Latest news with #GPT2
CNET
4 days ago
- Business
- CNET
OpenAI's New Models Aren't Really Open: What to Know About Open-Weights AI
Despite the company's name, OpenAI hasn't dropped an open version of its AI models since GPT-2 in 2019. That changed on Tuesday, as CEO Sam Altman shared two new open-weights, reasoning AI models, named gpt-oss-120b (120 billion parameters) and gpt-oss-20b (20 billion parameters). If open-weights is a new piece of AI jargon to you, don't worry. In the simplest possible terms, open-weights is a category of AI models that power products like chatbots, image and video generators. But they are philosophically different from the technology underpinning the AI tools you might use now. ChatGPT, Gemini and Copilot are all powered by closed models, which means we have no real insights into how those black box machines work. Open-weights models give us a peek at the mechanical wizard behind the curtain, so to speak. You don't need to be a developer or machine learning expert to understand how these open models work or even to run them yourself. Here's everything you need to know about open-weights and open-source AI models. What is an open-weights AI model? All AI models have weights, which are characteristics or elements. Models are trained to give certain connections more weight, or value. An open-weights model does exactly what its name implies -- the weights are publicly available, as defined by the Federal Trade Commission. Developers can see these weights and how they're used in the creation of AI models. "Arguably, the most valuable thing in large [language] models is actually the weights. You can do a lot if you have the weights, which is somewhat different from traditional software," Omar Khattab, assistant professor of computer science at MIT and researcher in its computer science and artificial intelligence lab (CSAIL), told CNET. For example, a chatbot is built to be really good at predicting the next logical word in a sentence. It's trained to string together words in its outputs that frequently show up next to each other in its training data, presumably in a logical order. Words that show up next to each other more frequently can be given more weight than words that don't often appear next to each other. These weights are just numbers but open-weights models also come with a map. "In open-weights [models], you get the weights, which are these numbers, and you get how to map those weights into the neural network structure, so the layers of the neural network, in order to actually be able to run it," said Khattab. The architecture of the model shows how a company structures its models, which is "incredibly valuable." Open-weights models are primarily aimed at developers, who can integrate the model into existing projects, like helping to build AI agents. For the "committed hobbyist," as Khattab put it, you can use the specs to run the model locally on your laptop, which could help alleviate privacy concerns that come, for example, with using AI through a company's mobile app. Researchers will also have a clearer look at how the AI works internally. Elena Uve/iStock/Getty Images Plus How do OpenAI's new open-weights models stack up? The new open-weights models come in two sizes, 120 billion parameters (128 experts and 128k context window) and 20 billion parameters (32 experts but the same 128k context window). Experts refer to the number of sub-neural networks a model has, and context windows describe how much information a model can process and include in its responses. Bigger numbers for both indicate a model is capable of more sophisticated answers and has more firepower. In terms of performance, OpenAI is reporting that the 120B model "achieved near-parity" with its latest reasoning model, o4-mini on core reasoning benchmarks while running on a single 80 gigabyte GPU. The 20B open-weights model performed similarly to o3-mini and ran on a 16 gigabyte device -- meaning, this smaller open-weights model could be run fairly well on laptops and some smartphones. (Like all AI models run locally, your speed will depend on your device's firepower.) The models will be available under the Apache 2.0 license, a type of open-source-friendly license. You can check out the more in-depth specs in the model card and paper on safety training, get tips from OpenAI's developer guidelines and see the weights for yourself now on HuggingFace and Github. CNET's Imad Khan installed gpt-oss-20B on a PC specifically built to run local AI models. Screenshot by Imad Khan/CNET Is open-weights the same as open-source AI? Open-weights models are related to open-source AI, but they aren't exactly the same. Open source as a concept refers to software that has no proprietary owners, whose source code is publicly available and can be used by most anybody under open-source licenses. The Open Source Initiative, a nonprofit that advocates for open-source software, defines open-source AI as "a system made available under terms that grant users the freedom to use, study, modify and share [it]." An open-weights AI model isn't the same as an open-source model. One way to think about the difference between the two is like with baking, Suba Vasudevan, chief operating officer at and senior vice president at Mozilla corporation, told CNET. "An open-weights model is someone giving you this baked cake and saying, 'Oh, it's made of flour, sugar and eggs.' Those are the weights. But if someone gave you the entire recipe, and all the instructions and the exact details of how much of each ingredient went into it, that's open source," said Vasudevan. For open-weights models, the types of things not disclosed are the data the model was trained on and the code used to train it. Training data is a point of contention between AI companies and the humans creating content; AI companies are ravenous for high-quality, human-generated content to refine and improve their models. Some companies collect this data through licensing agreements, but some publishers and creators have filed lawsuits alleging that AI companies are illegally acquiring their copyrighted content. (Disclosure: Ziff Davis, CNET's parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) Training data, regardless of its origin, is one of the most valuable things an AI company has. But it likely won't be included in any open-weights model release. "I think because of sheer scale [of data], because of liability, because a lot of it is licensed and you are not supposed to share it, I think it's probably fair to assume that no companies, for-profit at least, are going to be releasing that anytime soon, maybe ever," said Khattab. Truly open-source AI comes with more publicly available information, Vasudevan said. Open-weights models can be harder to retrain or inspect for bias without the additional information. "You're still a fair bit in the dark about how it was built or what data shaped it," Vasudevan said. Now Playing: How You Talk to ChatGPT Matters. Here's Why 04:12 Why should I care about open-source AI? These differences between open-source, open-weights and closed models likely won't affect your average experience using an AI chatbot. But they're important qualities for developers to consider when selecting and using certain AI models, and they're important more broadly for us to understand the technology that is infiltrating our online lives. There's no guarantee that the inner workings of its open-weights models will reflect what's inside its closed models, Khattab said. But for a company whose product is synonymous with gen AI, any light shed on how it works is sure to have an impact on the people who use it and those who design it behind the scenes. As people start to get into the weeds with the new models, we'll learn more and see what effect it has on the industry going forward. Overall, the underlying philosophy of open source is that technology gets better when more people can access it, like academics who can study it and security experts who can expose weaknesses. "You have to let people build, scale and innovate and be able to poke holes and make it better," said Vasudevan. For more, check out our beginner's guide to ChatGPT and how AI is changing our search experience.
Gizmodo
4 days ago
- Business
- Gizmodo
OpenAI Finally Lives Up to Its Name, Drops Two New Open Source AI Models
For the first time in five years, OpenAI has released two new free and open-source AI models that are lightweight and designed to be easily integrated into other software programs. In a blog post on Tuesday, the company characterized gpt-oss-120b and gpt-oss-20b as flexible but powerful AI algorithms that can perform a variety of tasks and be used in numerous settings. The company also included a feedback portal and a more extensive blog that further explains the models and how they work. OpenAI's CEO, Sam Altman, said via X on Tuesday that he hoped that the AI would assist with 'new kinds of research and the creation of new kinds of products.' Altman also seemed to champion the open source method: 'We believe in individual empowerment,' he said. 'Although we believe most people will want to use a convenient service like ChatGPT, people should be able to directly control and modify their own AI when they need to, and the privacy benefits are obvious.' Unlike most of ChatGPT's products, open-source models disclose the training parameters that were used to build their system. This level of transparency affords onlookers the benefit of knowing how the system functions and why it might behave in the ways that it does. The last time OpenAI released an open-source model was during Trump's first presidency (truly, that feels like decades ago), with the release of GPT-2. That was when OpenAI was still a little-known startup, and it would be several years before the launch of ChatGPT in 2022. Back in those days, the company was still routinely being referred to by monikers like 'Elon Musk's AI project,' despite the fact that Musk had parted ways with the company. Perhaps most importantly, you won't have to pay a dime to use these models. And as long as your computer's specs are up to snuff, you can run them locally instead of relying on OpenAI's servers. Any preppers out there looking for an AI doomsday machine might want to look into it. From the looks of things, there's a lot of promising stuff in the company's new releases. gpt-oss is built to integrate into agentic workflows, OpenAI says, which means that new types of automated work—conducted by so-called 'agents'—can be powered by the new algorithms. The new models also fall under the Apache 2.0 license, which allows users to create new software with the algorithms without worrying about getting sued. 'Build freely without worrying about copyleft restrictions or patent risk—whether you're experimenting, customizing, or deploying commercially,' OpenAI writes. The open-source ecosystem largely subsists on a plethora of such licensing agreements, allowing companies to build off free models. OpenAI also paid some lip service to AI safety in its announcement. The company claims that, in addition to 'running the models through comprehensive safety training and evaluations,' it also 'introduced an additional layer of evaluation by testing an adversarially fine-tuned version of gpt-oss-120b' using its Preparedness Framework, which is designed to assess and track risky behavior in large language models. In recent years, much criticism has been aimed at OpenAI over its decision to continue the 'walled garden' approach to software development. The company's LLM releases have stayed proprietary and, thus, shut off from public inspection. Now, OpenAI is obviously trying to prove the haters wrong and double down on its commitment to being a truly 'open' organization. It's anybody's guess as to whether the company will be able to maintain such a commitment to the FOSS ethos, given that it also happens to be an organization worth hundreds of billions of dollars on paper. The organized money behind OpenAI may continue to see a benefit in owning systems that are exclusive, closed, and, most importantly, only controlled by a select group of insiders. It's noteworthy that GPT-5, the company's most powerful and highly anticipated new model, will almost certainly be released in the same fashion as other recent GPT releases: closed.
