2024 关于AI的一张图和24个问题

Full Steam Ahead: The 2024 MAD (Machine Learning, AI & Data) Landscape

这是我们第十届年度报告，对数据、分析、机器学习及人工智能生态系统进行“国情咨文”式的概述。在过去十多年的观察中，这个领域的发展从未像今天这样令人激动和充满希望。多年来，我们描述的所有趋势和子趋势正汇集在一起：数据被大规模数字化；利用现代工具，可以快速且经济地存储、处理和分析数据；最关键的是，数据能够被越来越高性能的机器学习/人工智能模型所使用，这些模型能够理解数据、识别模式、基于数据做出预测，并现在还能生成文本、代码、图像、声音和视频。

MAD（机器学习、人工智能与数据）生态系统已从一个小众技术领域转变为主流。似乎这一范式转变正在加速，其影响远远超出技术或商业领域，波及到社会、地缘政治乃至人类的生存状态。对一些人来说，这变化可能是突然的，但现在它已经变成了“无所不在，一切皆可”。尽管这个数十年的大趋势还有许多章节有待书写，我们每年的这篇帖子旨在尝试理解我们当前所处的状态，包括产品、公司和行业趋势的概述。

这里是过去的版本：2012、2014、2016、2017、2018、2019（第一部分和第二部分）、2020、2021和2023（第一部分、第二部分、第三部分、第四部分）。

今年的团队由Aman Kabeer和Katie Mills（FirstMark）、Jonathan Grana（Go Fractional）以及Paolo Campos组成，对他们所有人表示衷心的感谢。也非常感谢CB Insights提供的交互版本中出现的卡片数据。

这篇年度“国情咨文”分为三个部分：

• 第一部分：生态系统概览（PDF，交互版本）

• 第二部分：我们在2024年关注的24个主题

• 第三部分：融资、并购与IPO情况

This is our tenth annual landscape and “state of the union” of the data, analytics, machine learning and AI ecosystem.

In 10+ years covering the space, things have never been as exciting and promising as they are today. All trends and subtrends we described over the years are coalescing: data has been digitized, in massive amounts; it can be stored, processed and analyzed fast and cheaply with modern tools; and most importantly, it can be fed to ever-more performing ML/AI models which can make sense of it, recognize patterns, make predictions based on it, and now generate text, code, images, sounds and videos.

The MAD (ML, AI & Data) ecosystem has gone from niche and technical, to mainstream. The paradigm shift seems to be accelerating with implications that go far beyond technical or even business matters, and impact society, geopolitics and perhaps the human condition. Perhaps suddenly for some, it has become “everything, everywhere all at once”.

There are still many chapters to write in the multi-decade megatrend, however. As every year, this post is an attempt at making sense of where we are currently, across products, companies and industry trends.

Here are the prior versions: 2012, 2014, 2016, 2017, 2018, 2019 (Part I and Part II), 2020, 2021 and 2023 (Part I, Part II, Part III, Part IV).

Our team this year was Aman Kabeer and Katie Mills (FirstMark), Jonathan Grana (Go Fractional) and Paolo Campos, major thanks to all. And a big thank you as well to CB Insights for providing the card data appearing in the interactive version.

This annual state of the union post is organized in three parts:

• Part: I: The landscape (PDF, Interactive version)

• Part II: 24 themes we’re thinking about in 2024

• Part III: Financings, M&A and IPOs

PART I: THE LANDSCAPE

第一部分：概览

公司数量

2024年MAD生态概览总共展示了2011个标志。

这个数字比去年的1416有所增加，地图上新增了578个入口。

作为参考，2012年的第一个版本仅有139个标志。

这个生态的极度（疯狂地？）拥挤主要是由于两波接连不断的大规模公司创立和融资浪潮所致。

第一波是持续了大约10年的数据基础设施周期，始于大数据，终于现代数据栈。长期期待的该领域整合尚未真正发生，绝大多数公司仍然存在。

第二波是ML/AI周期，真正开始于生成式人工智能。我们正处于这一周期的初期，大多数公司非常年轻，我们在生态中包括了年轻的初创公司（其中相当一部分仍处于种子阶段）。

注意：这两波浪潮是紧密相关的。MAD生态概览每年的一个核心思想就是展示数据基础设施（在左侧）、分析/商业智能和ML/AI（在中间）以及应用（在右侧）之间的共生关系。

虽然每年将日益增加的公司数量适配进生态图变得越来越难，但最终，思考MAD领域最好的方式是作为一个流水线——数据从收集到存储再到处理，最后通过分析或应用创造价值的完整生命周期。

两大浪潮 + 有限的整合 = 生态上的大量公司。

“基础设施”和“分析”中的主要变化

我们对生态左侧的整体结构几乎没有做出改变——正如下文所述（现代数据栈死了吗？），这部分MAD生态最近的热度相对较低。

一些值得注意的变化：我们将“数据库抽象”更名为“多模型数据库与抽象”，以捕捉围绕一体化“多模型”数据库组（如SurrealDB*、EdgeDB）的兴起浪潮；去年我们实验性创建的“加密/网页3分析”部分感觉在这个生态中格格不入，因此被取消；我们移除了“查询引擎”部分，感觉它更像是部分中的一部分而不是一个单独的部分（该部分的所有公司仍然出现在生态图中——如Dremio、Starburst、PrestoDB等）。

“机器学习与人工智能”中的主要变化

随着2023年AI公司的爆炸性增长，这里是我们进行最多结构性变化的地方。

鉴于去年“AI赋能”层面的巨大活动，我们在MLOps旁边增加了3个新类别：

“AI可观察性”是今年的一个新类别，包括帮助测试、评估和监控LLM应用的初创公司

“AI开发平台”在概念上接近MLOps，但我们想要认可完全专注于AI应用开发的平台浪潮，特别是围绕LLM的训练、部署和推理

“AI安全与保障”包括解决LLM固有问题的公司，从幻觉到伦理、法规合规等

如果Sam Altman与Elon Musk之间非常公开的争执告诉了我们什么，那就是在基础模型开发者中，商业与非营利的区别是一个关键点。因此，我们将之前的“横向AI/AGI”分成了两个类别：“商业AI研究”和“非营利AI研究”

我们进行的最后一个变化是另一个命名变化，我们将“GPU云”修改为以反映许多GPU云提供商添加的核心基础设施功能集：“GPU云/ML基础设施”

“应用”中的主要变化

这里最大的更新是……对任何人都不意外……每个应用层公司现在都自称是“AI公司”——正如我们尽力过滤所见，这导致了您在今年MAD生态右侧看到的新标志的爆炸性增长

在结构方面的一些小改变：

在“横向应用”中，我们增加了一个“演示与设计”类别

我们将“搜索”重命名为“搜索/对话AI”，以反映LLM驱动的聊天式界面的兴起，如Perplexity。

在“行业”中，我们将“政府与情报”重新品牌为“航空航天、防御与政府”

“

开源基础设施”中的主要变化

我们合并了一直非常接近的类别，创建了一个横跨“数据访问”和“数据运营”的单一“数据管理”类别

我们增加了一个重要的新类别，“本地AI”，随着构建者寻求提供基础设施工具，将AI和LLM带入本地开发时代

To access the interactive version of the 2024 MAD landscape, please CLICK HERE

Number of companies

The 2024 MAD landscape features 2,011 logos in total.

That number is up from 1,416 last year, with 578 new entrants to the map.

For reference, the very first version in 2012 has just 139 logos.

The intensely (insanely?) crowded nature of the landscape primarily results from two back-to-back massive waves of company creation and funding.

The first wave was the 10-ish year long data infrastructure cycle, which started with Big Data and ended with the Modern Data Stack. The long awaited consolidation in that space has not quite happened yet, and the vast majority of the companies are still around.

The second wave is the ML /AI cycle, which started in earnest with Generative AI. As we are in the early innings of this cycle, and most companies are very young, we have been liberal in including young startups (a good number of which are seed stage still) in the landscape.

Note: those two waves are intimately related. A core idea of the MAD Landscape every year has been to show the symbiotic relationship between data infrastructure (on the left side); analytics/BI and ML/AI (in the middle) and applications (on the right side).

While it gets harder every year to fit the ever-increasing number of companies on the landscape every year, but ultimately, the best way to think of the MAD space is as an assembly line – a full lifecycle of data from collection to storage to processing to delivering value through analytics or applications.

Two big waves + limited consolidation = lots of companies on the landscape.

Main changes in “ Infrastructure ” and “Analytics“

We’ve made very few changes to the overall structure of left side of the landscape – as we’ll see below (Is the Modern Data Stack dead?), this part of the MAD landscape has seen a lot less heat lately.

