Next Level of AI: Superintelligence — AI的新紀元： 超智慧
Superintelligence is a term you’ve likely never heard of. You may be hearing about it soon. In the meantime, if you are among us nerds who have a taste for this seemingly pseudoscientific topic, you are probably a fan of Elon Musk’s (in fact, Musk discusses it quite often). So what exactly is this ‘digital superintelligence’?
您可能從未聽說過“超智慧”這個詞語。 但您可能很快就會聽說。如果您是對這個看似偽科學的話題感興趣的怪咖，搞不好您是Elon Musk的粉絲（Musk很常在討論這個話題）。 所謂的“數位超智慧”到底是什麼呢？
Watch: Elon Musk on AI · 馬斯克在討論數位超智慧
馬斯克：無監督的AI發展可謂全人類物種風險。
Originally thought to be a quandary of science fiction, superintelligence is the concept of machines achieving a level of cognition far beyond that of the brightest humans.
Sound far out? It’s not, and here’s why.
數位超智慧最早被認為是科幻小說的概念，它指機器獲得的認知水平遠遠超過最聰明的人。
聽起來太誇張了嗎？ 其實不是，以下是原因。
Musk, most notably founder and CEO of both SpaceX and Tesla, has committed much of his resources to AI’s advancement with OpenAI and Tesla’s world-class AV (self-driving car) technology.
馬斯克 (以SpaceX及Tesla的創始人最為人所知) 已致力他多數的資源於OpenAI和Tesla世界一流的AV技術在AI方面的發展。
Edge Computing · 邊緣運算
With the dawn of the ever-expanding Internet of Things (IoT), there are, as of 2020, an estimated 31 billion connected devices, each serving as data collection points. IoT devices efficiently monitor human patterns and actions, ushering in massive-scale generation of big data. Ultra advanced, computationally-based behavioral analysis and prediction could not be possible without this big data. This is precisely where Edge AI (also known as edge computing) comes in.
隨著 物聯網（IoT）的不斷擴展，截至2020年，估計有 310億 個互聯裝置，每個裝置都充當數據收集點。物聯網裝置可有效地監控人類行為和習慣，從而產生大規模的大數據。 沒有這些大數據，就不可能進行基於計算的超高級行為分析和預測。 這正是Edge AI（也稱為邊緣運算）出現的強項。
Edge AI （邊緣運算）= AI + IoT
Edge computing stems from the need to perform near instantaneous calculations in different locations and situations, where connecting to the cloud would prove impractical and would be accompanied by too much overhead. As an example of edge computing, Amazon Alexa connects to an ever-growing response database that is constantly being trained by millions of different connected devices; each Alexa device also individually, locally trains to adapt to the needs of each owner (skill training). Edge AI requires small-scale calculations on basic machine learning models for quick-reaction systems; the data from these models simultaneously assists the advancement of larger scale models trained on supercomputers. Neural network technology has evolved to a stage in which developing the most advanced AI systems is not so much a matter of having the best AI engineers as it is having access to the most data. With China now leading the way and America closely following in the quantity of aggregate data generated, the resulting zettabytes of IoT data will be the fuel for this AI revolution.
邊緣運算源於需要在不同的地點以及情況下執行近乎即時的計算，而在這種情況下，連接到雲端在長時間下證實是不切實際的，並且伴隨著過多的開銷。作為邊緣運算的一個示例，亞馬遜Alexa連接到一個不斷增長的響應資料庫，該數據庫不斷受到數百萬種不同連接設備的培訓。每個Alexa裝置還分別進行本地培訓以適應每個所有者的需求（數位培訓）。 邊緣AI需要對快速反應系統的基本機器學習模型進行小規模計算；這些模型中的數據同時有助於在超級計算機上訓練的大規模模型的發展。神經網絡技術已經發展到一個階段，在該階段，開發最先進的AI系統並不在於擁有最好的AI工程師，反而是在於蒐集最多的數據。隨著中國現在處於生成的匯總數據方面的領先地位，和美國緊追再後，由此而來的IoT數據將達到Zettabytes（10⁹ TB!)，這將是這場AI革命的動力。
We are now in the advent where manual driving is being entirely replaced by autonomous transportation. Right now, Uber is mapping out Washington D.C. in preparation to implement its first AV’s. Elon Musk estimates that, by the end of 2020, Tesla’s AV’s will be on average far safer than any human driver. As AI acquires skills that were once conceived possible only in science fiction, and as an increasing number of jobs are replaced by machines, AI is becoming ever more replicative of humans. While professing on the ‘apocalyptic’ risk of superintelligence may seem to some as simple fearmongering, others consider this concern to be grounded but premature. A necessary step between our current (and still relatively tame) machine learning models and superintelligence is a computer that has at least human-level intelligence. Some argue not only have machines reached a human level of intelligence, but that in fact this level has been surpassed. Computers have instantaneous data (memory) recall (i.e. perfect recall) to an amount of data that far exceeds any human. Given a clear set of rules, AI can outcompete us in any game we train it for. With exponential advancement in AI technology and without, as of yet, any overseeing regulatory committee, AI’s capabilities may develop at a rate that we can’t control. This concept is known as singularity, and it may pose a far more imposing threat than the reality we currently face from AI: news fabrication and social engineering or the mass technological unemployment predicted by the year 2050.
我們現在處在一個新的AI時代，自動駕駛已要完全取代了手動駕駛。目前，Uber正在規劃將在華盛頓特區準備實施其首個AV實行。馬斯克估計，到2020年底，特斯拉的AV平均比任何人類駕駛員都安全得多。隨著AI獲得過去只能在科幻小說中才能想到的技能，並且隨著越來越多的工作被機器取代，AI越來越具有模仿和取代人類的能力。對某些人而言，聲稱自己具有“多性”的超智慧風險似乎只是簡單的恐懼心理，而另一些人則認為這種關懷是有根據的，就是為時過早。在我們目前（目前相對馴服）的機器學習模型和超級智慧之間，必不可少的步驟是擁有至少人為智慧的計算機。一些人認為，機器不僅達到了人類的智慧水準，反而實際上已經超過了這一水準。計算機具有即時的數據（內存）調用（即完完美記憶力）功能，其數據量遠遠超過任何一個人。只要有一套明確的規則，人工智慧就可以在我們為其訓練的任何遊戲中勝過我們。隨著AI技術的飛速發展，並且直到目前還沒有任何監督性的監管委員會，AI的能力可能會以我們無法控制的速度發展。這個概念被稱為“科技奇異點”，它可能構成比我們目前從AI面臨的現實要強大得多的威脅：新聞製作和社會工程或到2050年所預測的大規模技術性失業。
Authors (作者): Seth Austin Harding, Albert (Al) Tsai · 蔡沛承
This article is written by NTU AI Club, committed to the public understanding and advancement of artificial intelligence.
本文由台大AI社撰寫，致力於公眾對人工智慧的理解和發展。
