Meta的Llama 3：大語言模型技術的巨大飛躍

Meta 發布了最新的大型語言模型 (LLM) Llama 3，具有 80 億和 700 億個參數，即將推出的模型預計將擁有 4000 億個參數。 Llama 3 在超過 15 兆個代幣的龐大資料集上進行訓練，在超過 50% 的用例中超越了其前輩和競爭對手，展示了基於文字的建模功能的進步。 Meta 旨在透過雲端供應商和服務讓 Llama 3 廣泛使用，並推出了 Meta AI，這是一個專門的網站，並與其產品集成，為最終用戶提供人工智慧驅動的幫助。

Meta Unveils Llama 3: A Quantum Leap in Large Language Model Technology

Meta 推出 Llama 3：大語言模型技術的量子飛躍

On April 18, 2023, Meta Platforms, Inc. (Meta) announced the release of Llama 3, the latest iteration of its groundbreaking large language model (LLM), marking a significant milestone in the field of artificial intelligence (AI).

2023 年 4 月 18 日，Meta Platforms, Inc. (Meta) 宣布發布 Llama 3，這是其突破性大語言模型 (LLM) 的最新版本，標誌著人工智慧 (AI) 領域的一個重要里程碑。

Llama 3 represents a quantum leap over its predecessors, showcasing substantial advancements in natural language processing, code generation, and overall reasoning capabilities. The LLM boasts an unprecedented scale, boasting 8 billion and 70 billion parameters in its two initial models, with plans for future models featuring an astounding 400 billion parameters.

Llama 3 相對於其前輩來說是一個巨大的飛躍，展示了自然語言處理、程式碼產生和整體推理能力方面的重大進步。 LLM 擁有前所未有的規模，其兩個初始模型擁有 80 億和 700 億個參數，未來模型計劃擁有驚人的 4000 億個參數。

Meta meticulously trained Llama 3 on a colossal "large, high-quality training dataset" comprising over 15 trillion tokens, a seven-fold increase over Llama 2. This expansive dataset, containing four times the amount of code, ensures the model's exceptional comprehension and generation capabilities. Additionally, Llama 3 incorporates sophisticated filtering techniques, including NSFW (Not Safe For Work) filters, to safeguard data quality and mitigate potential biases.

Meta 在龐大的「大型、高品質訓練資料集」上精心訓練了Llama 3，該資料集包含超過15 兆個令牌，比Llama 2 增加了七倍。量，確保了模型卓越的理解力和生成能力。此外，Llama 3 還採用了複雜的過濾技術，包括 NSFW（工作不安全）過濾器，以確保資料品質並減少潛在偏差。

Benchmark tests conducted by Meta indicate that Llama 3 outperforms its predecessors and rivals in over half of 12 representative use cases. These use cases encompass a wide range of tasks, including dialogue generation, question answering, summarization, and code generation.

Meta 進行的基準測試表明，Llama 3 在 12 個代表性用例中超過一半的性能優於其前輩和競爭對手。這些用例涵蓋廣泛的任務，包括對話生成、問題回答、摘要和程式碼產生。

The initial releases of Llama 3 are text-based models, while future releases are expected to be multilingual and multimodal, providing enhanced capabilities in various languages and across different media formats. These upcoming models are anticipated to offer extended context, improved reasoning abilities, and superior coding performance, which Meta deems essential "core LLM capabilities."

Llama 3 的初始版本是基於文字的模型，而未來的版本預計將是多語言和多模式的，提供各種語言和跨不同媒體格式的增強功能。這些即將推出的模型預計將提供擴展的上下文、改進的推理能力和卓越的編碼性能，Meta 認為這是必不可少的「核心 LLM 功能」。

Meta intends to make Llama 3 universally accessible by deploying it on all major cloud providers, model API providers, and other services, with the objective of making the product ubiquitous. This strategic move aims to democratize AI technology and empower developers and researchers to leverage its vast capabilities.

Meta 打算透過將 Llama 3 部署在所有主要雲端供應商、模型 API 提供者和其他服務上，使 Llama 3 變得普遍可用，目標是讓產品無所不在。這項策略性舉措旨在使人工智慧技術民主化，並使開發人員和研究人員能夠利用其強大的功能。

Complementing the release of Llama 3, Meta is actively expanding access to AI services for end users across the United States and over 12 other countries. The company has launched a dedicated website, Meta AI, which provides users with a comprehensive suite of AI-powered tools, including writing assistance, trivia games, simulated job interviews, and homework assistance.

作為 Llama 3 發布的補充，Meta 正在積極擴大美國和其他超過 12 個國家的最終用戶對 AI 服務的存取。該公司推出了專門的網站Meta AI，為用戶提供一整套人工智慧驅動的工具，包括寫作幫助、問答遊戲、模擬工作面試和作業幫助。

Meta has also seamlessly integrated Meta AI with its existing products, such as Facebook, Instagram, WhatsApp, and Messenger, providing users with AI-enhanced experiences within their preferred social media and messaging platforms. Furthermore, Meta has integrated the service into its Ray-Ban Meta smart glasses and plans to extend its availability to the Meta Quest VR headset, offering users immersive and hands-free AI interactions.

Meta 還將 Meta AI 與其現有產品（例如 Facebook、Instagram、WhatsApp 和 Messenger）無縫集成，為用戶在其首選的社交媒體和訊息平台中提供人工智慧增強的體驗。此外，Meta 已將該服務整合到其雷朋 Meta 智慧眼鏡中，並計劃將該服務擴展到 Meta Quest VR 耳機，為用戶提供身臨其境的免持人工智慧互動。

"We envision Meta AI as the most intelligent AI assistant that is freely accessible to all users," remarked Mark Zuckerberg, Meta's CEO, in a recent announcement. "Our goal is to empower individuals to harness the transformative power of AI and unlock its potential to enhance their everyday lives."

Meta 執行長 Mark Zuckerberg 在最近的聲明中表示：“我們將 Meta AI 設想為最聰明的人工智慧助手，所有用戶都可以免費使用。” “我們的目標是讓個人能夠利用人工智慧的變革力量，釋放其潛力，改善他們的日常生活。”

The unveiling of Llama 3 and the expansion of Meta AI services come amidst an intensifying race among major tech companies to develop and deploy consumer-oriented AI services. Notably, ChatGPT recently upgraded to GPT-4 Turbo on April 11, while Microsoft Copilot incorporated GPT-4 Turbo in March, signaling a surge in the development and refinement of user-friendly AI solutions.

Llama 3 的推出和 Meta AI 服務的擴展正值各大科技公司開發和部署面向消費者的 AI 服務的激烈競爭之際。值得注意的是，ChatGPT 最近於 4 月 11 日昇級到 GPT-4 Turbo，而 Microsoft Copilot 於 3 月納入了 GPT-4 Turbo，這標誌著用戶友好的人工智慧解決方案的開發和完善激增。