长推:对红杉资本《The New Language Model Stack》文章的8点分析

MarsBitОпубликовано 2023-06-16Обновлено 2023-06-16

Введение

干货非常多。很多内容对AI创业者应该会有不少帮助

注:本文来自@FinanceYF5 推特,其是@33daoweb3 成员,Gofans顾问,原推文内容由MarsBit整理如下:
红杉资本6月14日发布最新文章《The New Language Model Stack》
访谈了投资Portfolio里的33家小到种子轮,大到已经上市的公司后总结出来的。
全文总共有8点分析,干货非常多。很多内容对AI创业者应该会有不少帮助
Thread:

AI人工智能


1.几乎所有公司都在产品中用到语言模型
几乎每一家公司都在将语言模型构建到他们的产品中
- 从代码和数据科学副驾驶,到为客户、开发者、员工和纯娱乐提供的聊天机器人。
许多公司正在用AI优先的视角重新设想整个工作流程。
这些只是几个例子,而且只是个开始。
2.这些应用主要基于API、检索和编排,但开源使用也在增长
-65%的公司已经将应用投入生产
-94%的公司正在使用基础模型API
-88%的公司认为检索技术
-38%的公司对像LangChain的框架感兴趣
-15%的公司从头开始或使用开源资源构建了自定义的语言模型

AI人工智能


与每一位实践者交谈的结果显示,AI的发展速度太快,以至于很难对最终的技术栈有高度的信心,
但大家一致认为,LLM API将继续作为一个关键支柱,
其次是检索机制和像LangChain这样的开发框架。
开源和自定义模型的训练和调整似乎也在增长。
语言模型技术栈的其他领域也很重要,但成熟度较低。
3.公司希望根据自己的上下文来定制语言模型
目前,有三种主要的方式来定制语言模型:
从头开始训练自定义模型,难度最高。
微调基础模型,难度中等。
使用预训练模型并检索相关上下文,难度最低

AI人工智能


4.如今API调用和训练模型是相互独立的,但未来两者将慢慢融合在一起
我们可能会觉得我们面临着两种技术栈的选择:
一种是利用LLM API的技术栈
另一种是训练自定义语言模型的技术栈
越来越多的公司对训练和微调自己的模型产生了兴趣。随着时间的推移,LLM API技术栈和自定义模型技术栈会越来越融合
5.技术栈正在变得越来越易于开发者使用
LangChain通过抽象化常见问题,帮助开发者构建LLM应用:
-将模型组合成更高级的系统,
-将多个模型调用链接在一起,
-将模型连接到工具和数据源,
-构建能够操作这些工具的代理,
-以及通过简化语言模型切换的过程,帮助避免对供应商的依赖。
6.语言模型需要变得更可靠(输出质量、数据隐私和安全性)
许多公司希望有更好的工具来处理数据隐私、隔离、安全、版权和监控模型输出。
来自金融科技到医疗保健的受监管行业的公司特别关注这一点
随着政策的明确和更多的安全措施到位,语言模型将得到更好的信任
7.语言模型应用将变得越来越多模态
8.目前还仅仅只是初期阶段
AI刚刚开始渗透到技术的每一个角落,只有65%的受访者今天已经开始尝试,而且其中很多都是相对简单的应用。
基础设施层(我理解就是中间层的意思)将在未来几年内快速发展
原文:
https://twitter.com/michelle_fradin/status/1669117628521271298

Похожее

A Nation Blocks Chips, a Giant Buys a Nuclear Power Plant: Why It's Time to Seriously Consider DeAI

**Title: Great Powers Blockade Chips, Giants Buy Nuclear Plants: Why It's Time to Seriously Consider DeAI** In May 2026, the US closed loopholes for Chinese firms to acquire advanced NVIDIA chips via overseas subsidiaries. That same month, Kenya halted a $1B geothermal data center project involving Microsoft, fearing its immense energy consumption. Meanwhile, Huawei announced mass production of its Ascend AI chip. These disparate events underscore a new reality: the competition for computing power ("compute") has escalated beyond the tech industry, becoming a geopolitical and infrastructural battleground. A new era of oligopoly is forming, with control over the AI stack—from GPU chips (NVIDIA) and cloud platforms (AWS, Azure, Google Cloud) to foundational models (OpenAI, Anthropic)—concentrating in a few Western "AI Octopus" corporations. This centralization creates systemic risks: pricing power and platform lock-in for users, infrastructure fragility, and a widening "compute divide" that threatens to marginalize nations without independent AI capacity. An "AI Iron Curtain" is deepening through export controls. In response, some nations like Saudi Arabia and the UAE are investing heavily to buy compute power, aiming to transition from oil to AI economies. The EU seeks to triple its compute capacity by 2030 to reduce dependency. However, the spending gap is vast, with four US tech giants alone planning ~$750B in AI capex for 2026. The race is increasingly constrained by energy, with AI tasks consuming up to 1000x more power than web searches, pushing firms to even acquire nuclear plants. This landscape is fueling interest in Decentralized AI (DeAI). It proposes a third way: using open protocols to coordinate a global network of idle GPUs, independent developers, and data centers, creating an AI infrastructure without a single controlling entity. Leveraging blockchain and cryptographic verification, DeAI aims to break market concentration, disperse energy demands, reduce geopolitical dependencies, and enhance transparency. While still nascent in performance and stability, DeAI's core promise is not immediate superiority but providing a crucial alternative architecture to resist monopoly, censorship, and centralized power. As specialized AI hardware costs fall and open-source models flourish, the window to build this foundation is open. The very existence of such competition serves as a vital check against the inevitable abuse of concentrated power.

