There seems to be a lot of interest in the idea of local models, and the idea of having a model that is trained on your own data or the data of your community. Particularly as the corporate foundation models make it increasingly apparent that they will always have a bias, likely towards the interests of their owners, whether that be a corporation or a government. The idea of having a model that is trained on your own data or the data of your community is appealing, and it raises questions about what it means to have a model that is truly representative of your own experiences and knowledge. This is particularly relevant in the context of the recent discussions around the biases present in large language models. Like Grok going crazy every so often and slipping its moorings and going off on South Africa and white genocide, or Claude doing its Golden Gate Bridge thing.

But what does this mean in the context of, you know — bedrock knowledge? What happens with the notion of ‘common sense’ when epistemologies are derived from diverse knowledge bases? What happens when the knoweldge base is not a single, monolithic entity but a collection of potentially conflicting perspectives?

Nominalism considers that the world is made up of individual objects, and that these objects are the only things that exist. This means that there is no such thing as a universal truth or a single reality. Instead, each object has its own unique properties and characteristics, and these properties are what define the object. In this context, the idea of a local model becomes even more interesting. If each model is based on its own unique set of data and experiences, then it is possible that each model could have its own unique perspective on the world. This raises questions about what it means to have a model that is truly representative of your own experiences and knowledge.

Large language models are 'foundational' often seemingly simulating universals. They emanate generalized and averaged ‘truths’ and ‘common sense’ that are not necessarily representative of the experiences and knowledge of a specific community or individual — or the specific and particularities of a given context.

THe particular — local(ly trained) langugage model — doesn't travel very well to the general, as I think we've been learning over the last few years — likely acutely, but probably an effect that has been around for much longer and less noticed or not as acute.

I think the idea of nominalism that we're working on in the OWG is exposed by this speculation about what local language models (or this design fiction of Heirloom Language Models) might look like. The idea that each model is based on its own unique set of data and experiences, and that these models could have their own unique perspectives on the world, is curious — and I'm not sure what world it builds. It feels like they make apparent a world where its been accepted that truth, ways of reasoning, ways of knowing and — as I try to intimate here — even language(s) always have to be named and constructued froma particular vantage point, or particular perspective.

In tension this contrast between foundational and nominalistic fractures the illusion that foundational LLMs are just ‘mirrors of the world’. Rather they are seen as constructors of worlds or world-building — each with their own naming logic, their own epistemological commitments, typically coming from direct or indirect preferences, biases, or organizational/institutional interests.

But — now what? Nominalist approaches to knowledge and representation is a bit of a mystery to me at the moment.

Incommensurability feels like its a real thing. What happens when two local models have fundamentally incompatible beliefs? Can they translate? Do we need meta-models to negotiate between them?

Epistemic tribalism which is basically what we might feel now, quite acutely. Do we risk digging deeper and hardening the silos of belief such that consensus, agreement, possibility of dialogue becomes nigh impossible? Authority. Like..wtf. If every group has its own "truth engine," who arbitrates between them? What happens to rule of one law? Anyone? Bueller? Is it just a free-for-all? Do we need a new kind of epistemic authority to help us navigate this world of local models? Or do we just accept that there will be multiple, potentially conflicting perspectives on the world, and learn — somehow — to live with that? Is this for a generation yet to come, or emerging that is born into such a world, rather than mine and yours in which we are trying to deal with a world we were definitely not born into?

Epistemologies in code, oh my. Is this what they mean by technoauthoritarianism? Or is this just a natural evolution of the way we think about knowledge and representation? Is it possible that the rise of local models could lead to a more pluralistic and inclusive understanding of knowledge, or will it just exacerbate existing divisions and conflicts?

Perhaps time for a workshop or General Seminar on the topic!

Heirloom Language Models is a speculative design fiction artifact that imagines a future where vintage language models are found at estate sales, auctions, perhaps at flea markets or found while excavating, renovating, in lock boxes drilled out after the owner has died. They may be inscrutible, like knowledge-based time capsules that can emanate and attempt to describe themselves, the contexts from which they derived/came, etc. This assumes that local models — not foundation or corporate models — become a thing, which trends seem to indicate certainly the interest and definitely the technical possibility. One can imagine the tools and resources necessary to train a model could diminish over time making it possible to 'train' at home. There would be a world of models with unique regional data and personalities. The implications in this speculative fictional artifact — a classified advertisement — are meant as a wink of a consideration of how local knowledge might be contained in such artifacts. Imagining a world in which there are not canonical models or models who might contain the biases of the models' owners/operators — but more local sensitivities. A cacocphany of models and their emanations...