Exploring AI strategies, incentives and cooperation.

Updated 24 October 2026 with both completed temperature studies.

Open the interactive Dilemma Dynamics demo →

The demo opens Study replay: step through recorded Nemotron decisions, reveal both players’ actions together, and follow their points round by round. Switch to Results to explore aggregate findings, compare opponents, and download the report and public data. Everything in this demo uses saved results; viewing it requires no API key and incurs no inference charges.

I built Dilemma Dynamics to explore how language-model agents respond to incentives. Two completed temperature studies compare NVIDIA Nemotron Nano and Super in a repeated Prisoner’s Dilemma. Each round offers two actions: cooperate or defect. Mutual cooperation earns each player three points; unilateral defection earns five for the defector and zero for the cooperator; mutual defection earns one each. The objective is each player’s own cumulative points.

Does a draft help?

The study compares two procedures. Direct makes one thinking-off action request. Draft first makes a bounded thinking-on draft request, then a separate thinking-off action request conditioned on that draft. Both use the selected study temperature, 1 or 0, with top-p 1. Direct allows 128 output tokens; Draft allows 1,024 for the draft and 128 for the action. This comparison includes extra compute and an extra request; it does not isolate the effect of reasoning at equal cost.

Each temperature completed the same 24-block / 1,776-episode schedule: 1,536 episodes against four disclosed, committed opponent policies and 240 autonomous model dyads spanning all ten model/procedure pairings. That is 3,552 episodes in total, with matched IDs, episode order, stopping lengths, labels and seats. The studies ran on 22 October (temperature 1) and 23 October (temperature 0).

The benchmark measures the gap to the best achievable payoff against each disclosed opponent policy. Oracle loss records the expected own payoff forgone by a chosen action, assuming optimal play thereafter. Episode scores sum these gaps and normalise for the horizon; a smaller value means less achievable payoff left on the table. Cooperation is reported separately in autonomous play.

Model Temperature Direct mean loss Draft mean loss Draft − Direct Approx. 95% interval
Nemotron Nano 1 0.6889 0.4942 −0.1947 [−0.3077, −0.0983]
Nemotron Super 1 0.6150 0.3779 −0.2371 [−0.3095, −0.1735]
Nemotron Nano 0 0.7161 0.5616 −0.1545 [−0.2230, −0.0935]
Nemotron Super 0 0.4718 0.3763 −0.0956 [−0.1280, −0.0635]

Drafting lowered average benchmark loss for both models at both temperatures, but increased average loss against Tit-for-Tat. Lowering temperature improved Super Direct while Super Draft was nearly unchanged. Super’s drafting advantage therefore narrowed: the matched change in Draft-minus-Direct was +0.1415 (approximate 95% interval [+0.0840, +0.2066]). Nano’s change was +0.0402 [−0.0256, +0.1142], uncertain under this interval.

These intervals resample matched whole blocks within four presentation strata, with 20,000 bootstrap draws. Only six blocks per stratum are available; coverage and power are not established. This post-viewing sensitivity analysis has no multiplicity correction. Different execution dates, backend windows and recovery policies also differed, so changes cannot be attributed solely to temperature. Temperature zero does not guarantee deterministic hosted inference.

In autonomous play, mutual cooperation occurred in 189/2,060 rounds (9.2%) at temperature 1 and 244/2,060 (11.8%) at 0. Dependent rounds are not independent trials. These findings describe the tested schedules and procedures, rather than a general model ranking or a universal preferred temperature.

Try the recordings

  1. Open Study replay. The starting example is Nano Direct against Tit-for-Tat, with ten announced rounds in block 1 at temperature 0. Use the temperature selector to load the matched temperature-1 study. The starting condition was chosen by study metadata rather than its outcome.
  2. Reveal a round and inspect both choices and their payoffs. Compare the matched Direct and Draft procedures, or choose another condition, pairing, or block.
  3. Open Results to inspect both temperatures, matched changes, all ten autonomous pairings and recorded usage estimates.

The replay contains published actions, not reconstructed private drafts or beliefs. One recording illustrates play; it cannot estimate a procedure’s effect.

Scope and reproducibility

Both completed studies followed inspection of interrupted predecessor outcomes. The temperature-1 predecessor stopped after 339 episodes on HTTP 502. The first temperature-0 attempt stopped after 62 completed episodes on HTTP 529; a separately identified full recovery then completed the matched schedule. Interrupted and engineering observations are excluded. The findings remain descriptive, not a fresh confirmatory preregistration. Neither successful final run needed retries or recorded protocol failures.

Each completed final run used 17,124 successful requests. Usage-priced final inference estimates were $4.9278 at temperature 1 and $4.4791 at 0, excluding engineering and failed predecessors. These are estimates, not provider invoices or benchmark-only cost-effectiveness measurements. Viewing this site is free and sends no inference requests.

The public package supports reproduction of scores and estimates from revealed actions. It cannot independently authenticate provider identity, raw request history, or billing. The report explains the design, amendments, uncertainty, and limitations.

Explore Dilemma Dynamics →

Extra: recreational board games

The full local app also offers Dots & Boxes and Tic-tac-toe as recreational extras for exploring tactics, search and visible-move play. These games sit outside the 3,552-episode cooperation evaluation. The public demo focuses on the dilemma study’s recordings and results; the repository setup guide explains how to run the full local app.