About
A small, open research effort
TransientAI is the creator of Transient. It is an independent effort, founded in 2026 and owned and run by Daws Macs. At the moment it is one person.
Mission
To make capable intelligence something almost anyone can have: running on hardware they already own, learning from them privately, and costing very little to keep. If only a few providers with huge infrastructure can offer capable AI, that is a choice of design and not a law of nature, and we want to find out how far another design can go.
We also share our progress and development openly, so anyone can replicate it, check it, build on it or prove it wrong. That is why the code is public under the AGPL, and why this site publishes failed ideas next to successful ones.
What we do, and do not, claim
Transient is early research. It has a working prototype, baby, that learns small, tidy, synthetic worlds and is measured in operations, memory and estimated energy. It is not a language model, and we do not claim it will become one. The long-term ambition is a system whose usefulness you would trust in everyday work and study, on an ordinary device. Whether this approach can reach that is unknown, and the work exists to find out by measurement.
Who is behind it
Daws Macs is the creator, founder and owner of TransientAI and of the Transient project, and the only person working on it for now. There are no plans to bring others into the core of the effort yet. Contributors to the repository are welcome: replications, bug reports, adversarial questions and patches. See how to contribute.
Note
A note on how the work is done. Much of the implementation and writing is done with an AI coding assistant (Claude, from Anthropic) working under Daws's direction, and the commit history says so. The direction, the decisions and the responsibility are Daws's. We mention it because several limits listed here, such as the mechanisms and thresholds being hand-chosen, are about exactly this kind of authorship.
How we work
- Honesty over polish. Wrong is reported apart from refused; caveats are stated plainly; mistakes are listed. See how we judge results.
- Pessimism is input, not a brake. Risks are treated as terrain: they tell us what to measure first. We do not promise success, and we do not stop because the work is hard.
- Measure first. Cost is counted in operations, memory and joules, not only in parameters.
- Rust-first, not Rust-only. See why Rust.
- Open by licence. AGPL-3.0-only, with a developer certificate of origin and no contributor licence agreement, so nobody, including the founder, can quietly close what others contribute.
Questions
Is this a chatbot?
No. The prototype answers short questions about the lessons it was taught, says how it knew, and refuses when it should.
Why is it called Transient?
That one stays a mystery for now.
Can I use or build on it?
Yes. The code is AGPL-3.0-only: you can use, study, change and share it, and anyone who distributes a modified version or lets others use it over a network must offer the source under the same terms.
Where do I report a problem or share a result?
In the repository's issue tracker. If you ran an experiment on different hardware, we would especially like to hear the numbers.
Where can I follow progress?
On the progress page and the blog, or by watching the repository.