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

MarsBitPublicado em 2023-06-16Última atualização em 2023-06-16

Resumo

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

Leituras Relacionadas

a16z Deep Dive: Stop Chasing the 'AI Smell', Here's a Practical Guide to Writing with AI

"Don't Obsess Over AI Detection: A Practical Guide to Writing Alongside AI" by Steph Zinn (a16z Crypto) This guide moves beyond the flawed premise that AI-generated text can be easily spotted by a set of "tells" and that these features automatically mean poor quality. Instead, it focuses on how writers and founders can use LLMs effectively by understanding, controlling, and editing the common stylistic tendencies of AI-assisted prose. The article breaks down AI writing "tells" into four key dimensions: **1. Rhetorical Features (Insight-Shaped Writing):** AI often produces semantically empty, "corporate-sounding" filler language—vague profundities, hedging phrases, excessive parallelism, and summary statements. The advice is to ruthlessly edit these out, using prompts to make language more specific and direct. **2. Voice Features (The Alexa Voice):** Default AI writing relies on a narrow, fungible vocabulary of low-friction, abstract words and cliché phrases that lack personality. While this generic voice is acceptable for support docs or mass communications, founders should preserve their unique voice for impactful writing. Use LLMs to identify and replace jargon, aiming for concrete, distinctive word choices. **3. Structural Features (Form Without Function):** AI tends towards over-structured text with excessive subheadings, lists, roadmaps, and the rigid "three-point" framework. While clear structure is good for readability and SEO, it shouldn't force ideas into unnatural containers. Choose a structure that serves the format and purpose, borrowing from effective examples. **4. Punctuation Features (Dash Panic):** The overuse of em dashes and colons has become a hallmark, but writers shouldn't avoid useful punctuation just to seem "human." The key is avoiding repetitive, distracting patterns. Use punctuation that is grammatically correct and supports the flow of your argument. The core argument is that many so-called AI flaws are just amplified versions of existing bad writing habits. The goal isn't to eliminate AI's role but to use it as a tool while maintaining editorial control. The final question shouldn't be "Can this be detected as AI?" but "Does this writing effectively do its job?"

marsbitHá 17m

a16z Deep Dive: Stop Chasing the 'AI Smell', Here's a Practical Guide to Writing with AI

marsbitHá 17m

Podcast Notes | Conversation with Tom Lee: Bitmine Acquiring Nearly 5% of Total ETH Supply Is Not the End Goal, ETH Price Target Set at $10,000

In a podcast interview, Tom Lee, Chairman of BitMine Immersion Technologies, discusses the company's strategy to accumulate nearly 5% of the total Ethereum supply within 14 months, using equity financing and avoiding debt. BitMine has consistently purchased ETH for over 60 consecutive weeks, with recent weeks combining buybacks with purchases. The company's substantial ETH holdings generate approximately $300 million in annual staking rewards, covering operational costs like the dividends for its 9.5% perpetual preferred stock (BMNP). Lee positions ETH as a store-of-value asset, likening it to stocks or land, rather than a pure cash-flow instrument. Looking ahead, Lee suggests BitMine may continue buying beyond the 5% target if institutional adoption grows. He outlines a bullish price target for ETH: surpassing $5,000 in a new crypto bull cycle and potentially exceeding $10,000 within 1-2 years, driven by Wall Street tokenization and AI-related demand. The discussion also covers BitMine's evolution into an ecosystem player, funding Ethereum Foundation spin-offs and developing its Maven staking platform. Lee acknowledges his significant financial interests are tied to ETH's price and BitMine's performance. The interview provides a framework for evaluating ETH as a long-term asset, emphasizing staking yield sustainability and future institutional demand, while noting the uncertainties surrounding macro cycles and real-world adoption.

marsbitHá 17m

Podcast Notes | Conversation with Tom Lee: Bitmine Acquiring Nearly 5% of Total ETH Supply Is Not the End Goal, ETH Price Target Set at $10,000

marsbitHá 17m

Pricing Risk Assets in 8 Hours: Tonight's PCE to Set the Discount Rate, Nvidia to Test Earnings Tomorrow Morning

"Pricing Risk Assets in 8 Hours: PCE to Set the Discount Rate Tonight, NVIDIA to Test Profits Tomorrow Morning" Risk asset prices hinge on two variables: the numerator (earnings expectations) and the denominator (the discount rate). Both will be recalibrated within eight hours. First, at 20:30 Beijing time, the US Bureau of Economic Analysis releases July PCE inflation data and the second estimate of Q2 GDP. Consensus expects mild core PCE growth, but a surge in key PPI components poses an upside risk. A hotter-than-expected print could push Treasury yields and the dollar higher, threatening the recent rally in Bitcoin (BTC) above $80K, which was fueled by falling yields. A benign reading would support risk assets. Market sentiment is already "greedy" (Fear & Greed Index at 74), making it vulnerable to disappointment. Second, around 04:20, NVIDIA reports its Q2 FY27 earnings. While consensus revenue of ~$91.85B slightly exceeds company guidance, the market has priced in a beat. The key will be the magnitude of the beat and, crucially, the Q3 guidance. As a bellwether for tech and AI narratives, NVIDIA's results will significantly impact overall risk appetite and AI-related crypto tokens. The combination creates four scenarios: 1) Benign PCE & strong NVIDIA guidance confirms the bullish trend. 2) Hot PCE & strong NVIDIA leads to conflicted signals and likely volatility for BTC. 3) Benign PCE & weak NVIDIA guidance pressures tech but offers some macro support for BTC. 4) Hot PCE & weak NVIDIA guidance presents a "double whammy," risking a sharp pullback in BTC toward the $75K-$76K support zone. The outcomes will set the tone for markets ahead of the upcoming Jackson Hole symposium.

marsbitHá 51m

Pricing Risk Assets in 8 Hours: Tonight's PCE to Set the Discount Rate, Nvidia to Test Earnings Tomorrow Morning

marsbitHá 51m

Trading

Spot
活动图片