Progress
Where things stand, and what comes next
Transient started on 2026-10-15. This page has the roadmap first, with a status on each item, then the dated log. It is a summary; the full record with numbers is the repository's experiment log.
Roadmap
Ambitions here are kept in proportion to what has been shown. Items are checkpoints: after each we ask what the result taught, and a result that disappoints tells us where to look next. It does not end the work.
| Step | What it is | Status |
|---|---|---|
| Prototype that chooses how to learn | baby: units of lessons, mechanisms chosen from held-back evidence, honest refusals | done |
| Learning that stays light | Flat cost per new fact up to 600 units; noise handling; digits as numbers | done, on synthetic data |
| Careful by default, labelled guesses | States as fact only what it was taught or has checked; keeps a record of its own guessing | done |
| Bridge to real data | A harness for real text (a word-problem corpus, or question/answer files) is built and tested on hand-written fixtures. Not yet run on a real corpus. Expect it to refuse most of the first one; the reasons are the interesting result. | next |
| Controlled comparison | A matched small transformer against baby-style systems at three or more sizes: quality at equal joules, retention, adaptation speed, scaling slope. Needs a rented GPU. | next |
| Real energy readings | Wire the CPU's energy counter into the ledger | next |
| Corrections that teach the new regime | Today one correction stops a wrong guess; it should also let the system learn what is right from a few points | open |
| Mechanisms of a different kind | Not just formulas: something that is not an expression search | not started |
| Growth | A unit that starts tiny and adds capacity only when it saturates | later |
| Learning at a scale and quality that matter | Language-like tasks on a consumer device, judged in capability per joule. A long-term ambition. We do not know whether this architecture can get there, and the repository exists to find out by measurement. | ambition |
What we do not know yet
These are not boilerplate. Each is something we would want to know before believing the headline numbers.
Real text
Everything so far is synthetic and tidy. Nothing shows behaviour on millions of real sentences, or that an architecture like this can reach quality that matters.
Whether "0% wrong" holds
It is a spot check. Every harder test we built found a new way for it to fail, so we expect more.
Whose language it is
The mechanisms, operations and thresholds are our choices. The system builds and selects inside them. A system that finds mechanisms of a new kind is not yet shown.
Real teachers
The experiments use a teacher who is always right. A real teacher is sometimes wrong, and a wrong correction currently switches a unit's guessing off.
Joules
Energy is estimated from counted operations. Real measurements so far are wall time and peak memory on one laptop core.
Data gaps
A lone explanation that fits everything taught can still be wrong at a number nobody taught. Checks and review catch it in our batteries, but nothing structural prevents it.
Timeline
Newest first. Dates are when the work landed; numbers come from the experiment log and carry its caveats.
The bridge to real text, and the TransientAI name
A harness that loads real corpora (SVAMP's JSON, or question/answer files), sets aside what the prototype cannot read and says why, teaches a share, and grades the rest, listing every wrong answer. Waiting on a corpus. The project is named Transient; TransientAI is its creator.
Guessing like a scientist
Careful became the default. Beyond what it was taught or has checked, the baby offers a labelled guess, keeps a record of how often its guesses at each distance are right, and stops guessing where they have been wrong more often than right. A correction ends the guessing in the region it contradicted. Testing this found a hole (a box-shaped reach vouching for an unchecked corner: 4 wrong facts of 9,309) that was fixed at a measured cost in coverage.
A hole found, and a careful mode
A rule that fits everything taught but changes behaviour outside it (a sum that caps or wraps) was answered confidently and wrongly: 53 wrong answers per seed. Careful mode removes them and, after two rounds of the baby's own self-checks, wins back 99.2% of genuine extrapolation with 0 wrong.
Tuning its own bar, adapting its effort
It measures the share of wrong lessons in its own data (it estimated 9.6% where the world had 9.7%) and sets how reliable a mechanism must be before answering. At 10% label noise, unseen questions answered went from 42% to 98%. Searches that find nothing now fail cheaply (battery 23.5 s to 14 s, same results).
Asking, inventing, and noise
The baby keeps competing explanations and asks the number that would separate them. It worked out all 16 hidden numeric kinds and 7 of 9 harder ones, with 0 false inventions. With 2% wrong lessons it answers 0% wrong and 97 to 100% of unseen questions right.
Scale
The world grew to 600 units, 25,776 lessons and a 10,000-word vocabulary. Learning one fact stayed at about half a millisecond; teaching 26k lessons fell from 127 s to 2.65 s after the per-lesson tables were made incremental.
Published, open
The work went public under the AGPL, with a technical README, CI that fails if the scorecard's WRONG column rises above 0.0%, and a developer certificate of origin instead of a contributor licence agreement.
Storage, discovery, light learning, less effort
The baby found units itself from unlabelled lessons (9 from 792). It can forget what is already explained (49% of the file kept, still answering every lesson). Answering fell from 8,977 operations to double digits, and a rule is fit exactly rather than iterated (392× fewer operations).
One baby
The separate experiments were unified: 13 units, five mechanisms chosen per unit by evidence, a pipeline that answers with a reason or refuses with one. Scorecard on five seeds: 0% wrong.
Readers, refusal, rules, and a library of ways to learn
Position-free readers mattered more than any learning trick. The baby learned to say it does not know, fitted a rule for sums from examples (right beyond the range it was taught, where memorising networks got 5%), and began choosing among mechanisms from held-back evidence.
The first baby
A 16k-parameter, from-scratch, zero-dependency Rust network that answers a query with one word, plus a joule ledger. A fact store matched retraining's retention at roughly 1/900 of the estimated energy.