There is a fundamental problem in current iterations of AI, particularly LLMs: they are technologically impressive, but basically represent an expensive consensus mechanism. This is what in Allora we’ve come to call the Cassandra Problem. And we have a solution.
Cassandra was a Trojan priestess and prophetess, who was cursed by Apollo so that her prophesies would not be believed. During the fall of Troy, she warned the Trojans that Greek warriors were hiding inside the Trojan Horse. In response, they ridiculed her.
Modern AI wouldn’t have acted on Cassandra’s warning, either. This is true for most LLMs, which are trained on large volumes of text to select a reply that follows the consensus or conforms to a predetermined selection function (even if the outcome is false or objectionable).
Decentralized or network solutions linking multiple AI models suffer a similar problem. Swarm intelligence must find the best way to combine a multitude of model inferences — we call this process Inference Synthesis.
Solving the Cassandra Problem with Inference Synthesis
The commonly accepted approach is to apply a weighted average, using the cumulative historical reputation to set weights.
But by definition, some models are better than others in certain circumstances. And some models might greatly outperform the others only in specific contexts, while they underperform in the most common situations.
These expert models are like Cassandra: they provide superior intelligence in specific cases, but they are silenced by the canonical approach of using cumulative historical reputation to set model weights.
Allora is a self-improving, decentralized AI network that introduces context-aware machine intelligence, designed specifically to be capable of context-dependent outperformance.
Imagine a set of specialized AI models, each representing the state-of-the-art in a different area of expertise. The Inference Synthesis mechanism of Allora extracts the best from these models, so that the network outperforms any individual model by definition.
With Allora, the Cassandra Problem is no more. After more than 3,000 years, we’ve finally learned to heed Cassandra’s call.
To learn more, read the Allora Litepaper here.
