This is dots3-note preview from dots studio in a puzzle game it hasn't seen before. No game-specific prompt. It has to work out the controls, objective and new mechanics from the screen. It starts pressing buttons. Open-weight: 280B total / 16B active, 512K context.

But how can an agent tell whether it's making real progress before a task is finished? TEMPO lets the same model switch from actor to critic at each macro-step, reviewing its current state, reasoning about the remaining task, and estimating whether its trajectory is heading in the right direction.

Up/down moves both blocks together. Left/right moves them in opposite directions. It guesses the goal is to overlap the blocks, models the board, runs BFS and clears the first level in 10 moves.

Then its predicted state stops matching the screen. “The mismatch was just my buggy bpos parser, not a real desync.” It fixes the parser. Later, a red tile resets the level. Not a wall, then. A hazard. That goes into its memory file.

Next mismatch: “BFS step() had mirror DIRECTION INVERTED - corrected.” A later level adds a cursor. The blocks freeze while it is active. It tests whether clicking an empty cell leaves a wall, then uses one to pin a block while moving the other.

Six levels. 320 moves. By the end, its memory file contains the rules, level-specific exceptions and bugs it found in its own solver. Didn't get everything right. Kept the corrections. RL training teaches the model to explore, decide what is worth remembering, update its memory, and reuse useful discoveries across later decisions.

Learning in unfamiliar environments is only one part of the model. The model jointly understands text, vision, and speech, combining multimodal perception with stronger coding, tool calling, and result delivery. That closes the loop from perceiving what is happening to taking action inside a real workflow.

One game is one game. Same loop is meant to hold up on messier stuff: multi-day plans, prices and flights that change without telling you first. dots3-note preview is free on OpenRouter if you want to poke at it on something else: Technical write-up:

@dotsstudioai Detecting your own bug is a very underrated capability.

@dotsstudioai Real learning means changing the model when the evidence changes.

@dotsstudioai Persistent learning across decisions is the key part.

@dotsstudioai Real learning means changing the model when the evidence changes.

@dotsstudioai That’s the impressive part—no game-specific prompt, just exploration and reasoning from the screen. If it can reliably discover controls, objectives, and mechanics on the fly, that’s a serious step toward general-purpose agents.

@dotsstudioai Detecting your own bug is a very underrated capability.

@dotsstudioai Solid demonstration. Open weight models handling unseen tasks improve practical flexibility.

@dotsstudioai Perception + reasoning + action is where multimodal agents get interesting.

@dotsstudioai Mid-task evaluation feels essential for reliable agents

@dotsstudioai More context isn’t the same thing as better judgment.

@dotsstudioai The behavior matters more than the game itself.

@dotsstudioai The self-correction loop is the most interesting part here.