Week 8: GTO versus exploitative play

What "unexploitable" actually means, formalized as a zero-exploitability property of week 7's equilibrium, versus the higher-EV but riskier alternative of deviating to exploit a fixed opponent

12 April 2027

What “optimal” means once an opponent adapts#

Week 7 derived x* = 25% and c* = 66.7% for one specific toy game, as the frequencies that make each player indifferent. This week asks: what is that indifference actually buying, and is it always the right target?

GTO, defined#

A pair of strategies is a Nash equilibrium — what “GTO” means here — if neither player can raise their own EV by unilaterally changing their strategy, given the other’s strategy stays fixed. The defining property that matters at the table: an equilibrium strategy’s EV cannot be pushed below the game’s value, no matter what the opponent does. That’s the technical content behind “unexploitable.”

Best-responding to a fixed opponent#

Contrast that with a different question: given a specific, known, fixed bluffing frequency x from villain — not villain playing optimally — what should hero actually do? Hero’s EV(call) − EV(fold) is linear in x (from week 7’s own formula), which means the best response is almost never a mix — it’s pure:

• If x > x* (villain overbluffs): call always (c = 100%), not 66.7%.
• If x < x* (villain underbluffs): fold always (c = 0%), not 66.7%.
• Only exactly at x = x* is hero indifferent, which is precisely the equilibrium case week 7 solved.

A known, fixed tendency is always more profitably exploited by a pure strategy than by the equilibrium mix — the mix is only correct when the opponent might also be adapting.

Exploitability, formalized#

Define the exploitability of hero’s calling frequency c as the most extra EV villain can extract by bluffing as often as profitable, given c:

exploitability(c) = max(0, pot − c · (pot + bet))

• At c = c* = 66.7%: exploitability = 100 − 0.667×150 = $0 — the equilibrium calling frequency cannot be exploited at all.
• At c = 50% (a hero who folds too much): exploitability = 100 − 0.5×150 = $25 per hand — the exact amount a villain willing to bluff 100% of the time can extract.

This is what “GTO guarantees the game’s value regardless of opponent” means numerically: zero exploitability is a property of one specific frequency, and every other frequency has some fixed dollar cost attached to it, computable in the same way.

The actual trade-off#

Exploitative play’s appeal: against villain’s real x = 50% (well above x* = 25%), calling every time earns hero $50 per hand, against the $33.33 that defending at the equilibrium 66.7% earns from that same villain. Its cost: that extra $16.67 depends entirely on villain actually bluffing 50% of the time continuing to be true. If villain notices hero calling too much and stops bluffing, hero’s edge becomes hero’s own exploitability instead — the same formula, now pointed the other way. GTO play never needs that assumption to hold; exploitative play is a bet that it will.

Before next week#

Every toy game this course has built has been exactly two players. Week 9 adds a third, and asks which parts of everything so far — pot odds, Bayesian updating, MDF — survive that change and which quietly assumed only two players all along.