The current AI discourse is dominated by talk about products, APIs, and closed platforms. This, in fact, distorts what AI is in a more academic and technical sense.

Allow me to introduce another approach:

The website Hugging Face is not just another AI tool. It is more of a library containing models, a workshop where you can tinker with LLMs, and an archive for storing them. Additionally, it represents infrastructure, commons, and material access to LLMs.

Started as an NLP (Natural Language Processing) company and focussed on open (vs proprietary) models and tooling, nowadays, it functions as a social layer around machine learning artifacts.

Not a chatbot, a single model, or even a consumer-facing AI product. This is for the tinkerers, the makers, the creative adventurers.

Whereas OpenAI / Anthropic / et al. offer finished systems, controlled interfaces, and opinionated usage, Hugging Face gives you raw components and multiple approaches, though with visible trade-offs.

If you are interested in learning how AI works, what it can do, and how you can adapt it to your own workflows and purposes, Hugging Face is worth a visit.

So, the models you find on the site are not "apps". They are building blocks, raw material with certain, distinct properties. Those properties can affect the LLM's output in regard to tone, bias, latency, size, failure modes, and strangeness/uniqueness.

As designers / creatives / artists let us think of the model hub as our palette, our library of instruments, or a set of lenses through which to observe the world.

We can make use of what the site offers in multiple ways for our creative praxis. We can browse models and regard them as cultural artifacts, and then playfully compare how different models interpret the same prompt. This way we can study bias, failure, and style difference. These learnings could get translated into workflows which inform and support our work.

The Spaces feature allows us to do live demos, create safe sandboxes and sketchpads. It's prototyping but without the engineering team! This way we can test interactions, explore behaviors, and validate ideas before going into production. I am sure you are already sensing how powerful this can be for solopreneurs and small teams / collectives.

LLMs use datasets, collections of training data. Those determine what an LLM "knows", how it "sees the world", how it reacts to prompts, etc. Hugging Face treats these datasets like authored objects. Others can curate, remix, restyle, alter these datasets and by doing so expose hidden assumptions within them.

Providers of LLM interfaces hide their datasets as best as they can, they are their "secret sauce". Hugging Face gives you a list of ingredients, a recipe book, and a kitchen to cook up your own.

When their creators do not think of the models' usefulness first, "weirder" ones can come into existence. They hallucinate differently, fail more poetically and by doing so reveal their structures instead of hiding them skillfully. Via experimentation with these models, we not only learn a lot about LLMs but also much about our own thinking. It was our very own thought processes which created AI, so it seems only natural that they get reflected back to us while using LLMs.

Our current culture is getting shaped by AI if we agree with that or not. If we leave that process to the manufacturers of centralized systems, the danger of cultural control via a small amount of players rises. Hugging Face preserves the much-needed multiplicity. It allows for forking paths, experiments, failure, and visible disagreement between systems.

This way AI stays interpretable, contestable, and culturally negotiable.

If we want to take part in how AI shapes our culture, we need access to its building blocks.

So, Hugging Face the AI utopia?
No. The site can be messy and uneven, beginners are often overwhelmed, regulars are too. You cannot bet on models working as expected, or at all. Documentation can be scarce, so one needs high motivation to get anywhere.
If you are so inclined, the chaos is not a negative. The friction you might experience reveals structure, lets you peek behind the curtain. Polished experiences hide power behind ease-of-use.

Hugging Face is not suited to completely replace AI products. It allows us though to learn at our own pace, under self-created conditions about LLMs. It gives us the much-needed critical distance and room for creative misuse. That freedom can lead to surprising results and newly informed decisions vis-à-vis the use of AI products.

Tl; DR:

If ChatGPT is AI as interface,
Hugging Face is AI as material.