Epoch AI has published a piece saying AI systems are improving rapidly at Slay the Spire and its sequel. The claim is Epoch’s own conclusion, and it sits on the organization’s publications page. It makes AI Slay the Spire performance a concrete thing to watch, though the headline carries no figures to judge it by.
A card-battling roguelike where outcomes depend on planning under randomness, Slay the Spire is a natural fit for the growing practice of testing AI agents on games. The material available to BriefFlash gives no scores, model names or methods, so this article explains what such a test can show and what the headline leaves open.
What is Slay the Spire, and why use it to test AI?
Slay the Spire is a deckbuilding roguelike. The player climbs a tower, battling enemies with cards, and builds a deck during the run. Enemies and events are random, and each run ends in victory or defeat.
That structure rewards planning several steps ahead and weighing risk, which differs from answering a question correctly once. Every card taken or skipped changes the odds in later fights, and mistakes compound.
An exam question has one right answer, checked once. A run of this game is a chain of linked decisions where luck and skill both play a part. In our reading, that is why one win or loss says little, and why game results usually need many runs to mean anything.
This is general background, not content from Epoch’s piece. Game environments are one of the few ways to measure long-horizon planning, an area where exam-style benchmarks say little.
What did Epoch AI actually report?
Epoch’s claim, as far as BriefFlash can see, is its headline: AI systems are improving rapidly at both games. The piece is titled AI systems are improving rapidly at Slay the Spire and its sequel and is filed on its publications page. We have the title and nothing else, so we cannot say how strong the evidence is.
The organization tracks trends in AI, publishing data on models and compute. That suggests the claim rests on measurement, though that is our inference, and a record with data is no substitute for seeing the numbers behind this result.
Rapid improvement is also a claim about a trend, and a trend needs a starting point and a time frame. Neither appears in the material we have.
Why does the sequel matter for the test?
In BriefFlash’s reasoning, including a newer sequel makes the test more informative. A system that does well on the original may owe part of that to how widely the game has been discussed in the text models learn from. A sequel offers a way to probe whether progress transfers.
That is our inference, not Epoch’s stated reason. The headline does not say why the sequel was included, and the training-data explanation is plausible but unconfirmed. The sequel’s name and release details are not in our material, so we call it only the sequel.
What does the AI Slay the Spire headline leave open?
Most of what would let a reader judge the claim is missing. The graphic below separates what the headline establishes from what it does not state.
What we do not know yet comes down to five questions:
- Which AI systems were tested, and how they were set up
- Any scores, win rates or difficulty levels
- The time period over which improvement is claimed
- Whether tools, memory or added scaffolding were used
- Whether the results were checked independently
The scaffolding question matters most. Game performance can depend heavily on how a system is connected to the game and what help it gets. A rapid-improvement claim could describe better models, better scaffolding or both, and the headline does not separate them.
How does this compare with the same-day study on agents?
On the same day, a separate study saying AI agents remain far from autonomous appeared, which makes a useful comparison point.
Neither source claims a link to the other, and from the headlines alone they do not contradict each other. A game has a clear win or loss; open-ended work does not, which matters if agents overstate their results, as that study’s headline says. We would expect that difference to matter, but it is a hypothesis, not a finding.
What should developers and benchmark watchers look for?
Success at one game does not by itself show general autonomy. For developers building agents, a game result is most useful when the setup resembles their own: similar tool access, memory and guidance. If improvement came mainly from added scaffolding, it says more about engineering than about the underlying models. A fast trend on one pair of games is worth noting, and worth testing against other environments before anyone generalizes from it.
When the full results are read, check whether the systems and settings are named, whether scores come from many runs rather than a handful, whether the baseline predates the improvement, and whether anyone outside Epoch has reproduced the numbers. The results sit on Epoch’s publications page.
Frequently asked questions
What is Slay the Spire and why use it to test AI?
Slay the Spire is a deckbuilding roguelike in which a player climbs a tower, battling enemies with cards and building a deck during the run. Enemies and events are random. It rewards planning ahead and weighing risk, so it tests long-horizon decision making in a way a single exam-style question does not.
What did Epoch AI actually report?
Epoch AI published a piece titled AI systems are improving rapidly at Slay the Spire and its sequel, on its publications page. The rapid-improvement claim is Epoch’s own conclusion. The material available to BriefFlash contains only that headline, with no scores, model names, methods or time period, so the strength of the evidence cannot yet be assessed.
Does doing well at a game mean an AI is capable at real work?
Not by itself. Success at one game does not show general autonomy. Game performance can depend heavily on how a system is connected to the game and what help it gets. A result is most informative about real work when the setup, tools and difficulty resemble the work in question, and when others have reproduced it.
Why does the sequel matter for the test?
In BriefFlash’s reasoning, a newer sequel lets researchers check whether progress carries beyond the original game, which models may know well from text discussing it. That is an inference. Epoch AI’s headline does not say why the sequel was included, so the training-data explanation is plausible but unconfirmed.