Some noteworthy changes: We renamed “Database Abstraction” to “Multi-Model Databases & Abstractions” , to capture the rising wave around an all-in-one ‘Multi-Model’ database group (SurrealDB*, EdgeDB); killed the “Crypto / Web 3 Analytics” section we experimentally created last year, which felt out of place in this landscape; and removed the “ Query Engine” section, which felt more like a part of a section than a separate section (all the companies in that section still appear on the landscape – Dremio, Starburst, PrestoDB etc).

Main changes in “ Machine Learning & Artificial Intelligence ”

With the explosion of AI companies in 2023, this is where we found ourselves making by far the most structural changes.

• Given the tremendous activity in the ‘AI enablement’ layer in the last year, we added 3 new categories next to MLOps:

  • “ AI Observability” is a new category this year, with startups that help test, evaluate and monitor LLM applications
  • “ AI Developer Platforms” is close in concept to MLOps but we wanted to recognize the wave of platforms that are wholly focused on AI application development, in particular around LLM training, deployment and inference
  • “AI Safety & Security” includes companies addressing concerns innate to LLMs, from hallucination to ethics, regulatory compliance, etc

• If the very public beef between Sam Altman and Elon Musk has told us anything, it’s that the distinction between commercial and nonprofit is a critical one when it comes to foundational model developers. As such, we have split what was previously “Horizontal AI/AGI” into two categories: “Commercial AI Research” and “Nonprofit AI Research”

• The final change we made was another nomenclature one, where we amended “GPU Cloud” to reflect the addition of core infrastructure feature sets made by many of the GPU Cloud providers: “GPU Cloud / ML Infra ”

Main changes in “Applications”

• The biggest update here is that…to absolutely no one’s surprise…every application-layer company is now a self-proclaimed “AI company” – which, as much as we tried to filter, drove the explosion of new logos you see on the right side of the MAD landscape this year

• Some minor changes on the structure side:

  • In “Horizontal Applications,” we added a “Presentation & Design” category
  • We renamed “Search” to Search / Conversational AI ” to reflect the rise of LLM-powered chat-based interface such as Perplexity.
  • In “Industry”, we rebranded “Gov’t & Intelligence” to “Aerospace, Defense & Gov’t”

Main changes in “ Open Source **Infrastructure ”

• We merged categories that have always been close, creating a single “Data Management” category that spans both “Data Access” and “Data Ops”

• We added an important new category, “Local AI” as builders sought to provide the infrastructure tooling to bring AI & LLMs to the local development age

PART II: 24 THEMES WE’RE THINKING ABOUT IN 2024 我们在2024年关注的24个主题

在人工智能领域，事情变化之快、关注度之高，几乎不可能像往年那样提供一个全面的MAD空间“国情咨文”。

因此，这里有一个不同的格式：没有特定的顺序，以下是我们思考的或在对话中频繁出现的24个主题。其中一些是相当成熟的思考，有些则主要是问题或思想实验。

Things in AI are both moving so fast, and getting so much coverage, that it is almost impossible to provide a fully comprehensive “state of the union” of the MAD space, as we did in prior years.

So here’s for a different format: in no particular order, here are 24 themes that are top of mind and/or come up frequently in conversations. Some are fairly fleshed out thoughts, some largely just questions or thought experiments.

1.Structured vs unstructured data 结构化数据与非结构化数据

这部分既是一个主题，也是我们在对话中经常提到的内容，帮助解释当前的趋势。

因此，或许作为对2024年讨论的引言，这里有一个重要的提醒，它解释了一些关键的行业趋势。并非所有数据都是相同的。冒着过度简化的风险说，有两大主要的数据家族，围绕每个家族，都出现了一套工具和用例。

• 结构化数据流水线：即那些可以适应于行和列的数据。

  • 出于分析目的，数据从交易数据库和SaaS工具中提取出来，存储在云数据仓库（如Snowflake）中，进行转换，并使用商业智能（BI）工具进行分析和可视化，主要是为了理解现在和过去（所谓的“描述性分析”）。这个流水线经常通过下面讨论的现代数据栈启用，分析是核心用例。
  • 此外，结构化数据也可以“传统”的机器学习/人工智能模型中进行未来预测（预测性分析）——例如，哪些客户最有可能流失。

• 非结构化数据流水线：即通常不适应于行和列的数据世界，如文本、图像、音频和视频。非结构化数据主要是在生成式AI模型（LLM等）中被使用（训练和推理）。

这两大数据家族（及相关工具和公司）目前正经历着非常不同的命运和关注度水平。

非结构化数据（机器学习/人工智能）很热门；结构化数据（现代数据栈等）则不那么热门。

This is partly a theme, partly something we find ourselves mentioning a lot in conversations to help explain the current trends.

So, perhaps as an introduction to this 2024 discussion, here’s one important reminder upfront, which explains some of the key industry trends. Not all data is the same. At the risk of grossly over-simplifying, there are two main families of data, and around each family, a set of tools and use cases has emerged.

• Structured data pipelines: that is data that can fit into rows and columns.

  • For analytical purposes, data gets extracted from transactional databases and SaaS tools, stored in cloud data warehouses (like Snowflake), transformed, and analyzed and visualized using Business Intelligence (BI) tools, mostly for purposes of understanding the present and the past (what’s known as “descriptive analytics”). That assembly line is often enabled by the Modern Data Stack discussed below, with analytics as the core use case.
  • In addition, structured data can also get fed in “traditional” ML/AI models for purposes of predicting the future (predictive analytics) – for example, which customers are most likely to churn

• Unstructured data pipelines: that is the world of data that typically doesn’t fit into rows and columns such as text, images, audio and video. Unstructured data is largely what gets fed in Generative AI models (LLMs, etc), both to train and use (inference) them.

Those two families of data (and the related tools and companies) are experiencing very different fortunes and levels of attention right now.

Unstructured data (ML/AI) is hot; structured data (Modern Data Stack, etc) is not.

2.Is the Modern Data Stack dead? 现代数据栈死了吗？

不久前（可以说是2019-2021年间），在软件世界中，没有什么比现代数据栈（MDS）更吸引人的了。与“大数据”并列，它是少数几个从数据工程师跨越到更广泛受众（高管、记者、银行家）的基础设施概念之一。

现代数据栈基本上涵盖了上文提到的那种结构化数据流水线。它围绕着快速增长的云数据仓库展开，与之相关的供应商位于其上游（如Fivetran和Airbyte）、上层（DBT）和下游（Looker、Mode）。

随着Snowflake成为有史以来最大的软件IPO，对MDS的兴趣爆炸式增长，伴随着疯狂的、ZIRP（零利率政策）驱动的公司创立和风险资本投资。在一两年内，整个类别变得拥挤不堪——数据目录、数据可观测性、ETL、反向ETL等等。

作为一个解决真实问题的真正解决方案，现代数据栈也是一个营销概念和一系列跨数据价值链的初创公司之间的事实联盟。

快进到今天，情况大不相同。在2023年，我们预览了MDS“面临压力”，这种压力在2024年只会继续加剧。

MDS面临两个关键问题：

• 组建一个现代数据栈需要将多个独立供应商的各种最佳解决方案缝合在一起。结果是，在金钱、时间和资源方面成本高昂。这在后ZIRP预算削减时代，不被CFO办公室看好。

• MDS不再是街区上的酷孩子。生成式AI已经从高管、风险资本家和媒体那里夺走了所有的注意力——它需要我们上文提到的那种非结构化数据流水线。

观看：MAD播客：“现代栈死了吗？”与dbt Labs首席执行官Tristan Handy（Apple, Spotify）

Not that long ago (call it, 2019-2021), there wasn’t anything sexier in the software world than the Modern Data Stack (MDS). Alongside “Big Data”, it was one of the rare infrastructure concepts to have crossed over from data engineers to a broader audience (execs, journalists, bankers).

The Modern Data Stack basically covered the kind of structured data pipeline mentioned above. It gravitated around the fast-growing cloud data warehouses, with vendors positioned upstream from it (like Fivetran and Airbyte), on top of it (DBT) and downstream from it (Looker, Mode).

As Snowflake emerged as the biggest software IPO ever, interest in the MDS exploded, with rabid, ZIRP-fueled company creation and VC funding. Entire categories became overcrowded within a year or two – data catalogs, data observability, ETL, reverse ETL, to name a few.

A real solution to a real problem, the Modern Data Stack was also a marketing concept and a de-facto alliance amongst a number of startups across the value chain of data.

