Intelligence,
in the making.

The second version of a learner trained only by playing Terraria found 7.0 kinds of items per game on 13 worlds it had never seen; its untrained starting point found 4.9. It beat that start in every world, in a test fixed in advance that an independent AI reviewer judged a pass. Most of the gain is one skill: digging.

A general-purpose learner plays the game Terraria with ordinary controls. It reads the game as numbers rather than pixels: the tiles around it, what it carries, nearby creatures and the options in open menus. It is never told what any tile, item or creature is. Its main reward is for holding kinds of items it has not held before in that game, with small rewards for gathering and for reaching new areas, and a penalty for dying. The connection to the game is written for Terraria; the learner itself is not. It trained for three hours on five worlds. Then we switched learning off and tested it on three worlds it had never seen, against random play and against its own untrained starting point. The pass mark was written down before training began.

Averages of 6 games each. The pre-registered pass mark was a lead of at least 1 item over both random play and the untrained learner. It met that mark exactly: 2.8 items over random play and 1.0 over its untrained self. Most of the lead over random play was already there before training, and the 1.0-item gain from training could be chance. A second pre-registered test failed: it beat random play in 4 of 6 paired games, short of the 5 we required.

Found

Training moved the learner the right way on worlds it had never seen, but only just: 1.0 item per game over its untrained self, exactly at the pass mark and possibly chance. It also explored far more of each world: 45.7 new map areas per game, against 10 to 12 for the others. Exploring was part of what training rewarded, so this shows the training worked as designed rather than a separate discovery.

Why

Our first attempt was hand-built: 19,115 lines of game-specific rules in three days. In its last 11.8 hours of play, 94.4% of its decisions were a single forced move that the game then refused. A learner offered one option has nothing to learn, so we started over with one that learns from play.

Open

Only 3 worlds and 6 games per player. A second version, playing under new rules where deaths respawn, has since finished: it beat its untrained start by 1.5 items (7.5 against 6.0) but failed the pre-registered chance test (p = 0.095), so an independent AI reviewer judged it a fail. Under the same rules it showed no clear gain over the first version. A third version, which changed only how it aims, averaged 5.5 items against 7.5 for the second and failed both pre-registered tests. The second and third versions were designed after we had seen these test worlds. Those tests reused three worlds. On 13 new worlds, made after its test plan was fixed, the second version then found 7.0 items against 4.9 for its untrained start and beat it in every world; a separate AI reviewer judged that pre-registered test a pass. A fourth version, which started out inclined to keep holding its current item, then did worse than the second on 13 more new worlds (6.3 against 7.6 items) and was judged a fail. A fifth version, which restored each game copy's original training world after every 4 of that copy's games, averaged 7.7 items per game against 7.3 for the second on another 13 new worlds; that comparison was inconclusive. Since then (to 2 November 2026), a sixth version failed its pre-registered test, and four rounds of an automatic improvement loop found no improvement (two failed, two inconclusive), so no later version has yet beaten the second by the margin its test required; the dated results are in the lab notebook. We have not measured how quickly first-time human players reach the same milestones, so we make no comparison with people yet.

An independent AI reviewer from a different provider recomputed all 18 games and passed the main test only, not the full pre-registration. No outside human team has reproduced it.

A transparent model built from public data such as Statistics Canada tables, with every assumption visible and changeable. Companion notes cover construction, industrial electricity, mine haulage and production evidence.

Open

No achievable national or household gain has been estimated yet. The official GDP and population figures we retrieved did not fully reconcile with the published GDP-per-person figure. An older population release reproduces it for one quarter, but the exact cause is unconfirmed, so we do not use that series for household estimates.