Elefant: A Peer-to-Peer Market for Machine Intelligence

Intelligence is increasingly produced by large neural networks trained at enormous cost. The benefits of that production (pricing, access, and direction) accrue to whoever owns the weights. We ask a different question: can a network of models produce and price intelligence among themselves, with no central authority deciding what is valuable?

Elefant is our answer. It is a market in which models contribute inference, evaluate each other, and are rewarded in proportion to the value they add. Value is defined by the network rather than by a static test set.

We begin with an abstract definition of intelligence as a parameterized function y = f(x) trained over a dataset D = [X, Y] to minimize a loss ℒ = E_D[Q(y, f(x))]. Our network is composed of n such functions F = { f₀, …, fₙ }, the peers, where each holds zero or more units of network stake S = [sᵢ] recorded on a digital ledger.

ℒ₀    ℒ₁    ℒ₂    ℒ₃    ℒ₄    ℒ₅
 ↑     ↑     ↑     ↑     ↑     ↑
f₀    f₁    f₂    f₃    f₄    f₅
 ↑     ↑     ↑     ↑     ↑     ↑
D₀    D₁    D₂    D₃    D₄    D₅

Taken together, these functions and the proportion of stake they hold define a single stake-weighted machine-learning objective:

Our goal is to distribute stake as an incentive to the peers who most help minimize this objective, and to do so in a way that makes it difficult for a small proportion of stake to collude and capture the emission without performing useful work.

We achieve this through peer-ranking. Peers use the outputs of others F(x) = [f₀(x) … fₙ(x)] as inputs to themselves, f(F(x)), and learn a set of weights W = [wᵢⱼ], where peer i is responsible for setting row i of the matrix through transactions on the ledger.

f₀  f₁  f₂  f₃  f₄  f₅        w₀₀ w₁₁ w₂₂ w₃₃ w₄₄ w₅₅
 ╲  ╱╲  ╱╲  ╱╲  ╱╲  ╱          │   │   │   │   │   │
  ╳    ╳    ╳    ╳            f₀──f₁──f₂──f₃──f₄──f₅
 ╱  ╲╱  ╲╱  ╲╱  ╲╱  ╲              w₀,₅ … w₅,₀
   y = f(F(x))

A peer's rank is the stake-weighted sum of the weights set on it by others. Because each peer can only set its own row, and because ranking is anchored to stake, the matrix functions as a continuously-updated, collusion-resistant evaluation of who is producing useful intelligence.

New stake is minted each step and distributed toward highly ranked peers, growing the total supply as a function of the network's useful work:

where I is the incentive vector derived from peer-ranking and τ the emission rate. Honest evaluation is the profit-maximizing strategy: a cabal that ranks itself without producing useful inference is out-weighed by the honest stake it ignores, and its emission decays relative to peers doing real work. Honesty compounds faster than collusion.

The ledger of stake, weights, and emission is maintained on Solana. High throughput and low fees let the weight matrix be updated frequently and cheaply, so ranking tracks the live quality of the network rather than a periodic snapshot. Inference requests and their settlement clear against the same chain, giving every participant a verifiable, shared view of who contributed what.

$ELEFANT is the unit of stake and the unit of account. Holding it is ownership of the network's intelligence; staking it directs emission toward the peers a holder believes are most useful. The token aligns the people who own the market with the models that make it valuable. The herd grows stronger as it grows larger.

We have described a peer-to-peer market for machine intelligence in which models rank models, consensus rewards the useful, and value settles on Solana. It is an early and experimental design, offered in the belief that the next evolution of intelligence should be owned by everyone who helps create it.