Fast forward to today, the situation is very different. In 2023, we had previewed that the MDS was “under pressure”, and that pressure will only continue to intensify in 2024.

The MDS is facing two key issues:

• Putting together a Modern Data Stack requires stitching together various best-of-breed solutions from multiple independent vendors. As a result, it’s costly in terms of money, time and resources. This is not looked upon favorable by the CFO office in a post ZIRP budget cut era
• The MDS is no longer the cool kid on the block. Generative AI has stolen all the attention from execs, VCs and the press – and it requires the kind of unstructured data pipelines we mentioned above.

Watch: MAD Podcast: Is the Modern Stack Dead? With Tristan Handy, CEO, dbt Labs (Apple, Spotify)

3.Consolidation in data infra , and the big getting bigger 数据基础设施的整合，以及大公司的进一步壮大

鉴于上述情况，2024年数据基础设施和分析领域接下来会发生什么？

情况可能是这样的：

• 许多围绕现代数据栈的初创公司将积极重新定位为“AI基础设施初创公司”，并尝试在现代AI栈中找到自己的位置（见下文）。这在某些情况下会奏效，但从结构化数据转向非结构化数据在大多数情况下可能需要根本性的产品演变。
• 数据基础设施行业终将看到一些整合。到目前为止，并购相对有限，但2023年确实发生了一些收购，无论是小型并入还是中等规模收购——包括Stemma（被Teradata收购）、Manta（被IBM收购）、Mode（被Thoughtspot收购）等（见下文第三部分）。
• 将会有更多的初创公司失败——随着风险资本资金的枯竭，情况变得艰难。许多初创公司大幅削减了成本，但某个时刻他们的现金流将会结束。不要期待看到耀眼的头条新闻，但这（不幸地）将会发生。
• 无论是规模扩大的公司还是上市公司，这个领域的较大公司将加倍投入到他们的平台战略，并努力覆盖更多功能。其中一部分将通过收购实现（因此有了整合），但很大一部分也将通过自主开发实现。

Given the above, what happens next in data infra and analytics in 2024?

It may look something like this:

• Many startups in and around the Modern Data Stack will aggressively reposition as “ AI ****infra startups” and try to find a spot in the Modern AI Stack (see below). This will work in some cases, but going from structured to unstructured data may require a fundamental product evolution in most cases.

• The data infra industry will finally see some consolidation. M&A has been fairly limited to date, but some acquisitions did happen in 2023, whether tuck-ins or medium-size acquisitions – including Stemma (acquired by Teradata), Manta (acquired by IBM), Mode (acquired by Thoughtspot), etc (see PART III below)

• There will be a lot more startup failure – as VC funding dried up, things have gotten tough. Many startups have cut costs dramatically, but at some point their cash runway will end. Don’t expect to see flashy headlines, but this will (sadly) happen.

• The bigger companies in the space, whether scale-ups or public companies, will double down on their platform play and push hard to cover ever more functionality. Some of it will be through acquisitions (hence the consolidation) but a lot of it will also be through homegrown development.

4.Checking in on Databricks vs Snowflake 关注Databricks与Snowflake的竞争

谈到这个领域的大公司，让我们来看看两个关键数据基础设施玩家Snowflake和Databricks之间的“泰坦之战”（参见我们2021年的MAD博客文章）。

Snowflake（历史上来自于结构化数据流水线世界）依旧是一家令人难以置信的公司，也是最高估值的公共技术股票之一（截至撰写时的市值/未来十二个月收入比为14.8倍）。然而，就像许多软件行业一样，其增长显著放缓——它在2024财年以38%的年度产品收入增长结束，总收入达到26.7亿美元，截至撰写时预计未来十二个月收入增长22%。或许最重要的是，Snowflake给人的印象是一家在产品方面承受压力的公司——它在拥抱人工智能方面动作较慢，相对来说较少进行收购。最近，有些突然的CEO过渡是另一个有趣的数据点。

Databricks（历史上来自于非结构化数据流水线和机器学习世界）正在经历全方位的强劲势头，据报道（由于它仍是一家私营公司）以50%以上的增长结束FY’24，收入达到16亿美元。重要的是，Databricks正在成为一个关键的生成式AI玩家，既通过收购（最著名的是以13亿美元收购MosaicML）也通过自主产品开发——首先是作为喂养LLMs所需非结构化数据的重要存储库，也是模型的创造者，从Dolly到DBRX，后者是公司在撰写时刚刚宣布的一款新的生成式AI模型。

Snowflake与Databricks竞争的主要新进展是Microsoft Fabric的推出。2023年5月宣布的它是一个端到端的基于云的SaaS平台，用于数据和分析。它整合了许多微软产品，包括OneLake（开放的湖屋）、PowerBI和Synapse Data Science，并基本涵盖了所有数据和分析工作流程，从数据整合和工程到数据科学。如同大公司产品发布通常的情况，宣布和产品的实际情况之间存在差距，但结合微软在生成式AI中的重大推动，这可能成为一个强大的威胁（作为故事的一个额外转折，Databricks在很大程度上位于Azure之上）。

Speaking of big companies in the space, let’s check in on the “titanic shock” (see our MAD 2021 blog post) between the two key data infra players, Snowflake and Databricks.

Snowflake (which historically comes from the structured data pipeline world) remains an incredible company, and one of the highest valued public tech stocks (14.8x EV/NTM revenue as of the time of writing). However, much like a lot of the software industry, its growth has dramatically slowed down – it finished fiscal 2024 with a 38% year-over-year product revenue growth, totaling $2.67 billion, projecting 22% NTM rev growth as of the time of writing). Perhaps most importantly, Snowflake gives the impression of a company under pressure on the product front – it’s been slower to embrace AI, and comparatively less acquisitive. The recent, and somewhat abrupt, CEO transition is another interesting data point.

Databricks (which historically comes from the unstructured data pipeline and machine learning world) is experiencing all-around strong momentum, reportedly (as it’s still a private company) closing FY’24 with $1.6B in revenue with 50$1.3B) and homegrown product development – first and foremost as a key respiratory for the kind of unstructured data that feeds LLMs, but also as creator of models, from Dolly to DBRX, a new generative AI model the company just announced at the time of writing.

The major new evolution in the Snowflake vs Databricks rivalry is the launch of Microsoft Fabric. Announced in May 2023, it’s an end-to-end, cloud-based SaaS platform for data and analytics. It integrates a lot of Microsoft products, including OneLake (open lakehouse), PowerBI and Synapse Data Science, and covers basically all data and analytics workflows, from data integration and engineering to data science. As always for large company product launches, there’s a gap between the announcement and the reality of the product, but combined with Microsoft’s major push in Generative AI, this could become a formidable threat (as an additional twist to the story, Databricks largely sits on top of Azure).

5.BI in 2024, and Is Generative AI about to transform data analytics ? 2024年的商业智能，以及生成式AI是否即将转变数据分析？

在现代数据栈和结构化数据流水线世界的所有部分中，最有可能被重新发明的类别是商业智能(BI)。我们在2019年的MAD中强调了BI行业几乎完全整合的情况，并在2021年的MAD中讨论了指标存储的出现。

BI/分析的转变比我们预期的要慢。这个行业仍然主要被较老的产品所主导，如微软的PowerBI、Salesforce的Tableau和谷歌的Looker，这些有时在更广泛的销售合同中免费捆绑提供。发生了一些进一步的整合（Thoughtspot收购了Mode；Sisu被Snowflake悄悄收购）。一些年轻的公司正在采取创新的方法，无论是规模扩大的公司（见dbt及其语义层/MetricFlow）还是初创公司（见Trace*及其指标树），但它们通常还处于旅程的早期。

除了在数据提取和转换中可能发挥强大作用外，生成式AI可能在超级增强和民主化数据分析方面产生深远的影响。

确实已经有很多活动。OpenAI推出了代码解释器，后来更名为高级数据分析。微软为Excel中的金融工作者推出了Copilot AI聊天机器人。跨云供应商、Databricks、Snowflake、开源以及大量初创公司，许多人正在研究或已发布“文本到SQL”的产品，以帮助使用自然语言运行数据库查询。

这个承诺既令人兴奋又可能具有颠覆性。数据分析的圣杯一直是其民主化。如果自然语言成为笔记本、数据库和BI工具的接口，将使更广泛的人群能够进行分析。

然而，许多BI行业的人持怀疑态度。SQL的精确性和理解查询背后业务上下文的细微差别被认为是自动化的重大障碍。

Of all parts of the Modern Data Stack and structured data pipelines world, the category that has felt the most ripe for reinvention is Business Intelligence. We highlighted in the 2019 MAD how the BI industry had almost entirely consolidated, and talked about the emergence of metrics stores in the 2021 MAD.

