Related, I heard about curriculum learning for LLMs quite often but I couldn’t find a library to order training data by an arbitrary measure like difficulty, so I made one[0].
What you get is an iterator over the dataset that samples based on how far you are in the training.
> To learn, agents must experience high-value states, which are hard (or impossible) for untrained agents to reach. The endgame-only envs were the final piece to crack 65k. The endgame requires tens of thousands of correct moves where a single mistake ends the game, but to practice, agents must first get there.
This seems really similar to the motivations around masked language modeling. By providing increasingly-masked targets over time, a smooth difficulty curve can be established. Randomly masking X% of the tokens/bytes is trivial to implement. MLM can take a small corpus and turn it into an astronomically large one.
That's effectively what you get in either case. With MLM, on the first learning iteration you might only mask exactly one token per sequence. This is equivalent to starting learning at a later state. The direction of the curriculum flows toward more and more of these being masked over time, which is equivalent to starting from earlier and earlier states. Eventually, you mask 100% of the sequence and you are starting from zero.
What I like about this writeup is that it quietly demolishes the idea that you need DeepMind-scale resources to get “superhuman” RL. The headline result is less about 2048 and Tetris and more about treating the data pipeline as the main product: careful observation design, reward shaping, and then a curriculum that drops the agent straight into high-value endgame states so it ever sees them in the first place. Once your env runs at millions of steps per second on a single 4090, the bottleneck is human iteration on those choices, not FLOPs.
The happy Tetris bug is also a neat example of how “bad” inputs can act like curriculum or data augmentation. Corrupted observations forced the policy to be robust to chaos early, which then paid off when the game actually got hard. That feels very similar to tricks in other domains where we deliberately randomize or mask parts of the input. It makes me wonder how many surprisingly strong RL systems in the wild are really powered by accidental curricula that nobody has fully noticed or formalized yet.
Related, I heard about curriculum learning for LLMs quite often but I couldn’t find a library to order training data by an arbitrary measure like difficulty, so I made one[0].
What you get is an iterator over the dataset that samples based on how far you are in the training.
0: https://github.com/omarkamali/curriculus
Curriculum learning helped me out a lot in this project too https://www.robw.fyi/2025/12/28/solve-hi-q-with-alphazero-an...
Star the puffer https://github.com/PufferAI/PufferLib
> To learn, agents must experience high-value states, which are hard (or impossible) for untrained agents to reach. The endgame-only envs were the final piece to crack 65k. The endgame requires tens of thousands of correct moves where a single mistake ends the game, but to practice, agents must first get there.
This seems really similar to the motivations around masked language modeling. By providing increasingly-masked targets over time, a smooth difficulty curve can be established. Randomly masking X% of the tokens/bytes is trivial to implement. MLM can take a small corpus and turn it into an astronomically large one.
perhaps I'm missing something. Why not start the learning at a later state?
That's effectively what you get in either case. With MLM, on the first learning iteration you might only mask exactly one token per sequence. This is equivalent to starting learning at a later state. The direction of the curriculum flows toward more and more of these being masked over time, which is equivalent to starting from earlier and earlier states. Eventually, you mask 100% of the sequence and you are starting from zero.
If the goal is to achieve end-to-end learning that would be cheating.
If you sat down to solve a problem you’ve never seen before you wouldn’t even know what a valid “later state” looking like.
What I like about this writeup is that it quietly demolishes the idea that you need DeepMind-scale resources to get “superhuman” RL. The headline result is less about 2048 and Tetris and more about treating the data pipeline as the main product: careful observation design, reward shaping, and then a curriculum that drops the agent straight into high-value endgame states so it ever sees them in the first place. Once your env runs at millions of steps per second on a single 4090, the bottleneck is human iteration on those choices, not FLOPs.
The happy Tetris bug is also a neat example of how “bad” inputs can act like curriculum or data augmentation. Corrupted observations forced the policy to be robust to chaos early, which then paid off when the game actually got hard. That feels very similar to tricks in other domains where we deliberately randomize or mask parts of the input. It makes me wonder how many surprisingly strong RL systems in the wild are really powered by accidental curricula that nobody has fully noticed or formalized yet.
Those are not hard tasks ...