Litepaper
An overview of the Ambient Protocol, Proof of Logits, and our vision for a decentralized, verifiable AI economy. The summary below covers the essentials. Read the full paper for the deep dive.
Read the full litepaperFrom request to proof.
Three steps turn an ordinary API call into a provable one, without the latency tax you'd expect from verification.
Inference request
Your API call is routed to an open GPU node on the Ambient network, compatible with OpenAI and Anthropic SDKs.
Proof of Logits
The node generates a cryptographic proof over the output logits, verifying the exact model produced the exact output.
Verify
Anyone can independently re-check the proof.
import OpenAI from 'openai';
const client = new OpenAI({
baseURL: 'https://api.ambient.xyz/v1',
apiKey: process.env.AMBIENT_KEY
});
Abstract
As artificial intelligence becomes the primary computing interface of the internet, trust in model execution is paramount. Currently, users must blindly trust centralized API providers to run the exact model requested without tampering, quantization, or substitution.
Ambient introduces a decentralized inference network secured by Proof of Logits, a novel cryptographic protocol that provides high-certainty verification of correct execution for large language models.
1. Introduction
The transition from deterministic software to probabilistic AI models breaks traditional verification methods. When an API returns a string of text, the caller has no way to prove which model generated it, or if the weights were modified.
Ambient solves this by decoupling execution from verification through a highly optimized economic and cryptographic protocol.
2. Proof of Logits
Instead of generating a full ZK proof for billions of operations (which currently takes hours for a single LLM forward pass), Ambient utilizes cryptographic commitments over the output logits layer.
- Commitment: Nodes commit to the logit distribution before sampling.
- Verification: The network independently verifies the executed outputs.
- Penalties: Nodes found manipulating outputs or substituting cheaper models are removed from the network.
3. Network Architecture
The network distributes inference across independent GPU nodes, achieving latencies comparable to centralized providers while maintaining strong security guarantees. Every request is fulfilled by a reliable operator and validated against the protocol.