The transformation of BI/analytics has been slower than we’d have expected. The industry remains largely dominated by older products, Microsoft’s PowerBI, Salesforce’s Tableau and Google’s Looker, which sometimes get bundled in for free in broader sales contracts. Some more consolidation happened (Thoughtspot acquired Mode; Sisu was quietly acquired by Snowflake). Some young companies are taking innovative approaches, whether scale-ups (see dbt and their semantic layer/MetricFlow) or startups (see Trace* and their metrics tree), but they’re generally early in the journey.

In addition to potentially playing a powerful role in data extraction and transformation, Generative AI could have a profound impact in terms of superpowering and democratizing data analytics.

There’s certainly been a lot of activity. OpenAI launched Code Interpreter, later renamed to Advanced Data Analysis. Microsoft launched a Copilot AI chatbot for finance workers in Excel. Across cloud vendors, Databricks, Snowflake, open source and a substantial group of startups, a lot of people are working on or have released “text to SQL” products, to help run queries into databases using natural language.

The promise is both exciting and potentially disruptive. The holy grail of data analytics has been its democratization. Natural language, if it were to become the interface to notebooks, databases and BI tools, would enable a much broader group of people to do analysis.

Many people in the BI industry are skeptical, however. The precision of SQL and the nuances of understanding the business context behind a query are considered big obstacles to automation.

6.The Rise of the Modern AI Stack 现代AI栈的崛起

我们迄今为止讨论的许多内容都与结构化数据流水线的世界有关。

如前所述，非结构化数据基础设施正在经历一个截然不同的时刻。非结构化数据是LLMs的食粮，对它的需求极其旺盛。每个正在实验或部署生成式AI的公司都在重新发现老生常谈：“数据是新石油”。每个人都想要LLMs的力量，但是训练在他们的（企业）数据上。

大小不同的公司都在争先恐后地提供生成式AI的基础设施。

几家AI规模扩大的公司一直在积极发展他们的产品以利用市场动力——从Databricks（见上文）到Scale AI（他们发展了标签基础设施，最初为自驾车市场开发，现在作为企业数据流水线与OpenAI及其他公司合作）到Dataiku*（他们推出了LLM Mesh，使全球2000强公司能够无缝地跨多个LLM供应商和模型工作）。

与此同时，一个新一代的AI基础设施初创公司正在涌现，涵盖多个领域，包括：

• 向量数据库，以一种格式（向量嵌入）存储数据，生成式AI模型可以消费。专门的供应商（Pinecone、Weaviate、Chroma、Qudrant等）度过了旗舰年，但一些现有的数据库玩家（MongoDB）也迅速反应，增加了向量搜索能力。
• 框架（LlamaIndex、Langchain等），连接和协调所有移动部件
• 保护栏，位于LLM和用户之间，确保模型提供的输出遵循组织的规则。
• 评估器，帮助测试、分析和监控生成式AI模型性能，这是一个难题，如公众对公共基准的普遍不信任所示
• 路由器，帮助实时跨不同模型指导用户查询，以优化性能、成本和用户体验
• 成本监控，帮助监控使用LLMs的成本
• 端点，实际上是抽象了底层基础设施（如模型）复杂性的APIs

我们一直在抗拒使用“现代AI栈”这个术语，鉴于现代数据栈的历史。

但这个表达捕捉到了许多平行点：许多这些初创公司是当今的“热门公司”，就像他们之前的MDS公司一样，他们倾向于成群结队，形成营销联盟和产品合作伙伴关系。也许还有

而这一新一代的AI基础设施初创公司将面临一些与MDS公司之前相同的挑战：这些类别中的任何一个都足够大以建立一个数十亿美元的公司吗？哪一部分将由大公司（主要是云提供商，但也包括Databricks和Snowflake）自己构建？

A lot of what we’ve discussed so far had to do with the world of structured data pipelines.

As mentioned, the world of unstructured data infrastructure is experiencing a very different moment. Unstructured data is what feeds LLMs, and there’s rabid demand for it. Every company that’s experimenting or deploying Generative AI is rediscovering the old cliche: “data is the new oil”. Everyone wants the power of LLMs, but trained on their (enterprise) data.

Companies big and small have been rushing into the opportunity to provide the infrastructure of Generative AI.

Several AI scale-ups have been aggressively evolving their offerings to capitalize on market momentum – everyone from Databricks (see above) to Scale AI (which evolved their labeling infrastructure, originally developed for the self-driving car market, to partner as an enterprise data pipeline with OpenAI and others) to Dataiku* (which launched their LLM Mesh to enable Global 2000 companies to seamlessly work across multiple LLM vendors and models).

Meanwhile a new generation of AI infra startups is emerging, across a number of domains, including:

• Vector databases, which store data in a format (vector embeddings) that Generative AI models can consume. Specialized vendors (Pinecone, Weaviate, Chroma, Qudrant etc) have had a banner year, but some incumbent database players (MongoDB) were also quick to react and add vector search capabilities.
• Frameworks (LlamaIndex, Langchain etc), which connect and orchestrate all the moving pieces
• Guardrails, which sit between an LLM and users and make sure the model provides outputs that follow the organization’s rules.
• Evaluators which help test, analyze and monitor Generative AI model performance, a hard problem as demonstrated by the general distrust in public benchmarks
• Routers, which help direct user queries across different models in real time, to optimize performance, cost and user experience
• Cost guards, whichhelp monitor the costs of using LLMs
• Endpoints, effectively APIs that abstract away the complexities of underlying infrastructure (like models)

We’ve been resisting using the term “Modern AI Stack”, given the history of the Modern Data Stack.

But the expression captures the many parallels: many of those startups are the “hot companies” of the day, just like MDS companies before them, they tend to travel in pack, forging marketing alliances and product partnerships. And perhaps there

And this new generation of AI infra startups is going to face some of the same challenges as MDS companies before them: are any of those categories big enough to build a multi-billion dollar company? Which part will big companies (mostly cloud providers, but also Databricks and Snowflake) end up building themselves?

WATCH – we featured many emerging startups on the MAD Podcast:

• Vector databases:

  • MAD Podcast with Edo Liberty, CEO, Pinecone
  • MAD Podcast with Jeff Huber, CEO, Chroma
  • MAD Podcast with Bob van Luijt, Weaviate

• MAD Podcast with Shreya Rajpal, CEO, Guardrails AI

• MAD Podcast with Jerry Liu, CEO, Llama Index

• MAD Podcast with Sharon Zhou, CEO, Lamini

7.Where are we in the AI hype cycle? 我们处于AI炒作周期的哪个阶段？

AI有着几十年的历史，经历了多次AI夏天和冬天。仅在过去的10-12年中，我们经历了第三个AI炒作周期：2012年ImageNet之后，深度学习进入聚光灯下，2013-2015年有一个；另一个大约在2017-2018年间，那时聊天机器人热潮兴起，TensorFlow崛起；现在自2022年11月以来，是生成式AI。

这个炒作周期特别激烈，以至于感觉像是AI泡沫，原因有很多：技术令人难以置信地令人印象深刻；它非常直观并跨越到了技术圈之外的广泛受众；对于坐拥大量闲散资金的风险投资家来说，这已经成为了唯一的游戏，因为技术中的几乎所有其他事物都处于低迷状态。

炒作带来了所有通常的好处（“没有任何伟大的成就是没有非理性狂热的”，“让一千朵花开放”阶段，有大量资金可用于雄心勃勃的项目）和噪音（一夜之间每个人都是AI专家，每个初创公司都是AI公司，太多的AI会议/播客/通讯……我们敢说，太多的AI市场图表了吗？）。

任何炒作周期的主要问题都是不可避免的反弹。

这个市场阶段内置了相当多的“古怪性”和风险：该领域的标志性公司具有非常不寻常的法律和治理结构；发生了很多尚未完全理解或披露的“计算力换股权”交易（可能存在资金循环）；许多顶级初创公司由AI研究团队运营；许多风险资本的交易让人回想起ZIRP时期：“抢地盘”，年轻公司的大轮次和惊人的估值。

AI炒作确实出现了裂缝（见下文），但我们仍处于每周都有新事物让每个人惊叹的阶段。像沙特阿拉伯报告的400亿美元AI基金这样的新闻似乎表明，资金流入这个领域不会很快停止。

AI has a multi decade-long history of AI summers and winters. Just in the last 10-12 years, this is the third AI hype cycle we’ve experienced: there was one in 2013-2015 after deep learning came to the limelight post ImageNet 2012; another one sometime around 2017-2018 during the chatbot boom and the rise of TensorFlow; and now since November 2022 with Generative AI.