WIRED
4 days ago
- Business
- WIRED
OpenAI Just Released Its First Open-Weight Models Since GPT-2
Aug 5, 2025 1:00 PM The models, gpt-oss-120b and gpt-oss-20b, represent a major shift for the AI company. Photograph:OpenAI just dropped its first open-weight models in over five years. The two language models, gpt-oss-120b and gpt-oss-20b, can run locally on consumer devices and be fine-tuned for specific purposes. For OpenAI, they represent a shift away from its recent strategy of focusing on proprietary releases, as the company moves towards a wider, and more open, group of AI models that are available for users. "We're excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible," said OpenAI CEO Sam Altman in an emailed statement. Both gpt-oss-120b and gpt-oss-20b are officially available to download for free on Hugging Face, a popular hosting platform for AI tools. The last open-weight model released by OpenAI was GPT-2, back in 2019. What sets apart an open-weight model is the fact that its 'weights' are publicly available, meaning that anyone can peek at the internal parameters to get an idea of how it processes information. Rather than undercutting OpenAI's proprietary models with a free option, cofounder Greg Brockman sees this release as 'complementary' to the company's paid services, like the application programming interface currently used by many developers. 'Open-weight models have a very different set of strengths,' said Brockman in a briefing with reporters. Unlike ChatGPT, you can run a gpt-oss model without a connection to the internet and behind a firewall. Both gpt-oss models use chain-of-thought reasoning approaches, which OpenAI first deployed in its o1 model last fall. Rather than just giving an output, this approach has generative AI tools go through multiple steps to answer a prompt. These new text-only models are not multimodal, but they can browse the web, call cloud-based models to help with tasks, execute code, and navigate software as an AI agent. The smaller of the two models, gpt-oss-20b, is compact enough to run locally on a consumer device with more than 16 GB of memory. The two new models from OpenAI are available under the Apache 2.0 license, a popular choice for open-weight models. With Apache 2.0, models can be used for commercial purposes, redistributed, and included as part of other licensed software. Open-weight model releases from Alibaba's Qwen as well as Mistral also operate under Apache 2.0. Publicly announced in March, the release of these open models was initially delayed for further safety testing. Releasing an open-weight model is potentially more dangerous than a closed-off version since it removes barriers around who can use the tool, and anyone can try to fine-tune a version of gpt-oss for unintended purposes. In addition to the evaluations OpenAI typically runs on its proprietary models, the startup customized the open-weight option to see how it could potentially be misused by a 'bad actor' who downloads the tool. 'We actually fine-tuned the model internally on some of these risk areas,' said Eric Wallace, a safety researcher at OpenAI, 'and measured how high we could push them.' In OpenAI's tests, the open-weight model did not reach a high level of risk, as measured by its preparedness framework. How do these models perform compared to OpenAI's other releases? 'The benchmark scores for both of these models are pretty strong,' said Chris Koch, an OpenAI researcher, in the briefing. Speaking about gpt-oss-120b, the researcher compared its performance as closely similar to OpenAI's o3 and o4-mini models, which are proprietary, and even out-performing them in certain evaluations. The model card for gpt-oss goes into detail about how exactly it stacks up to the company's other offerings. In a pre-launch press briefing, staff members of OpenAI also focused on the latency offered by gpt-oss and the cheaper cost to run these models. At the beginning of this year, the Chinese startup DeepSeek stunned Silicon Valley with the release of its cheap-to-run model that was open-weight. While the release blog about gpt-oss does not mention DeepSeek or any other Chinese AI company directly, Altman is clear that he wants innovation around open-weight models to happen in the United States. "Going back to when we started in 2015, OpenAI's mission is to ensure AGI that benefits all of humanity,' said Altman in a statement. 'To that end, we are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit." In the US, the open-weight leader has been Meta. The tech giant released the first of its Llama series of models back in 2023, with Meta's most recent release, Llama 4, arriving a few months ago. With that in mind, Meta is currently hyper-focused on building AI that can surpass human cognition, often called superintelligence by AI insiders. The company recently launched a new, internal lab focused on this lead by Alexandr Wang, the former CEO of Scale. Mark Zuckerberg has signaled that the company may move away from open-source for future models, citing potential safety concerns. The gpt-oss release also comes as the AI talent war between companies, like OpenAI and Meta, continues to ramp up. In 2025, AI researchers who have in-demand talents are being presented with astronomical offers to switch companies. The latest releases from OpenAI could be stiff competition for Meta, depending on how the gpt-oss models are received by developers.
