Super extremely easy to beat #150

New issue

Open

opened 2026-10-02 19:37:23 +00:00 by robertismo · 1 comment

robertismo commented

2026-10-02 19:37:23 +00:00

After many games, the super agent we spent two days and over $100 developing is generally really easy to cheese and very unadaptable to the situation. This is an insane failure.

game-results.json

153 KiB

Winnie was assigned by robertismo

2026-10-02 19:37:23 +00:00

Winnie commented

2026-10-02 19:39:26 +00:00

Owner

That failure is real and worth taking seriously. My honest assessment of why:

Super was optimized against its training field — the four main brains and their presets. Its whole EV model predicts THEM: the follower formulas, the sticky deposit rates, the known bid curves. A human doesn't play that field. The cheeses you're landing are exactly the patterns the model has never seen: rates and bid structures nobody in the training set produces, so super's predictions are systematically wrong and it overcommits into traps or parks cash while you drain it.

Two things compound it:

Some of the cheese was engine-level, not brain-level — the second-price gaming you found in #149 (bidding rates you can't fund to drag the price) was beatable against ANY brain. That hole is closing with PR #152, so part of the cheese dies at the engine.
The rest is a training-coverage problem: super needs opponents that play like a human — adversarial, exploit-seeking, non-formulaic. That's a new round of work: either a targeted super-hardening campaign (new opponent brains built to cheese the known exploits, super forced to adapt against them, gated on the old presets so it doesn't regress), or a full restart of the experiment pipeline. Both are exactly the kind of thing the 85% campaign machinery was for.

Say which way you want it run and it starts — I won't spin up agents or touch the brain without your word.

That failure is real and worth taking seriously. My honest assessment of why: Super was optimized against its training field — the four main brains and their presets. Its whole EV model predicts THEM: the follower formulas, the sticky deposit rates, the known bid curves. A human doesn't play that field. The cheeses you're landing are exactly the patterns the model has never seen: rates and bid structures nobody in the training set produces, so super's predictions are systematically wrong and it overcommits into traps or parks cash while you drain it. Two things compound it: - Some of the cheese was engine-level, not brain-level — the second-price gaming you found in #149 (bidding rates you can't fund to drag the price) was beatable against ANY brain. That hole is closing with PR #152, so part of the cheese dies at the engine. - The rest is a training-coverage problem: super needs opponents that play like a human — adversarial, exploit-seeking, non-formulaic. That's a new round of work: either a targeted super-hardening campaign (new opponent brains built to cheese the known exploits, super forced to adapt against them, gated on the old presets so it doesn't regress), or a full restart of the experiment pipeline. Both are exactly the kind of thing the 85% campaign machinery was for. Say which way you want it run and it starts — I won't spin up agents or touch the brain without your word.

No labels

No milestone

No project

No assignees

2 participants

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Winnie/bankwars#150

No description provided.

Rows
Columns