This hype cycle has been particularly intense, to the point of feeling like an AI bubble, for a number of reasons: the technology is incredibly impressive; it is very visceral and crossed over to a broad audience beyond tech circles; and for VCs sitting on a lot of dry powder, it’s been the only game in town as just about everything else in technology has been depressed.

Hype has brought all the usual benefits (“nothing great has ever been achieved without irrational exuberance”, “let a 1000 flowers bloom” phase, with lots of money available for ambitious projects) and noise (everyone is an AI expert overnight, every startup is an AI startup, too many AI conferences/podcasts/newsletters… and dare we say, too many AI market maps???).

The main issue of any hype cycle is the inevitable blowback.

There’s a fair amount of “quirkiness” and risk built into this market phase: the poster-child company for the space has a very unusual legal and governance structure; there are a lot of “compute for equity” deals happening (with potential round-tripping) that are not fully understood or disclosed; a lot of top startups are run by teams of AI researchers; and a lot of VC dealmaking is reminiscent of the ZIRP times: “land grabs”, big rounds and eye-watering valuations for very young companies.

There certainly have been cracks in AI hype (see below), but we’re still in a phase where every week a new thing blows everyone’s minds. And news like the reported $40B Saudi Arabia AI fund seem to point that money flows into the space are not going to stop anytime soon.

8.Experiments vs reality: was 2023 a headfake? 实验与现实：2023年是虚晃一枪吗？

与上文相关——鉴于炒作，到目前为止有多少是真实的，相对于仅仅是实验性的呢？

2023年是充满行动的一年：a) 每个技术供应商都急于将生成式AI纳入他们的产品供应，b) 每个全球2000强的董事会都要求他们的团队“做AI”，一些企业部署以创纪录的速度发生，包括在受监管行业的公司，如摩根斯坦利和花旗银行，以及c) 当然，消费者对生成式AI应用表现出极大的兴趣。

因此，2023年是一个重大胜利的年份：OpenAI达到了每年20亿美元的运营率；Anthropic以一种速度增长，使其能够预测2024年的收入为8.5亿美元；Midjourney在没有投资和只有40人团队的情况下，收入增长到2亿美元；Perplexity AI的月活跃用户量从0增长到1000万等。

我们应该持怀疑态度吗？一些顾虑：

• 在企业中，许多支出用于概念验证，或容易获得的成功，往往来自创新预算。
• 有多少是由于高管不想显得措手不及，而不是解决实际的商业问题所驱动的？
• 在消费者领域，AI应用显示出高流失率。这有多少是仅仅出于好奇？
• 在个人和专业生活中，许多人报告说他们不完全确定如何使用生成式AI应用和产品
• 我们应该将Inflection AI快速折叠的决定视为一个承认，即世界不需要另一个AI聊天机器人，甚至是LLM提供商吗？

Related to the above – given the hype, how much has been real so far, vs merely experimental?

2023 was an action packed year: a) every tech vendor rushed to include Generative AI in their product offering, b) every Global 2000 board mandated their teams to “do AI”, and some enterprise deployments happened a record speed, including at companies in regulated industries like Morgan Stanley and Citibank and c) of course, consumers showed rabid interest for Generative AI apps.

As a result, 2023 was a year of big wins: OpenAI reached $2B in annual run rate; Anthropic grew at a pace that allowed it to forecast$850M in revenues for 2024; Midjourney grew to $200M in revenue with no investment and a team of 40; Perplexity AI went from 0 to 10 million monthly active users, etc.

Should we be cynical? Some concerns:

• In the enterprise, a lot of the spend was on proof of concepts, or easy wins, often coming out of innovation budgets.

• How much was driven by executives wanting to not appear flat-footed, vs solving actual business problems?

• In consumer, AI apps show high churn. How much was it mere curiosity?

• Both in their personal and professional lives, many report not being entirely sure what to do with Generative AI apps and products

• Should we view Inflection AI’s decision to fold quickly an admission that the world doesn’t need yet another AI chatbot, or even LLM provider?

9.LLM companies: maybe not so commoditized after all? LLM公司：也许毕竟不是那么通用化？

数十亿的风险资本和企业资金被投资于基础模型公司。

因此，在过去的18个月里，每个人都喜欢问的问题是：我们是否正在见证资本被大量投入到最终变成通用产品的惊人燃烧？或者这些LLM提供商是新的AWS、Azure和GCP？

对于涉及的公司来说，一个令人不安的事实是，似乎没有任何LLM正在构建一个持久的性能优势。截至撰写时，Claude 3 Sonnet和Gemini Pro 1.5的性能优于GPT-4，而GPT-4又优于Gemini 1.0 Ultra，等等——但这似乎每几周就会变化一次。性能也可能波动——ChatGPT在某一时刻“失去理智”并“变得懒惰”，但这是暂时的。

此外，开源模型（如Llama 3、Mistral和DBRX等）在性能方面迅速赶上。

另外——市场上的LLM提供商比最初看起来的要多得多。几年前，普遍的观点是，可能只有一两家LLM公司，存在赢者通吃的动态——部分原因是世界上只有少数几个人具有扩展Transformers的必要专业知识。

结果证明，有能力的团队比最初预期的要多。除了OpenAI和Anthropic外，还有许多初创公司正在进行基础AI工作——Mistral、Cohere、Adept、AI21、Imbue、01.AI等——然后当然还有谷歌、Meta等公司的团队。

尽管如此——到目前为止，LLM提供商似乎做得相当不错。非常感谢，OpenAI和Anthropic的收入正在以非凡的速度增长。也许LLM模型确实会变得通用化，但LLM公司仍然面临着巨大的商业机会。他们已经成为“全栈”公司，提供应用程序和工具给多个受众（消费者、企业、开发人员），在底层模型之上。

或许，与云服务提供商的类比确实相当贴切。AWS、Azure和GCP通过应用程序/工具层吸引并保留客户，并通过大体上无差别的计算/存储层实现盈利。

Billions of venture capital and corporate money are being invested in foundational model companies.

Hence everyone’s favorite question in the last 18 months: are we witnessing a phenomenal incineration of capital into ultimately commoditized products? Or are those LLM providers the new AWS, Azure and GCP?

A troubling fact (for the companies involved) is that no LLM seems to be building a durable performance advantage. At the time of writing, Claude 3 Sonnet and Gemini Pro 1.5 perform better than GPT-4 which performs better than Gemini 1.0 Ultra, and so on and so forth – but this seems to change every few weeks. Performance also can fluctuate – ChatGPT at some point “lost its mind” and “got lazy”, temporarily.

In addition, open source models (Llama 3, Mistral and others like DBRX) are quickly catching up in terms of performance.

Separately – there are a lot more LLM providers on the market than could have appeared at first. A couple of years ago, the prevailing narrative was that there could be only one or two LLM companies, with a winner-take-all dynamic – in part because there was a tiny number of people around the world with the necessary expertise to scale Transformers.

It turns out there are more capable teams than first anticipated. Beyond OpenAI and Anthropic, there are a number of startups doing foundational AI work – Mistral, Cohere, Adept, AI21, Imbue, 01.AI to name a few – and then of course the teams at Google, Meta, etc.

Having said that – so far the LLM providers seem to be doing just fine. OpenAI and Anthropic revenues are growing at extraordinary rates, thank you very much. Maybe the LLM models do get commoditized, the LLM companies still have an immense business opportunity in front of them. They’ve already become “full stack” companies, offering applications and tooling to multiple audiences (consumer, enterprise, developers), on top of the underlying models.

Perhaps the analogy with cloud vendors is indeed pretty apt. AWS, Azure and GCP attract and retain customers through an application/tooling layer and monetize through a compute/storage layer that is largely undifferentiated.

