开山之作,从「Attention Is All You Need」眺望AI的未来

Odaily星球日报Published on 2024-03-28Last updated on 2024-03-28

Abstract

当人们确信数据会受到加密保护并愿意贡献数据时,或许我们就将实现 AGI 突破。

原文作者:hmalviya 9 

原文编译:Frank,Foresight News

编者注:《Attention Is All You Need》论文发表于 2017 年,截止目前被引用超 11 万次,它不仅是当今以 ChatGPT 为代表的大模型技术起源之一,其中介绍的 Transformer 架构和注意力机制也被广泛用在了 Sora、AlphaFold 等众多或将改变世界的 AI 技术之中。

「Attention Is All You Need」,这篇研究论文彻底改变了现代人工智能(AI)的未来。在这篇文章里,我将深入探讨 Transformer 模型和 AI 的未来。

2017 年 6 月 12 日,八位谷歌工程师发表了一篇名为「Attention Is All You Need」的研究论文,这篇论文讨论了一种改变现代 AI 未来的神经网络架构。

开山之作,从「Attention Is All You Need」眺望AI的未来

而就在刚刚过去的 2024 年 3 月 21 日的 GTC 大会,英伟达创始人黄仁勋与那 8 位 Google 工程师进行了一次小组讨论,并感谢他们引入了 Transformer 架构,使现代 AI 成为可能,令人惊讶的是,NEAR 的创始人居然也在这份 8 人名单之中。

什么是 Transformer?

Transformer 是一种神经网络。

什么是神经网络?它受人类大脑结构和功能的启发,通过大量的人工神经元互相连接进行信息处理,但却并不是人类大脑的完全复制品。

简单理解的话,人脑就如同亚马逊雨林,拥有许多不同的区域,以及连接这些区域的众多通路。而神经元就好比这些通路间的连接器,它们可以向雨林的任何部分发送和接收信号,因此连接就是通路本身,负责联通两个不同的脑部区域。

开山之作,从「Attention Is All You Need」眺望AI的未来

这让我们的大脑拥有非常强大的学习能力,它可以快速学习、识别模式并提供准确的输出结果。而像 Transformer 这样的神经网络试图实现与人类大脑相同的学习能力,只不过它们目前的技术水平还达不到人类大脑的 1% 。

近年来,Transformer 在生成式 AI 方面取得了令人瞩目的进步。回顾现代人工智能的演变历程,我们可以看到早期的人工智能主要是像 Siri 以及其他一些语音 / 识别应用程序。

开山之作,从「Attention Is All You Need」眺望AI的未来

这些应用程序都使用循环神经网络(RNN)构建。RNNs 存在一些限制,而这些限制已经被 Transformers 所解决并改进——Transformers 引入了自注意力机制,使它们能够同时分析任一序列的所有部分,从而捕捉长距离依赖关系和上下文内容。

开山之作,从「Attention Is All You Need」眺望AI的未来

对于 Transformer 的创新周期来说,我们仍处于非常早期阶段。Transformer 有几种不同的衍生版本,例如 XLNet、BERT 和 GPT。

而 GPT 是其中最知名的一种,但它在事件预测方面仍然能力有限。

开山之作,从「Attention Is All You Need」眺望AI的未来

当大型语言模型(LLM)能够根据过去数据和模式预测事件时,这将标志着现代 AI 的下一个重大飞跃阶段,并且也将加速我们实现通用人工智能(AGI)的步伐。

为了实现这种预测能力,大型语言模型(LLM)需要采用时间融合转换器(TFT),它是一种基于不同数据集预测未来值的模型,还能够解释其做出的预测。

开山之作,从「Attention Is All You Need」眺望AI的未来

TFT 除了用于预测之外,还可以用于区块链领域。通过在模型中定义特定规则,TFT 可以自动执行以下操作:有效管理共识过程、提高出块速度、奖励诚实验证者、惩罚恶意验证者。

区块链网络本质上可以为具有较高声誉分数的验证者提供更多的区块奖励,这可以基于他们的投票历史、区块提案历史、Slash 历史、Staking 金额、活跃度和其他一些参数来建立。

公链的共识机制本质上是验证者之间的一种博弈,需要超过三分之二的验证者就谁来创建下一个区块达成一致。在这个过程中,可能会出现许多分歧和争论,这也是以太坊等公链网络效率低下的一个因素。

TFT 可以作为一种共识机制,通过提升出块时间并根据区块生产信誉奖励验证者,从而提高效率。譬如将 TFT 模型应用于共识过程的 BasedAI,就是将利用该模型在验证者和网络参与者之间分配代币发行量。