Yahoo
18-02-2025
- Business
- Yahoo
'Hopeless' to potentially handy: law firm puts AI to the test
Artificial intelligence (AI) tools have got significantly better at answering legal questions but still can not replicate the competence of even a junior lawyer, new research suggests. The major British law firm, Linklaters, put chatbots to the test by setting them 50 "relatively hard" questions about English law. It concluded OpenAI's GPT 2, released in 2019, was "hopeless" but its o1 model, which came out in December 2024, did considerably better. Linklaters said it showed the tools were "getting to the stage where they could be useful" for real world legal work - but only with expert human supervision. Law - like many other professions - is wrestling with what impact the rapid recent advances in AI will have, and whether it should be regarded as a threat or opportunity. The international law firm Hill Dickinson recently blocked general access to several AI tools after it found a "significant increase in usage" by its staff. There is also a fierce international debate about how risky AI is and how tightly regulated it needs to be. Last week, the US and UK refused to sign an international agreement on AI, with US Vice President JD Vance criticising European countries for prioritising safety over innovation. This was the second time Linklaters had run its LinksAI benchmark tests, with the original exercise taking place in October 2023. In the first run, OpenAI's GPT 2, 3 and 4 were tested alongside Google's Bard. The exam has now been expanded to include o1, from OpenAI, and Google's Gemini 2.0, which was also released at the end of 2024. It did not involve DeepSeek's R1 - the apparently low cost Chinese model which astonished the world last month - or any other non-US AI tool. The test involved posing the type of questions which would require advice from a "competent mid-level lawyer" with two years' experience. The newer models showed a "significant improvement" on their predecessors, Linklaters said, but still performed below the level of a qualified lawyer. Even the most advanced tools made mistakes, left out important information and invented citations - albeit less than earlier models. The tools are "starting to perform at a level where they could assist in legal research" Linklaters said, giving the examples of providing first drafts or checking answers. However, it said there were "dangers" in using them if lawyers "don't already have a good idea of the answer". It added that despite the "incredible" progress made in recent years there remained questions about whether that would be replicated in future, or if there were "inherent limitations" in what AI tools could do. In any case, it said, client relations would always be a key part of what lawyers did, so even future advances in AI tools would not necessarily bring to an end what it called the "fleshy bits in the delivery of legal services". What is AI and how does it work? Musk v Altman: What might really be behind failed bid for OpenAI
BBC News
18-02-2025
- Business
- BBC News
'Hopeless' to potentially handy: law firm tests chatbots
Artificial intelligence (AI) tools have got significantly better at answering legal questions but still can not replicate the competence of even a junior lawyer, new research major British law firm, Linklaters, put chatbots to the test by setting them 50 "relatively hard" questions about English concluded OpenAI's GPT 2, released in 2019, was "hopeless" but its o1 model, which came out in December 2024, did considerably said it showed the tools were "getting to the stage where they could be useful" for real world legal work - but only with expert human - like many other professions - is wrestling with what impact the rapid recent advances in AI will have, and whether it should be regarded as a threat or international law firm Hill Dickinson recently blocked general access to several AI tools after it found a "significant increase in usage" by its is also a fierce international debate about how risky AI is and how tightly regulated it needs to week, the US and UK refused to sign an international agreement on AI, with US Vice President JD Vance criticising European countries for prioritising safety over innovation. A future for 'fleshy bits' This was the second time Linklaters had run its LinksAI benchmark tests, with the original exercise taking place in October the first run, OpenAI's GPT 2, 3 and 4 were tested alongside Google's exam has now been expanded to include o1, from OpenAI, and Google's Gemini 2.0, which was also released at the end of did not involve DeepSeek's R1 - the apparently low cost Chinese model which astonished the world last month - or any other non-US AI test involved posing the type of questions which would require advice from a "competent mid-level lawyer" with two years' newer models showed a "significant improvement" on their predecessors, Linklaters said, but still performed below the level of a qualified the most advanced tools made mistakes, left out important information and invented citations - albeit less than earlier tools are "starting to perform at a level where they could assist in legal research" Linklaters said, giving the examples of providing first drafts or checking it said there were "dangers" in using them if lawyers "don't already have a good idea of the answer".It added that despite the "incredible" progress made in recent years there remained questions about whether that would be replicated in future, or if there were "inherent limitations" in what AI tools could any case, it said, client relations would always be a key part of what lawyers did, so even future advances in AI tools would not necessarily bring to an end what it called the "fleshy bits in the delivery of legal services".
