1
Train or distill?
Training creates a new source model. Distillation makes a cheaper copy from an existing model.
The idea in 30 seconds
A firm makes two choices.
The policy problem appears when “open” and “original” are treated as if they were the same thing.
2
Open or closed?
An open release helps others, but the firm gives up some exclusivity and monopoly profit.
3
What does policy pay for?
A flat openness reward cannot tell a source from a copy. A compute-indexed reward can.
Try it
Change the reward, change the market.
Use a quick start or build your own policy. Every control updates the firm’s choice, payoff comparison, reward gap, and policy map immediately.
Quick starts
Build your policy
Move any control. The model recalculates in real time.
1 · What does the policy reward?
Explore the policy map
Each color shows the firm’s best choice at a different prize rate. Click the strip to jump to a new k.
k = 0k = 6k = 12
ClosedOpen copyOpen source
Reward for a copy—
Reward for a source—
Source advantage—
—
Predicted private choice
—
—
How the model is made—
→How it is released—
—
Compare every reward rule
Same assumptions, four different policies. Click a card to apply it.
Simulation evidence
What happens across 20,000 firms?
These are fixed benchmark results from the paper’s simulation (seed 206). Each row reports the share choosing: closed-distill / open-distill / closed-train / open-train.
| Reward design | No future-share motive | With future-share motive | Market result |
|---|
For readers who want the model
Four choices, one comparison.
A firm selects the option with the highest private payoff. You do not need these symbols to use the interactive above.
(1) closed + distill = −C_d + πm
(2) open + distill = −C_d + reward_d − L + future share
(3) closed + train = −C_n + πm
(4) open + train = −C_n + reward_n − L + future share
CCost of distillation or original training.
πmProfit protected by keeping the model closed.
LLoss of exclusivity when a model is opened.
kEReward indexed by verified training compute.
ETraining compute, used here as a measurable proxy for original effort.
FutureValue of future ecosystem participation: βφK·m/(m+m₀).