for a company whose product is synonymous with the AI generation, any light hoof on its operation is sure to have an impact on the people who use it and those who design it behind the scenes. While people start to get into the weeds with new models, we will learn more and see what effect it has on industry in the future.

Overall, the underlying philosophy of open source is that technology improves when more people can access it, such as academics who can study IT and security experts who can expose weaknesses. “You have to let people build, evolve and innovate and be able to prick holes and improve it,” said Vasudevan.

To find out more, see our guide for the chatgpt beginner and how AI changes our research experience.