激勵網絡：人工智慧與現實世界之間缺失的聯繫

去中心化網路讓一切變得更加複雜，但去中心化網路可以處理複雜的事情。當談到解決人工智慧（AI）對運算能力的貪婪需求時，問題可能足夠複雜，以至於需要去中心化。

Decentralized networks complicate everything, but they can also handle complicated things. When it comes to artificial intelligence’s (AI) insatiable demand for computing power, the problem might just be knotty enough for decentralization.

去中心化網路使一切變得複雜，但它們也可以處理複雜的事情。當談到人工智慧（AI）對運算能力的永不滿足的需求時，問題可能已經足夠棘手，以至於需要去中心化。

Incentive networks are a type of decentralized network that reward individual behavior that benefits the network as a whole, fostering an “ecosystem” mentality. What distinguishes a simple ecosystem from an incentive network is its intentionality and mechanisms. An ecosystem is often a fortunate accident, the sum of competing forces deciding they are better off working within certain parameters than outside the group. An incentive network is designed for shared success from the get-go.

激勵網絡是一種去中心化網絡，它獎勵有利於整個網絡的個人行為，培養「生態系統」心態。簡單生態系統與激勵網絡的差異在於其意圖與機制。生態系統往往是一個幸運的意外，競爭力量的總和決定了它們在某些參數內工作比在群體之外工作更好。激勵網絡從一開始就旨在實現共同成功。

But how does this relate to AI? Think of scalable AI applications as mechanical entities that churn out simple answers from ludicrously large data sets using computational power, like gas in a car. The more data you haul and the faster you want the answers, the more gas you burn, and the biggest and most complex AI models burn several times what smaller ones do: OpenAI’s GPT-4 cost $78 million in compute to train, while Google’s Gemini Ultra cost $191 million. With numbers that big, a system that can reduce hardware investments and dynamically allocate resources to reduce overall costs while remaining impartial to the participants is crucial — and that’s what incentive networks do.

但這與人工智慧有什麼關係呢？將可擴展的人工智慧應用視為機械實體，它們使用運算能力從極其龐大的數據集中產生簡單的答案，就像汽車中的汽油一樣。您傳輸的資料越多，想要得到答案的速度越快，消耗的Gas 就越多，而最大、最複雜的AI 模型的消耗量是較小模型的幾倍：OpenAI 的GPT-4 的訓練計算成本為7800 萬美元，而Google 的Gemini Ultra耗資1.91億美元。由於數量如此之大，一個能夠減少硬體投資、動態分配資源以降低整體成本、同時對參與者保持公正的系統至關重要——這就是激勵網路的作用。

Continue Reading on Coin Telegraph

繼續閱讀硬幣電報