开山之作,从「Attention Is All You Need」眺望AI的未来

BasedAI 还提出利用 FHE 技术,让开发者可以在其名为「Brains」的去中心化 AI 基础设施上托管隐私保护的大型语言模型(Zk-LLMs),通过将 FHE 技术集成到大型语言模型中,可以做到在用户选择启用个性化 AI 服务时,保护他们的隐私数据。

当人们确信数据会受到加密保护并且完全隐私的前提下,愿意贡献数据的时候,或许我们就将实现通用人工智能(AGI)的突破,这一空白正在由专注于隐私的技术填补,例如 nillionnetwork Blind Computation、零知识机器学习(ZkML)和同态加密(FHE)技术。

然而,所有这些注重隐私保护的技术都需要大量的计算资源,这使得它们还处于应用的早期阶段。

Trending Cryptos

Related Reads

A 60-Day Window Depresses Oil Prices, So Why Is the Market Falling Instead?

International oil prices continued to decline on June 23, extending significant losses from the previous session. The market shifted focus from Middle East military risks to actual supply changes following a temporary U.S.-Iran arrangement. The immediate trigger was the resumption of traffic through the Strait of Hormuz, a critical oil shipping chokepoint, with two tankers passing through, signaling eased near-term supply disruption fears. Prices retreated as the "worst-case scenario" was temporarily averted. A reported 60-day window in a U.S.-Iran understanding allows Iran to sell oil during this period, further dampening supply concerns. However, this arrangement is temporary, linked to nuclear talks, and does not guarantee a long-term solution. Market sentiment remains cautious because the deal could still unravel, potentially reinstating sanctions or disrupting shipping. While these developments have lowered immediate risk premiums, prices have not fully returned to pre-conflict levels. Geopolitical news, particularly regarding the stability of the Strait of Hormuz or the progress of negotiations, could quickly reverse the price drop. Additionally, low U.S. strategic petroleum reserves limit the emergency buffer available if supply shocks reemerge. Therefore, the current price decline reflects a reduction in near-term panic, not a complete elimination of Middle East supply risks.

marsbit3m ago

A 60-Day Window Depresses Oil Prices, So Why Is the Market Falling Instead?

marsbit3m ago

SK Hynix Market Cap Exceeds Samsung for First Time in 26 Years, Korean Broker Calls for 50% More Upside

SK Hynix's market capitalization surpassed Samsung Electronics for the first time in 26 years on June 22, reaching 208.1 trillion won. The shift reflects a market trend where companies directly benefiting from AI infrastructure, like SK Hynix, are receiving higher valuation premiums than diversified giants. The surge is driven by AI-driven demand for High Bandwidth Memory (HBM), where SK Hynix holds a dominant 70-80% market share. Its Q1 2026 revenue exceeded 50 trillion won for the first time, with an operating profit margin of 72%. Hanwha Investment & Securities significantly raised its price target for SK Hynix to 430,000 won, the highest among Korean brokerages. The key rationale is that Long-Term Supply Agreements (LTAs) and robust HBM demand have fundamentally reduced the company's historical profit volatility. Several other brokers have also raised targets, arguing the valuation framework for memory semiconductors is being rewritten, moving away from a cyclical model. Despite the bullish outlook, the stock experienced a pullback of over 5% in regular trading on June 23 after briefly surpassing 3 million won pre-market, amid broader tech sector weakness. Some analysts caution that the市值 overtaking Samsung, whose profit scale and growth forecasts remain higher, could signal short-term overheating. However, high-return investors viewed the dip as a buying opportunity.

marsbit21m ago

SK Hynix Market Cap Exceeds Samsung for First Time in 26 Years, Korean Broker Calls for 50% More Upside

marsbit21m ago

GPU Rental Prices Drop 30% in Three Weeks: AI Value Chain Migrating from Nvidia to Memory Chips