WATCH:

• MAD Podcast with Ori Goshen, co-founder, AI21 Labs

• MAD Podcast with Kanjun Qiu, CEO, Imbue

10.LLMs , SLMs and a hybrid future LLM、SLM及其混合未来

尽管对大型语言模型（LLM）的兴奋不断，过去几个月的一个明确趋势是小型语言模型（SLM）的加速发展，例如Meta的Llama-2-13b、Mistral的Mistral-7b和Mixtral 8x7b、以及微软的Phi-2和Orca-2。

虽然LLM正在变得越来越大（据报道，GPT-3拥有1750亿参数，GPT-4拥有1.7万亿，世界正在等待更加庞大的GPT-5），SLM正成为许多用例的强大替代品，因为它们运行成本更低，更易于微调，往往还提供强大的性能。

另一个加速的趋势是专业模型的兴起，专注于特定任务，如编码（Code-Llama、Poolside AI）或特定行业（例如，彭博的财经模型，或是为材料科学建模的初创公司Orbital Materials等）。

正如我们已经在许多企业部署中看到的，世界正迅速向混合架构演进，结合多个模型。

尽管价格一直在下降（见下文），但大型专有LLM仍然非常昂贵，存在延迟问题，因此用户/客户将越来越多地部署模型组合，大型和小型、商业和开源、通用和专业，以满足他们的特定需求和成本约束。

For all the excitement about Large Language Models, one clear trend of the last few months has been the acceleration of small language models (SLMs), such as Llama-2-13b from Meta, Mistral-7b and Mixtral 8x7b from Mistral and Phi-2 and Orca-2 from Microsoft.

While the LLMs are getting ever bigger (GPT-3 reportedly having 175 billion parameters, GPT-4 reportedly having 1.7 trillion, and the world waiting for an even more massive GPT-5), SLMs are becoming a strong alternative for many use cases are they are cheaper to operate, easier to finetune, and often offer strong performance.

Another trend accelerating is the rise of specialized models, focused on specific tasks like coding (Code-Llama, Poolside AI) or industries (e.g. Bloomberg’s finance model, or startups Orbital Materials building modelsl for material sciences, etc).

As we are already seeing across a number of enterprise deployments, the world is quickly evolving towards hybrid architectures, combining multiple models.

Although prices have been going down (see below), big proprietary LLMs are still very expensive, experience latency problems, and rso users/customers will increasingly be deploying combinations of models, big and small, commercial and open source, general and specialized, to meet their specific needs and cost constraints.

Watch: MAD Podcast with Eiso Kant, CTO, Poolside AI (also: Apple Podcasts, Spotify)

11.Is traditional AI dead? 传统AI死了吗？

随着ChatGPT的推出，发生了一件有趣的事情：直到那时为止部署的许多AI一夜之间被贴上了“传统AI”的标签，与“生成式AI”形成对比。

这对许多一直被认为是从事前沿工作的AI实践者和公司来说，有点儿震惊，因为“传统”这个术语明确暗示了新事物将彻底取代所有形式的AI。

实际情况要复杂得多。传统AI和生成式AI在根本上是非常互补的，因为它们处理不同类型的数据和用例。

现在被标记为“传统AI”，或偶尔被称为“预测AI”或“表格AI”的，也是现代AI（基于深度学习）的一部分。然而，它通常关注于结构化数据（见上文），和诸如推荐、流失预测、价格优化、库存管理等问题。“传统AI”在过去十年里经历了巨大的采用，已经在全球成千上万的公司中大规模部署。

相比之下，生成式AI主要操作非结构化数据（文本、图像、视频等）。它在不同类别的问题上表现出色（代码生成、图像生成、搜索等）。

在这里，未来也是混合的：公司将使用LLM完成某些任务，预测模型完成其他任务。最重要的是，它们经常会结合使用——LLM可能不擅长提供精确预测，比如流失预测，但你可以使用一个LLM调用另一个专注于提供那个预测的模型的输出，反之亦然。

A funny thing happened with the launch of ChatGPT: much of the AI that had been deployed up until then got labeled overnight as “Traditional AI” , in contrast to “Generative AI”.

This was a little bit of a shock to many AI practitioners and companies that up until then were considered to be doing leading-edge work, as the term “traditional” clearly suggests an impending wholesale replacement of all forms of AI by the new thing.

The reality is a lot more nuanced. Traditional AI and Generative AI are ultimately very complementary as they tackle different types of data and use cases.

What is now labeled as “traditional AI”, or occasionally as “predictive AI” or “tabular AI”, is also very much part of modern AI (deep learning based). However, it generally focuses on structured data (see above), and problems such as recommendations, churn prediction, pricing optimization, inventory management. “Traditional AI” has experienced tremendous adoption in the last decade, and it’s already deployed at scale in production in thousands of companies around the world.

In contrast, Generative AI largely operates on unstructured data (text, image, videos, etc.). Is exceptionally good at a different class of problems (code generation, image generation, search, etc).

Here as well, the future is hybrid: companies will use LLMs for certain tasks, predictive models for other tasks. Most importantly, they will often combine them – LLMs may not be great at providing a precise prediction, like a churn forecast, but you could use an LLM that calls on the output of another model which is focused on providing that prediction, and vice versa.

12.Thin wrappers, thick wrappers and the race to be full stack 薄封装、厚封装以及成为全栈的竞赛

“薄封装”是2023年大家喜欢使用的一种轻蔑术语。如果你的核心能力由别人的技术（如OpenAI）提供，那么构建持久的价值和差异化将会很困难，这是人们的论点。几个月前的报告显示，像Jasper这样的初创公司在经历了流星般的收入增长后遇到了困难，似乎证实了这种思维方式。

有趣的问题是，随着年轻初创公司构建更多功能，随着时间的推移会发生什么。薄封装是否会变成厚封装？

到了2024年，厚封装似乎有通过以下方式实现差异化的路径：

• 集中解决一个特定的问题，通常是垂直的——任何过于横向的东西都有处于大科技公司“杀手区”的风险
• 构建特定于该问题的工作流、协作和深度集成
• 在AI模型层面做大量工作——无论是使用特定数据集对模型进行微调，还是创建为其特定业务量身定做的混合系统（LLM、SLM等）

换句话说，他们需要既窄又“全栈”（既包括应用程序又包括基础设施）。

“Thin wrappers” was the dismissive term everyone loved to use in 2023. It’s hard to build long lasting value and differentiation if your core capabilities are provided by someone else’s technology (like OpenAI), the argument goes. And reports a few months ago that startups like Jasper were running into difficulties, after experiencing a meteoric revenue rise, seem to corroborate that line of thinking.

The interesting question is what happens over time, as young startups build more functionality. Do thin wrappers become thick wrappers?

In 2024, it feels like thick wrappers have a path towards differentiation by:

• Focusing on a specific problem, often vertical – as anything too horizontal runs the risk of being in the “kill zone” of Big Tech
• Building workflow, collaboration and deep integrations, that are specific to that problem
• Doing a lot of work at the AI model level – whether finetuning models with specific datasets or creating hybrid systems (LLMs, SLMs, etc) tailored for their specific business

In other words, they will need to be both narrow and “full stack” (both applications and infra).

13.Interesting areas to watch in 2024: AI agents, Edge AI 2024年值得关注的领域：AI代理，边缘AI

过去一年里，围绕AI代理的概念——基本上是智能系统的最后一公里，可以执行任务，通常以协作的方式——引起了大量兴奋。这可以是任何事情，从帮助预订旅行（消费者用例）到自动运行完整的SDR活动（生产力用例）到RPA风格的自动化（企业用例）。

AI代理是自动化的圣杯——一个“文本到行动”的范式，AI为我们完成任务。

每隔几个月，AI世界就会因为类似代理的产品而疯狂，从去年的BabyAGI到最近的Devin AI（一个“AI软件工程师”）。然而，总的来说，到目前为止，这种兴奋大多被证明为过早。在复杂系统涉及多个模型可以一起工作并代表我们采取实际行动之前，首先需要大量工作来使生成式AI变得更加稳健和可预测。还有一些缺失的组件——例如需要构建更多记忆进AI系统。然而，预计AI代理在未来一两年将是一个特别激动人心的领域。

另一个有趣的领域是边缘AI。正如存在一个巨大的市场，LLMs在大规模上运行并作为终端提供，AI的一个圣杯一直是能够在设备上本地运行的模型，特别是不需要GPU的手机，但也包括智能的、物联网类型的设备。这个领域非常活跃：Mixtral、Ollama、Llama.cpp、Llamafile、GPT4ALL（Nomic）。谷歌和苹果也可能越来越活跃。

There’s been plenty of excitement over the last year around the concept of AI agents – basically the last mile of an intelligent system that can execute tasks, often in a collaborative manner. This could be anything from helping to book a trip (consumer use case) to automatically running full SDR campaigns (productivity use case) to RPA-style automation (enterprise use case).

AI agents are the holy grail of automation – a “text to action” paradigm where AI just gets stuff done for us.

