@OpenGradient
I traced one transaction through the alpha testnet expecting a single execution path. It split into three before it even confirmed.
The transaction landed in what they call the Inference Mempool. Instead of just sitting there waiting its turn, the mempool simulated it, pulled out the inference call buried inside, and sent that off to run in parallel while the transaction itself stayed pending.
That ordering caught me off guard. I had been assuming inference always happens after a transaction lands, the same way an oracle update happens after a price moves. Here it runs before the transaction finishes, not after.
Once the model result came back, the original transaction resumed with that result already baked in, then went into the next block. No separate callback, no waiting on an external oracle round trip. The result is just part of the same atomic operation.
What this actually buys is removing oracle delay for ML-driven logic, plus letting hundreds of pending inferences run side by side instead of queuing behind each other.
I'll say the obvious part though. This is alpha testnet only, not even live on the official testnet yet. Models have to be ONNX format, which rules out a lot of larger architectures. And I haven't seen anything in the docs addressing what happens if the simulated inference and the actual execution environment disagree by the time the transaction resumes.
#OPG $OPG
$SYN $BAS
I traced one transaction through the alpha testnet expecting a single execution path. It split into three before it even confirmed.
The transaction landed in what they call the Inference Mempool. Instead of just sitting there waiting its turn, the mempool simulated it, pulled out the inference call buried inside, and sent that off to run in parallel while the transaction itself stayed pending.
That ordering caught me off guard. I had been assuming inference always happens after a transaction lands, the same way an oracle update happens after a price moves. Here it runs before the transaction finishes, not after.
Once the model result came back, the original transaction resumed with that result already baked in, then went into the next block. No separate callback, no waiting on an external oracle round trip. The result is just part of the same atomic operation.
What this actually buys is removing oracle delay for ML-driven logic, plus letting hundreds of pending inferences run side by side instead of queuing behind each other.
I'll say the obvious part though. This is alpha testnet only, not even live on the official testnet yet. Models have to be ONNX format, which rules out a lot of larger architectures. And I haven't seen anything in the docs addressing what happens if the simulated inference and the actual execution environment disagree by the time the transaction resumes.
#OPG $OPG
$SYN $BAS