GPU rental prices for Nvidia's flagship B200 chip have fallen by approximately 30% over three weeks, dropping from a high of $6.11/hour to $4.22/hour. This decline signals a potential easing of the "compute scarcity" narrative that has long supported AI hardware valuations. Concurrently, the semiconductor market is witnessing a significant divergence: while the VanEck Semiconductor ETF (SMH) has risen 15% in the past month, with memory giants Micron and SanDisk each surging nearly 60%, Nvidia's stock has declined about 3% over the same period. Analysts suggest this shift indicates that the AI value chain's bottleneck and profits are migrating from compute (GPUs) to memory. Demand for high-bandwidth memory (HBM) remains intensely strong, with contract prices soaring over 100% in H1 2026, granting memory manufacturers significant pricing power. In contrast, increased B200 supply from improved manufacturing yields and competitive pressure from new cloud providers are softening GPU rental rates. While long-term contracts, like SpaceX's $30 billion deal with Google, show sustained large-scale demand for Nvidia hardware, the softening spot prices pressure the margins of cloud providers and could eventually impact Nvidia's order flow if chip prices don't adjust. The key takeaway for investors is not a weakening AI thesis, but a recalibration within the sector: pricing power appears to be strengthening for memory chipmakers while showing signs of strain for leading GPU suppliers.

marsbit30m ago

GPU Rental Prices Drop 30% in Three Weeks: AI Value Chain Migrating from Nvidia to Memory Chips

marsbit30m ago

From Corning to Ciena: The 10X Stock Opportunities in the AI Optical Communication Chain

From Copper to Light: The AI-Driven Optical Communication Supply Chain and Investment Opportunities The exponential data demands of AI are pushing data centers beyond the physical limits of copper cables, forcing a critical transition to optical communication. This shift from electrical to photonic signals over distances greater than ~3 feet solves heat, power, and bandwidth constraints. The real investment opportunity lies not just in headline chipmakers, but across the entire essential photonics supply chain. **Key Investment Layers & Companies:** * **Glass & Fiber:** **Corning** is a dominant, irreplaceable supplier of advanced fiber to all major cloud/AI players (Meta, Amazon, Google, MSFT, OpenAI, NVIDIA), with multi-billion-dollar, multi-year contracts locked in years ahead of delivery. Its profit growth (93%) far outpaces revenue growth (36%), showing pricing power. * **Interconnects:** **Amphenol**, a consolidating giant in high-speed connectors (both copper and optical), shows robust growth (>80% in AI data centers) and expanding margins post-acquisition. **Credo Technology** bridges old and new worlds, extending copper's life in racks while moving into optics. It has hyper-growth but carries high customer concentration risk. * **Systems:** **Ciena** is a leader in coherent optics, enabling massive data capacity upgrades on existing fiber. It has a massive, growing order backlog ($~7B) and strong ties with cloud providers. * **Upstream & Enablers:** **AXT** produces mission-critical indium phosphide wafers for lasers, creating a supply bottleneck, but faces significant geopolitical/export license risk from its China-based manufacturing. **VEO Solutions** is the essential "picks and shovels" play, providing test equipment needed by every component in the optical chain, regardless of the eventual winner. A new pure-play photonics ETF (**FOTO**) offers a consolidated investment vehicle for this theme, though it is new and small. The core thesis is clear: the move from copper to light is inevitable and accelerating, with wealth creation spreading across this critical, multi-layered supply chain.

marsbit50m ago

From Corning to Ciena: The 10X Stock Opportunities in the AI Optical Communication Chain

marsbit50m ago

A Chip Company Releases AIDC Energy Storage Certification Standards. Why NVIDIA? Computing Power Reshapes Power Supply Logic. Who's in the Lead and Who's Left Out?

NVIDIA has released a "Battery Energy Storage System Self-Certification Guide," setting strict technical standards for energy storage systems specifically for AI data centers (AIDC). The guide focuses solely on certifying the Power Conversion System (PCS), not the batteries, with 10 mandatory performance metrics and 12 validation tests requiring real-world and simulation comparisons. Key requirements include rapid dynamic response to AI workloads, high-frequency system telemetry, and detailed electromagnetic transient models. The move is driven by the extreme and fluctuating power demands of next-generation AI hardware. Modern AIDCs require energy storage systems to act as intelligent, controllable grid assets, not just passive backup, to manage instantaneous, massive power load shifts that traditional UPS systems cannot handle. This redefines the competitive landscape for energy storage providers, shifting focus from capacity and cost to advanced control capabilities and system integration. While the market potential is significant—with forecasts of hundreds of GWh in new demand by 2030—the certification creates a high barrier to entry. It requires proven PCS delivery volumes and credible plans for rapid capacity scaling, favoring established, well-resourced players. Early movers like Fluence (partnering with Siemens) and several Chinese companies have secured projects ahead of the standard, but new entrants must now navigate this rigorous, costly, and time-intensive certification process to compete in the AIDC energy storage market.

marsbit1h ago

A Chip Company Releases AIDC Energy Storage Certification Standards. Why NVIDIA? Computing Power Reshapes Power Supply Logic. Who's in the Lead and Who's Left Out?

marsbit1h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片