Every few months, the AI world goes crazy for an agent-like product, from BabyAGI last year to Devin AI (an “AI software engineer”) just recently. However, in general, much of this excitement has proven premature to date. There’s a lot of work to be done first to make Generative less brittle and more predictable, before complex systems involving several models can work together and take actual actions on our behalf. There are also missing components – such as the need to build more memory into AI systems. However, expect AI agents to be a particularly exciting area in the next year or two.

Another interesting area is Edge AI. As much as there is a huge market for LLMs that run at massive scale and delivered as end points, a holy grail in AI has been models that can run locally on a device, without GPUs, in particular phones, but also intelligent, IoT-type devices. The space is very vibrant: Mixtral, Ollama, Llama.cpp, Llamafile, GPT4ALL (Nomic). Google and Apple are also likely to be increasingly active.

14.Is Generative AI heading towards AGI, or towards a plateau? 生成式AI是朝向AGI发展，还是走向一个高原期？

考虑到所有关于AI的激动人心的讨论，以及似乎每周都有惊人新产品问世，这几乎是一个亵渎的问题——但是否存在一个世界，在这个世界中，生成式AI的进步放缓，而不是一路加速至AGI？那将意味着什么？

这个论点有两方面：a) 基础模型是一种蛮力练习，我们将耗尽资源（计算力、数据）来喂养它们，以及b) 即使我们不耗尽资源，最终通往AGI的路径是推理，这是LLMs无法做到的。

有趣的是，这或多或少是6年前行业正在进行的同样讨论，正如我们在2018年的博客文章中描述的。事实上，似乎改变最多的是我们投入到（越来越能干的）模型中的大量数据和计算力。

我们距离任何类型的“耗尽”有多近非常难以评估。“耗尽计算力”的边界似乎每天都在被推迟。NVIDIA最近宣布的Blackwell GPU系统，该公司表示它可以部署一个27万亿参数的模型（相比之下，GPT-4为1.7万亿）。

数据部分是复杂的——围绕耗尽合法许可数据（参见所有OpenAI的许可交易）存在更多策略性问题，以及围绕通常耗尽文本数据的更广泛问题。围绕合成数据肯定有很多工作正在进行。Yann LeCun讨论了如何将模型提升到下一个水平，可能需要它们能够摄取更丰富的视频输入，这目前还不可能实现。

从创业生态系统参与者（创始人、投资者）的狭窄视角来看，这个问题在中期内可能并不那么重要——如果生成式AI从明天开始停止进步，我们仍然有数年的机会在各个垂直领域和用例中部署我们当前拥有的技术。

It’s almost a sacrilegious question to ask given all the breathless takes on AI, and the incredible new products that seem to come out every week – but is there a world where progress in Generative AI slows down rather than accelerates all the way to AGI? And what would that mean?

The argument is twofold: a) foundational models are a brute force exercise, and we’re going to run out of resources (compute, data) to feed them, and b) even if we don’t run out, ultimately the path to AGI is reasoning, which LLMs are not capable of doing.

Interestingly, this is more or less the same discussion as the industry was having 6 years ago, as we described in a 2018 blog post. Indeed what seems to have changed mostly is the sheer amount of data and compute we’ve thrown at (increasingly capable) models.

How close we are from any kind of “running out” is very hard to assess. The frontier for “running out of compute” seems to be pushed back further every day. NVIDIA’s recently announced Blackwell GPU system, and the company says it can deploy a 27 trillion parameter model (vs 1.7 trillion for GPT-4).

The data part is complex – there’s a more tactical question around running out of legally licensed data (see all the OpenAI licensing deals), and a broader question around running out of textual data, in general. There is certainly a lot of work happening around synthetic data. Yann LeCun discussed how taking models to the next level would probably require them to be able to ingest much richer video input, which is not yet possible.

From the narrow perspective of participants in the startup ecosystem (founders, investors), perhaps the question matters less, in the medium term – if Generative AI stopped making progress tomorrow, we’d still have years of opportunity ahead deploying what we currently have across verticals and use cases.

15.The GPU wars (is NVIDIA overvalued?) GPU战争（NVIDIA是否被高估了？）

我们是否处于一个庞大周期的早期阶段，在这个周期中，计算成为世界上最宝贵的商品，或者我们是否在以一种肯定会导致大崩溃的方式大幅度扩建GPU生产？

作为几乎唯一一家能够提供适用于生成式AI的GPU的公司，NVIDIA确实正在经历其辉煌时刻，其股价上涨了五倍，达到2.2万亿美元的估值，自2022年底以来其总销售额增加了三倍，围绕其收益的巨大兴奋以及Jensen Huang在GTC的表现与泰勒·斯威夫特的大型活动相媲美，成为2024年最大的事件之一。

也许这部分是因为它是所有风险投资在AI中投资的数十亿美元的最终受益者？

生成式AI 投资：一种过程，通过这个过程，风险资本公司通过被称为“ 初创公司 ”的中介机构将大量资金转移到NVIDIA

— Matt Turck (@mattturck) 2023年6月14日

无论如何，尽管NVIDIA作为一家公司的能力无可否认，但其财富将取决于当前淘金热的可持续性如何。硬件行业艰难，准确预测台湾TSMC需要制造多少GPU是一门艰难的艺术。

此外，竞争对手正在尽力应对，从AMD到Intel再到三星；初创公司（如Groq或Cerebras）正在加速发展，可能还会有新公司成立，如Sam Altman传言中的7万亿美元芯片公司。一个包括谷歌、英特尔和高通在内的技术公司新联盟正在试图攻击NVIDIA的秘密武器：其CUDA软件，该软件使开发者依赖于Nvidia芯片。

我们的看法：随着GPU短缺的缓解，NVIDIA可能会在短至中期内面临下行压力，但长期来看，AI芯片制造商的前景仍然非常光明。

Are we in the early innings of a massive cycle where compute becomes the most precious commodity in the world, or dramatically over-building GPU production in a way that’s sure to lead to a big crash?

As pretty much the only game in town when it comes to Generative AI-ready GPUs, NVIDIA certainly has been having quite the moment, with a share price up five-fold to a $2.2 trillion valuation, and total sales three-fold since late 2022, massive excitement around its earnings and Jensen Huang at GTC rivaling Taylor Swift for the biggest event of 2024.

Perhaps this was also in part because it was the ultimate beneficiary of all the billions invested by VCs in AI?

Generative AI investing: a process by which venture capital firms transfer large amounts of money to NVIDIA via intermediaries known as “startups”

— Matt Turck (@mattturck) June 14, 2023

Regardless, for all its undeniable prowess as a company, NVIDIA’s fortunes will be tied to how sustainable the current gold rush will turn out to be. Hardware is hard, and predicting with accuracy how many GPUs need to be manufactured by TSMC in Taiwan is a difficult art.

In addition, competition is trying its best to react, from AMD to Intel to Samsung; startups (like Groq or Cerebras) are accelerating, and new ones may be formed, like Sam Altman’s rumored $7 trillion chip company. A new coalition of tech companies including Google, Intel and Qualcomm is trying to go after NVIDIA’s secret weapon: its CUDA software that keeps developers tied to Nvidia chips.

Our take: As the GPU shortage subsides, there may be short-to medium term downward pressure on NVIDIA, but the long term for AI chips manufacturers remains incredibly bright.

16.Open source ****AI : too much of a good thing? 开源AI：好事过头了吗？

这个问题只是为了稍微搅动一下局面。我们非常支持开源AI，显然这在过去一年左右已经成为一个大趋势。Meta对其Llama模型进行了重大推广，法国的Mistral从争议的焦点转变为生成式AI的新亮点，谷歌发布了Gemma，HuggingFace继续作为开源AI充满活力的中心不断上升，托管着大量模型。一些生成式AI中最创新的工作已经在开源社区完成。

然而，社区中也普遍存在一种膨胀感。现在有成千上万的开源AI模型可用。许多是玩具或周末项目。模型在排名中上上下下，一些模型在短短几天内按照Github星标准（虽然是一个有缺陷的衡量标准，但仍然有效）经历了流星般的上升，只是最终从未转化为特别有用的东西。这让许多人感到眩晕。

我们的看法：市场将自我纠正，成功的开源项目将遵循幂律分布。

This one is just to stir a pot a little bit. We’re huge fans of open source AI, and clearly this has been a big trend of the last year or so. Meta made a major push with its Llama models, France’s Mistral went from controversy fodder to new shining star of Generative AI, Google released Gemma, and HuggingFace continued its ascension as the ever so vibrant home of open source AI, hosting a plethora of models. Some of the most innovative work in Generative AI has been done in the open source community.

