OpenAI recently released gpt-oss, a powerful open-weight large language model (LLM) family, under the Apache 2.0 license. This marks the company’s first open weights model launch since GPT-2 in 2019. Following the release, Jack Morris, a PhD student at Cornell Tech, unveiled a modified version, gpt-oss-20b-base, which removes the “reasoning” behaviors of the original model, reverting it to a base version for quicker, uncensored responses.
Morris’s adaptation is available on Hugging Face under an MIT License, permitting both research and commercial use. Understanding this development requires some context regarding the differences between OpenAI’s models and what researchers refer to as “base models.” OpenAI’s models, such as gpt-oss, are “post-trained,” meaning they undergo an additional phase to refine behaviors based on curated examples. In contrast, base models are pretrained without specific alignment, allowing them to produce varied and less constrained outputs.
Morris aimed to restore the model to a form closer to its original pretrained state. He described this process as reversing OpenAI’s alignment training, resulting in a model that generates text without engaging in structured reasoning. This new model produced more varied outputs, including content that the aligned models would typically avoid, raising questions about safety and ethical considerations.
Despite the model’s ability to reproduce copyrighted material and potentially risky content, some elements of alignment remain. If prompted in a structured format, it may still behave in a more reserved manner. In terms of technical execution, significant computational resources were used, and while Morris claims to have recovered the distribution of the base model, he clarifies that the internal weights of the model were not extracted.
The initial response to OpenAI’s gpt-oss has been mixed, with some praising its permissive license and efficiency, while others express concerns about its underlying training data and safety features. Morris’s approach provides a concrete example of how open-weight models can be adapted quickly after release, entering ongoing discussions within the AI community regarding the balance between innovation and safety.
Source: https://venturebeat.com/ai/this-researcher-turned-openais-open-weights-model-gpt-oss-20b-into-a-non-reasoning-base-model-with-less-alignment-more-freedom/