Sequoia's AI Take
(中文在下）
The recent AI Ascent Conference video from Sequoia highlighted the significant potential for future AI applications. The presentation clearly outlined AI's current state, developments for this year, and future trends.
Current Landscape
Vast Opportunity Gap in AI Application: Compared to the saturated cloud and mobile markets, AI applications remain an untapped treasure trove: AI application companies generating over $1B in revenue are still limited to just ChatGPT (excluding additional revenue from big companies due to AI product launches), which is far less compared to the cloud and mobile markets where application-end revenue exceeds $1B.
AI Application haven’t reached strong PMF!? The numbers show that people are interested in AI, but they're not sticking around as much as they could. This means there's loads of room to make AI stuff more appealing and useful, so users really get hooked.
Reflecting on the iPhone's Journey: Looking back at the iPhone's development, it also took two to three years before applications exploded: The first iPhone was launched in 2007, the AppStore appeared in 2008, but it wasn't until after 2010 that phenomenon-level applications like Instagram, WhatsApp, and Instacart emerged.
2024 Prediction
From Co-pilot to Agent: By 2024, AI will transition from a supportive role to an independent action agent, eliminating the need for human intervention and functioning more like a colleague. This change is already becoming apparent in fields like coding and customer service.
Stronger Reasoning and Planning: One of the primary critiques of large language models is their reliance on mimicking statistical patterns rather than truly reasoning through problems. Recent advances in research, including real-time data access and game-theoretic approaches (make agent to debate to each other), are addressing this by allowing models to 'think' before acting.
More Reliability: Shift from casual consumer apps with tolerable errors to critical enterprise applications in healthcare and defense where accuracy is crucial. These methods range from RLHF, prompt chaining and the use of vector databases.
From Prototyping to Production: In 2024, expect a surge of AI prototypes advancing to production, highlighting key concerns like latency, cost, and ownership of models and data. This will drive a shift in computational focus towards inference rather than pre-training.
What’s next?
Radical Cost Reduction: AI's potential to dramatically reduce costs across various sectors could lead to unprecedented levels of operational efficiency. This reduction is particularly impactful in areas like education and healthcare, where costs have historically risen faster than inflation. AI's ability to drive down these costs represents a significant societal benefit.
AI-Infused Companies: The concept of AI-infused companies suggests a future where businesses operate similarly to neural networks, with interconnected processes enhancing efficiency and decision-making. This framework allows even individuals to manage complex operations effectively, leveraging AI for comprehensive support across various tasks and challenges.
之前提到的 Sequoia 最近公開的 AI Ascent Conference 影片中，有談及 AI 應用未來的巨大潛力。看完完整影片後，我發現這個簡報非常清楚地描述了 AI 的現狀、今年的發展展望以及未來的趨勢。以下是一些重點摘要，完整內容可回去看 YouTube 影片。
現況分析：
相比於已飽和的雲端和移動市場，AI 應用仍舊是一片待挖掘的寶地：營收超過$1B的AI應用公司仍只有ChatGPT一間（忽略大公司因為推出AI產品提高ARPU帶來的額外營收)，相較雲端和移動市場，應用端營收超過$1B公司數量是遠超基礎設施的。
大部分的 AI 應用還沒有達到 PMF：雖然大家對 AI 非常感興趣，但真正能留住用戶的產品還不夠多。DAU/MAU和留存率仍然落後於最好的雲端/移動應用程式，顯示真正到達PMF還有一段路要走。
回顧 iPhone 的發展歷程也是經歷兩三年後才迎來應用的爆發: iPhone 2007 年推出第一代，2008年出現 AppStore，但也直到 2010年之後才出現像 Instagram、WhatsApp、Instacart等現象級的應用。
2024 年展望：
從 Copilot 到 Agent：到 2024 年，AI 將從協助角色轉變為獨立的行動代理人，無需人類介入，更像是你的同事。這種轉變已在軟體工程和客服等領域浮現。
強化模型推理與規劃能力：一直以來對大型語言模型的批評在於它們依靠模仿統計模式，而非真正解決問題。近期的研究進展，例如引入即時數據訪問和博弈論策略（讓機器自我辯證），正助力模型在行動前進行深思。
提升模型推理可靠性：從錯誤尚可接受的一般消費者應用，過渡到在醫療和國防等需要模型推論結果精確度極高的企業應用。這些領域正在運用諸如基於人類反饋的強化學習 (RLHF)、提示串連 (Prompt Chaining) 和向量數據庫 (Vector Database) 等技術來提高推論精準度。
AI應用規模化：預計在 2024 年，將有更多 AI 應用原型成功轉入規模化階段，公司門將不只在意模型表現與幻覺問題，他們將更看重延遲、成本以及模型與數據所有權等關鍵問題，且硬體資源關注的重點將從 pre-training 轉向 inference。
長期展望：
大幅降低成本： AI 在降低各行各業成本上的潛力，有望帶來前所未有的運營效率。在我們的社會中，一些重要的事物，如教育和醫療保健的價格，上漲速度超過通脹率，而人工智能有望幫助降低這些關鍵領域的成本。
AI Infused Company：未來的公司可能會像神經網絡一樣運作，過程相互連接，提高效率和決策能力。這樣的架構使個人也能有效地管理複雜的業務，借助 AI 在面對各種任務和挑戰時提供全面的支援。
