TransientAI · independent research

Learning systems that run on ordinary hardware.

TransientAI is a small research effort building Transient, an open, experimental learning system written in Rust. The aim is capable AI that almost anyone can run, study and teach on their own machine. We publish everything, mistakes included, so others can replicate it, check it and reuse it.

Read the first post Meet the prototype Source on GitHub

Active research · early stage Started 2026-10-15 Rust, zero dependencies AGPL-3.0-or-later 92 tests, CI on every push

What we are working on

Most AI today is trained once in a data centre, frozen, and served from one. We are looking at a different corner of the design space: small systems that keep learning on the device they run on, and whose cost we can count.

Adaptive learning

A system that learns while it is used, and decides for itself how to learn each kind of thing: memorise it, fit a rule, or search for a small program.

Measured in joules

We count operations, memory traffic and estimated energy, not only parameters, because energy is what decides whether "on your own hardware" is real.

Honest about being wrong

Wrong answers are reported apart from refusals. Guesses are labelled. Mistakes and rejected ideas are written down as they happen.

The prototype, in action

baby is the first working piece of Transient. It is taught from scratch in about a third of a second, then answers a question and says how it knew, or refuses and says why. This is copied from a real run of the current build.

baby ask

baby ask "what is two times four"

eight

(unit 'times', worked out by an expression I built, (first x second) (100% on facts I held back))

baby ask "the capital of peru"

lima

(unit 'capital', recalled, not worked out: nothing I tried could work these out, so I only remember what I was told)

baby ask "what is seven times one"

probably seven

(a guess, not a fact: unit 'times', for this kind I was only taught (or have checked) numbers from 0 to 5, and this asks about 7: I won't state it as fact unchecked ...)

baby ask "three minus five"

I don't know this one: unit 'minus', I couldn't work it out: it has a word or a shape I haven't learned how to handle. Teach me: baby teach "three minus five" <answer>

baby ask "what is -3 plus 5"

I only read plain whole numbers, not '-3' (signs, decimals, fractions, percentages).

Multiplication was taught on the numbers 0 to 5 only, so "seven times one" is a labelled guess and not stated as fact. The long explanation after it is trimmed here. The last two are refusals on purpose: a negative result and a number it cannot read are things it must not bluff.

A few numbers, with their caveats

From the synthetic lesson set the prototype is developed on, on one CPU core of an ordinary laptop with no GPU. Energy figures are estimates from counted operations; trust the ratios, not the absolute joules.

0.0%

wrong answers on every row of the scorecard (taught, reworded, never taught, beyond range). It refuses or labels a guess instead.

8,977 → 57

operations per answer after changing how it finds the right kind of lesson: about 85× less estimated energy, same answers.

2.65 s

to teach 25,776 lessons across 600 kinds of lesson (56 MB peak) in the scale test.

~0.5 ms

to learn one more fact, at every size tested from 20 to 600 units.

These come from tidy, synthetic lessons. Nothing here shows behaviour on real text at scale, or that this approach can reach the quality of today's language models. What we do not know yet

Latest from the blog

2026-10-21

How Transient began, and where it stands

The starting hunch, the first week of work, what the prototype does and does not show, and what comes next.

Read the post

Open on purpose

Everything is public under the AGPL, so that anyone who runs a modified version, including as a service, has to share their changes. There is no contributor licence agreement: contributions stay under the same licence.

Right now TransientAI is one person. There are no plans to grow the core team, but contributors to the repository are welcome.

How to contribute