OpenAI Aims for Best-in-Class 'Open' AI Model

OpenAI's Forthcoming Open Language Model

OpenAI announced in late March its plans to launch its first fully “open” language model since GPT‑2 at some point this year. Initial specifics regarding this model are now emerging from discussions with the AI developer community.

Development and Timeline

Aidan Clark, OpenAI’s Vice President of Research, is currently spearheading the development of this open model, which remains in its initial phases, according to sources who have spoken with TechCrunch. The anticipated release timeframe is early summer.

The company’s objective is to ensure the model achieves superior performance compared to other publicly available reasoning models when assessed against established benchmarks.

Licensing and Accessibility

OpenAI is considering a highly permissive licensing structure for the model, minimizing restrictions on its usage and commercial applications. Existing open models, such as Llama and Google’s Gemma, have faced criticism for imposing stringent requirements. OpenAI appears to be actively working to circumvent these concerns.

Competitive Landscape

The ChatGPT developer is experiencing increased competitive pressure from companies like DeepSeek, a Chinese AI laboratory, which have embraced an open-source approach to model deployment. Unlike OpenAI’s traditional strategy, these competitors freely distribute their models to the AI community for both experimentation and potential commercial use.

Success of Open-Source Models

This strategy has demonstrably proven successful for several organizations. Meta, with its substantial investment in the Llama series of open AI models, reported over 1 billion downloads in early March. Similarly, DeepSeek has rapidly gained a significant global user base and garnered investor interest.

Model Capabilities and Hardware

Sources indicate that OpenAI intends for its open model – designed for “text in, text out” functionality – to operate effectively on advanced consumer-grade hardware. Developers may also have the option to enable or disable the model’s “reasoning” capabilities, mirroring features found in recently released reasoning models from Anthropic and other developers.

Should the initial release be well-received, OpenAI may subsequently introduce additional models, potentially including smaller, more streamlined versions.

Shift in OpenAI's Open-Source Philosophy

Sam Altman, OpenAI’s CEO, has previously expressed his belief that OpenAI may have erred in its past approach to open-sourcing its technologies.

During a Reddit Q&A session in January, Altman stated, “[I personally think we need to] develop a different open source strategy.” He acknowledged that not all within OpenAI share this viewpoint and that it isn’t currently the company’s primary focus. He further noted that while OpenAI will continue to produce advanced models, its competitive advantage may diminish in the coming years.

Safety and Evaluation

Altman has also emphasized that the forthcoming open model will undergo rigorous safety testing and evaluation. Sources confirm that OpenAI plans to publish a model card – a comprehensive technical report detailing the results of both internal and external benchmarking and safety assessments.

“[B]efore release, we will evaluate this model according [to] our preparedness framework, like we would for any other model,” Altman posted on X last month. “[A]nd we will do extra work given that we know this model will be modified post-release.”

Past Concerns Regarding Safety

OpenAI has faced scrutiny from some AI ethicists regarding the perceived haste in safety testing for its recent models and the lack of publicly released model cards for others. Altman has also been accused of providing misleading information to OpenAI executives concerning model safety reviews prior to his temporary removal from his position in November 2023.

Seeking Comment

TechCrunch has contacted OpenAI for a statement and will update this article upon receiving a response.