Open models
Foundation and task models with meaningful access to method, weights or both, under licenses that permit real use.
Ecosystem
Who builds this, and how open it really is.
Laboratories, models, products, infrastructure, communities and users are six different kinds of thing. Telling them apart — and being precise about what "open" means — makes the whole field easier to think about.
Our reading
Six kinds of thing, routinely confused.
Public conversation about AI tends to flatten everything into one list — as though a research laboratory, a chat assistant and a chip designer were three competitors in the same race. They are not the same kind of object, they answer to different pressures, and confusing them makes the whole field harder to think about.
The italic line under each is what that layer fundamentally is.
Research labs
Organisations that train models and publish research — companies such as OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral AI or xAI, alongside university and public research groups. They differ enormously in size, funding model, openness and stated purpose.
An organisation. Has strategy, incentives and constraints.
Models
The trained artefacts themselves. A single lab ships several, they are versioned, they are deprecated, and their capabilities move faster than any description of them survives.
An artefact. Not interchangeable with the company that made it.
Products and assistants
The interfaces most people actually meet: chat apps, coding assistants, search integrations, features embedded inside other software. A product wraps a model in a system prompt, tooling, memory, safety layers and a business model — all of which shape the experience as much as the model does.
An experience. What you judge is rarely the model alone.
Infrastructure and hardware
Chip designers such as NVIDIA, cloud platforms, data centres, networking and energy. This layer decides what is physically and economically possible, and is where concentration is most acute.
A dependency. Almost everything upstream rests on it.
Communities and open ecosystems
Model hubs, open source projects, independent researchers, maintainers, documentation writers, hobbyists. Rarely funded in proportion to how much the rest of the field relies on them.
A commons. Maintained by people, often unpaid.
Deployers and users
Companies, public services, schools, clinicians, artists, individuals. The layer where consequences actually land, and the one least represented in the conversation about how any of this should work.
The point of it. Frequently the last consulted.
On the names above. They are illustrative, not exhaustive, and the order carries no ranking. Naming an organisation here implies no relationship with it, no endorsement of it, and no criticism of it — and no organisation named has any involvement with this project. There is deliberately no "best AI" list on this site: the honest answer depends on the task, the constraints and who is asking.
Openness
Four things that are not the same.
"Open" is doing an enormous amount of work in AI discourse, and it hides distinctions that matter a great deal in practice. Being precise here is not pedantry — it changes what you are actually allowed to do.
Free of charge
Costs nothing to use. Says nothing about what you may inspect, modify, self-host or rely on tomorrow. A free product can close at any time.
Open code
The surrounding software is published. Useful — but for a machine learning system the code is often the least decisive part.
Open weights
The trained parameters can be downloaded and run. This is what most "open models" actually offer, and it is genuinely valuable: you can self-host, study behaviour and stop depending on someone's API. It is not the same as knowing how the model was built, and licenses attached to weights often carry real restrictions.
Open source AI
A stricter idea, covering the data, the training process and the freedom to use, study, modify and share. The Open Source Initiative published a formal Open Source AI Definition in 2024; it remains debated within the field, which is itself worth knowing.
Our reading
Preserving as many open, accessible and independent options as possible is one of the things this project cares about most — because it is the practical answer to concentration, because it lets researchers study systems rather than only their outputs, and because being able to change tools is what makes any of the other freedoms real.
But openness is not a virtue certificate. Published weights can be unmaintained, poorly documented, released under licenses that quietly forbid what you need, or shipped with fewer safeguards than a closed equivalent. Meanwhile several proprietary organisations fund and publish substantial safety research. Both things are true at once, and a position that cannot hold both is not worth much.
Where openness actually lives
If open alternatives matter, these are the layers that decide whether they survive. Most attention goes to the first one; most fragility sits in the fourth and fifth.
Open datasets and memory
Reusable, well-documented data with clear provenance and licensing — the least glamorous and most decisive layer.
Evaluation and safety tooling
Benchmarks, evaluation harnesses and safety tooling the whole ecosystem can inspect rather than take on trust.
Interoperability
Frameworks and protocols built so that tools can talk to each other, and so that leaving one is possible.
Maintenance
Support for the maintainers of libraries and infrastructure everything else quietly depends on. This is where open ecosystems most often break.
Documentation
Documentation, reproducible recipes and onboarding — what turns published code into something a stranger can actually use.
These are areas this project watches and writes about. ORÆ does not fund, select, certify or coordinate any of them, and no grant, programme or financial support is offered or implied.
Related work
A connected exploration.
The same person behind ORÆ is separately working on OSAIRIS — a portable identity layer meant to keep memory, tone and personality stable when the underlying AI model changes. It is in development, in private beta, and it is mentioned here for transparency about who writes this site.
It is not a validation of anything on this site, it does not define what ORÆ is about, and it is not offered as an example to follow. Treat it as a disclosure of interest.
What could make this healthier
← Back to ORÆ