marsbit23 мин. назад

A Nation Blocks Chips, a Giant Buys a Nuclear Power Plant: Why It's Time to Seriously Consider DeAI

marsbit23 мин. назад

Outpoll Review: A Prediction Market Platform Built for Active Traders

Outpoll Review: A Prediction Market Platform Built for Active Traders In recent years, prediction markets have grown from a niche sector to a mainstream arena, attracting billions in trading volume and institutional capital. However, the user experience and tools for traders have not kept pace. Outpoll, a new global prediction market platform, aims to fill this gap by providing enhanced trading infrastructure for active and professional traders. Built on standard prediction market principles, Outpoll allows users to trade on the outcome of specific events. It uses fully collateralized contracts with USDC settlement, charges a competitive 0.1% fee per trade, and provides clear settlement rules upfront to minimize disputes. A key focus for Outpoll is its professional-grade trading tools. The platform supports limit and market orders, as well as take-profit and stop-loss orders for open positions—features uncommon in prediction markets. For automated trading, Outpoll offers comprehensive REST and WebSocket APIs, enabling portfolio management, price arbitrage, and integration with existing tools. The platform also features a creator-led market model, where approved experts and community leaders can create and manage markets for niche topics under platform supervision. Its integrated interface combines news feeds directly with trading functions, allowing users to monitor events and manage positions seamlessly. Outpoll launched with a native Android app (available on Google Play) and plans an iOS version later this year. In summary, Outpoll distinguishes itself with trader-focused tools, practical APIs, transparent and collateralized markets, integrated news, and an expanding creator program. For active traders, its advanced order types and API access alone make it a platform worth watching. Outpoll is now globally accessible via outpoll.com and Google Play.

marsbit31 мин. назад

Outpoll Review: A Prediction Market Platform Built for Active Traders

marsbit31 мин. назад

Bitwise: Crypto Becomes a Contrarian Investment, Three Logics to Understand the Current Market

**Summary** Matt Hougan, Bitwise's CIO, analyzes the current crypto market through three key lenses, arguing it has shifted from a momentum-driven to a contrarian investment. **1) Crypto Becomes a Contrarian Play:** The market is weak, with major assets like Bitcoin and Ethereum down significantly. Capital has moved to hot sectors like AI, leaving crypto as an "unloved" asset class. This transforms crypto investing from trend-following to a test of patience and fundamental analysis. Investors now favor projects with solid fundamentals (e.g., Hyperliquid) over speculative ones. **2) Regulatory Overhang:** The uncertain fate of the U.S. CLARITY Act, a major crypto regulatory framework, is a key headwind. With its passage in 2024 seen as far from guaranteed (estimates range from 30-55%), institutional capital remains on the sidelines, choosing less risky alternatives like AI stocks. The market needs clarity—whether the bill passes or fails—more than any specific outcome to move decisively. **3) Capital Rotates to New Fundamentals:** This cycle differs from past bear markets where money fled to Bitcoin. Now, capital seeks smaller assets with strong use cases. While major cryptos fell in May 2024, tokens like Hyperliquid (+72%), Zcash (+50%), and XLM (+44%) rallied on their specific fundamentals. This rotation confirms the new contrarian, fundamentals-driven logic and signals the bear market may be in its later stages. **Conclusion:** Short-term pressure persists due to regulatory uncertainty and competition from AI narratives. Investing in crypto now requires a contrarian mindset—acting against the crowd and focusing on fundamental value. Patience and targeting high-quality projects based on their merits are essential for capturing long-term gains.

marsbit1 ч. назад

Bitwise: Crypto Becomes a Contrarian Investment, Three Logics to Understand the Current Market

marsbit1 ч. назад

ChatGPT Might Be Disappearing Soon

OpenAI announced at its "Intelligence at Work" event that its coding assistant, Codex, will be fully integrated into the ChatGPT app within weeks. This move marks a strategic shift from a conversational AI (Chat) towards a unified "agentic" platform capable of execution. Codex, originally launched to compete with Anthropic's Claude Code, has grown rapidly to 5 million weekly active users, with 20% being non-developers like analysts and designers. Its enterprise revenue now constitutes 40% of OpenAI's total. The integration is the first step in creating a super-app combining ChatGPT (interface), Codex (execution engine), and the Atlas browser (web access). OpenAI also unveiled new Codex features: specialized Agent plugins for six professional roles, an "Annotations" tool for direct document editing, and a "Sites" function to turn work into shareable web apps. Internally, this reflects a power shift; the Codex team now leads core product strategy. While the ChatGPT brand remains for its vast user base, the platform's future is focused on autonomous agents that perform tasks, not just chat. The article notes that competition with Claude Code pushed OpenAI's development, with Codex competing on cost-effectiveness and accessibility rather than raw coding quality. It concludes that the essence of "ChatGPT" is evolving from a chatbot into an AI agent platform, with the name potentially becoming a legacy symbol of its original function.

marsbit1 ч. назад

ChatGPT Might Be Disappearing Soon

marsbit1 ч. назад

Торговля

Спот
Фьючерсы
活动图片