Data moved off-chain, transactions batched, ClaimReward 80x faster, 20,000 GNK for a bug.
AI prompts and responses are no longer stored in blocks. Only hashes are recorded on the blockchain, while the data itself is stored separately. This removes the block size limit (22 MB) and significantly increases network throughput.
GitHub ↗StartInference and FinishInference are now sent in batches. Fewer transactions → less blockchain load → faster request processing.
GitHub ↗Reward claim request processing has been optimized from 1.6 seconds to ~20 milliseconds. Reservoir sampling and debouncing were used. For a million inferences — the difference between minutes and seconds.
GitHub ↗Proof of Computation parameters have been moved from config to on-chain governance. Now the community can vote to change PoC difficulty without waiting for code updates.
GitHub ↗Participant maksimenkoff discovered a bug in Confirmation PoC: nodes that did not send batches retained confirmation weight. The bug has been fixed, the contributor received 20,000 GNK from the community pool.
GitHub ↗