New open weight standard untethers the subconscious of AI
Framework enables the sharing of deep parameters which guide a model's behaviour and shape its responses.
The Open Source Alliance has released a new definition for an open-weight standard which allows the free sharing and redistribution of the "subconscious minds" of AI models.
Weights are the parameters AIs such as large language models (LLMs) learn during their childhood and then use to generate responses when they're mature and fully educated.
ChatGPT told us that weights could be compared to the subconscious - a dark, almost mechanical part of the psyche that controls behaviours without the knowledge or even consent of the conscious mind.
In Freud's model, the subconscious is a tangle of habits, emotions, instincts, memories and routines that power bodily functions such as breathing, as well as deep foundational beliefs and reactions that profoundly shape behaviour.
AI weights similarly contain “knowledge” the system has learned from training data. This knowledge governs how the AI makes decisions, but it is not directly interpretable or consciously accessible to the AI or humans.
Open Weight models "bridge the gap" between closed and open source models by allowing users to download and deploy advanced AI technologies independently without charge or needing to ask for permission.
Setting a new standard for open weights
The new definition grants access to the weights but not the data originally used to generate them.
"Open Weight models are indispensable tools for open innovation, allowing anyone to download and deploy cutting-edge AI models independently," said Sam Johnston, Convenor of the Open Source Alliance.
"We opted to base the draft open weight definition on the tried and tested open source definition because it addresses vendors who aspire to label their products as Open Source but are not yet willing or able to deliver the data - and for AI, the data is the source."
What is the definition of open source?
To be open source, software must adhere to four key requirements:
- Access: Source code must be made available allowing users to view, modify, and improve the software.
- Redistribution: Software can be freely shared and distributed without restriction, including selling or giving it away as part of a larger package.
- Modification: Users are allowed to modify the source code and distribute derivative works under the same licence as the original.
- Open to all: The licence must not discriminate against any person, group, or field of endeavour, ensuring equal access and usability for all.
The open weight definition protects two of the four core freedoms of open source software: the ability to use and share, but not necessarily to study or modify a model. This "lowers the barrier to entry, providing a level of flexibility and accessibility to vendors not yet able to satisfy the Open Source Definition covering all four", the Open Source Alliance wrote.
The draft Open Weight Definition is now open for public consultation at openweight.org. Additionally, collaboration is welcomed at the linked GitHub repository.
"The announcement is a welcome one"
Amanda Brock, CEO at OpenUK, responded at length to the new standard. She said that the definition is only in its early stages of development.
"To my mind... the approach of trying to define 'open source AI' is the wrong one," she said. "Rather we should follow this disaggregated approach to the challenge and look at the underlying 'technology' including the training data, and what it means to be open. Open source doesn't define law and it should not. It's about what enables anyone to use the technology's 'source' including data for any purpose.
"But this is subject to law. And if laws - whether privacy or otherwise, or contractual relationship - mean an element of the data cannot be opened up, that frankly is irrelevant to any attempt to define 'openness'.
"We should not be defining openness in any form, trying to second guess legislation whether from any one or multiple countries. The definition of fully open may mean that there are different legal or other choices creating qualifiers, particularly for data, meaning that there is also a level of partial openness.
"The reality and accuracy of this must be understood in assessing risk and liability. So for today, The Alliance announcement of a definition of open weight is a welcome one."
Have you got a story or insights to share? Get in touch and let us know.