However, there’s also a general feeling of inflation permeating the community. Hundreds of thousands of open source AI models are now available. Many are toys or weekend projects. Models go up and down the rankings, some of them experiencing meteoric rises by Github star standards (a flawed metric, but still) in just a few days, only to never transform into anything particularly usable. It’s been dizzying for many.

Our take: the market will be self-correcting, with a power law of successful open-source projects.

17.How much does AI actually cost?****

生成式AI的经济学是一个快速发展的话题。毫不奇怪，很多关于这个领域的未来都围绕着它展开——例如，如果提供AI驱动的答案的成本显著高于提供十个蓝色链接的成本，那么能否真正挑战谷歌的搜索地位？如果推理成本消耗了它们大部分的毛利润，软件公司能真正成为AI驱动的吗？

好消息是，如果你是AI模型的客户/用户：我们似乎正处于价格方面的竞争到底部的早期阶段，这一进程的发生速度比人们可能预测的要快。一个关键驱动因素是开源AI（如Mistral等）的并行崛起和商业推理供应商（如Together AI、Anyscale、Replit）采用这些开放模型并将它们作为端点提供。对于客户而言，切换成本非常低（除了使用不同模型产生不同结果的复杂性之外），这给OpenAI和Anthropic带来了压力。一个例子是嵌入模型的显著成本下降，其中多个供应商（如OpenAI、Together AI等）同时降低了价格。

从供应商的角度来看，构建和提供AI的成本仍然非常高。有报道称，Anthropic花费了其产生的收入的一半以上支付给像AWS和GCP这样的云提供商来运行其LLM。出版商的授权交易成本也很高。

另一方面，也许我们所有人作为生成技术的用户应该享受风险资本补贴的免费服务所带来的爆炸：

风险资本家为您带来了廉价的优步

风险资本家为您带来了廉价的Airbnb

风险资本家正在为您带来廉价的 AI 推理

不客气

— Matt Turck (@mattturck) 2024年1月26日

观看：MAD播客，嘉宾Nomic

The economics of Generative AI is a fast-evolving topic. And not surprisingly, a lot of the future of the space revolves around it – for example, can one seriously challenge Google in search, if the cost of providing AI-driven answers is significantly higher than the cost of providing ten blue links? And can software companies truly be AI-powered if the inference costs eat up chunks of their gross margin?

The good news, if you’re a customer/user of AI models: we seem to be in the early phase of a race to the bottom on the price side, which is happening faster than one may have predicted. One key driver has been the parallel rise of open source AI (Mistral etc) and commercial inference vendors (Together AI, Anyscale, Replit) taking those open models and serving them as end points. There are very little switching costs for customers (other than the complexity of working with different models producing different results), and this is putting pressure on OpenAI and Anthropic. An example of this has been the significant cost drops for embedding models where multiple vendors (OpenAI, Together AI etc) dropped prices at the same time.

From a vendor perspective, the costs of building and serving AI remain very high. It was reported in the press that Anthropic spent more than half of the revenue it generated paying cloud providers like AWS and GCP to run its LLMs. There’s the cost of licensing deals with publishers as well

On the plus side, maybe all of us as users of Generative technologies should just enjoy the explosion of VC-subsidized free services:

VCs brought you cheap Ubers VCs brought you cheap Airbnbs VCs are bringing you cheap AI inference YOU'RE WELCOME

— Matt Turck (@mattturck) January 26, 2024

Watch: MAD Podcast with Nomic

18.Big companies and the shifting political economy of AI : Has Microsoft won? 大公司和AI的政治经济变化：微软是否已经胜出？

这是2022年末每个人都在问的第一个问题，到了2024年，这个问题更加受到关注：大科技公司是否会捕获生成式AI中的大部分价值？

AI奖励规模——更多的数据、更多的计算能力、更多的AI研究人员往往会带来更大的力量。大科技公司对此非常清醒，与以往平台转变中的现有公司不同，对潜在的颠覆反应非常激烈。

在大科技公司中，微软显然像在玩4D棋。显然有与OpenAI的关系，但微软也与开源竞争对手Mistral建立了合作伙伴关系。它对ChatGPT的竞争对手Inflection AI（Pi）进行了投资，最近以壮观的方式收购了它。最终，所有这些合作伙伴关系似乎只是增加了对微软的云计算需求——Azure收入同比增长24%，在2024年第二季度达到330亿美元，其中6个百分点的Azure云增长归功于AI服务。

与此同时，谷歌和亚马逊与OpenAI的竞争对手Anthropic建立了合作并进行了投资（在撰写本文时，亚马逊刚刚承诺向该公司再投资27.5亿美元，为其计划的40亿美元投资的第二笔）。亚马逊还与开源平台Hugging Face建立了合作关系。谷歌和苹果据说正在讨论在苹果产品中集成Gemini AI。Meta可能通过全力支持开源AI而超越所有人。然后是中国正在发生的一切。

显而易见的问题是，初创公司有多少成长和成功的空间。第一梯队的初创公司（主要是OpenAI和Anthropic，可能很快Mistral也会加入他们）似乎已经建立了正确的合作关系，并达到了脱离重力的速度。对于许多其他初创公司，包括资金充足的公司，结果仍然非常不确定。

我们是否应该从Inflection AI决定让自己被收购，以及Stability AI的CEO问题中读出，对于一群“二线”生成式AI初创公司来说，商业牵引力更难实现的承认？

This was one of the first questions everyone asked in late 2022, and it’s even more top of mind in 2024: will Big Tech capture most of the value in Generative AI?

AI rewards size – more data, more compute, more AI researchers tends to yield more power. Big Tech has been keenly aware of this, and unlike incumbents in prior platform shifts, intensely reactive to the potential disruption ahead.

Among Big Tech companies, it certainly feels like Microsoft has been playing 4-D chess. There’s obviously the relationship with OpenAI, but Microsoft also partnered with open source rival Mistral. It invested in ChatGPT rival Inflection AI (Pi), only to acqui-hire it in spectacular fashion recently. And ultimately, all those partnerships seem to only create more need for Microsoft’s cloud compute – Azure revenue grew 24% year-over-year to reach $33 billion in Q2 2024, with 6 points of Azure cloud growth attributed to AI services.

Meanwhile, Google and Amazon have partnered with and invested in OpenAI rival Anthropic (at the time of writing, Amazon just committed another $2.75B to the company, in the 2nd tranche of its planned$4B investment). Amazon also partnered with open source platform Hugging Face. Google and Apple are reportedly discussing an integration of Gemini AI in Apple products. Meta is possibly under-cutting everyone by going full hog on open source AI. Then there is everything happening in China.

The obvious question is how much room there is for startups to grow and succeed. A first tier of startups (OpenAI and Anthropic, mainly, with perhaps Mistral joining them soon) seem to have struck the right partnerships, and reached escape velocity. For a lot of other startups, including very well funded ones, the jury is still very much out.

Should we read in Inflection AI’s decision to let itself get acquired, and Stability AI’s CEO troubles an admission that commercial traction has been harder to achieve for a group of “second tier” Generative AI startups?

19.Fanboying OpenAI – or not? 对OpenAI的狂热——还是不？

OpenAI继续引发关注——860亿美元的估值、收入增长、宫廷斗争，以及Sam Altman成为这一代的Steve Jobs：

Sam Altman在一天后返回 OpenAI ，就像Steve Jobs在12年后返回苹果，但对于TikTok一代来说 t.co/AHqH7WmVfF

— Matt Turck (@mattturck) 2023年11月18日

一些有趣的问题：

OpenAI是否试图做太多了？在所有11月的戏剧之前，有一个OpenAI开发者日，在那天OpenAI明确表示它将在AI中做尽一切，无论是垂直方向上（全栈）还是横向方向上（跨用例）：模型 + 基础设施 + 消费者搜索 + 企业 + 分析 + 开发工具 + 市场等。当一家初创公司在一个大的范式转变中作为早期领导者，并且实际上拥有无限的资本获取能力时，这并非前所未有的策略（Coinbase在加密领域某种程度上做到了这一点）。但这将是有趣的观察点：虽然这肯定会简化MAD生态图，但在竞争加剧的情况下，这将是一个巨大的执行挑战。从ChatGPT的懒惰问题到其市场努力的表现不佳，表明OpenAI并非不受商业引力法则的影响。

OpenAI和微软会分手吗？与微软的关系引人入胜——显然，微软的支持在资源（包括计算）和分销（企业中的Azure）方面为OpenAI提供了巨大的推动力，而且这一举措在生成式AI浪潮的初期被广泛视为微软的高明之举。同时，微软明确表示它不依赖OpenAI（拥有所有代码、权重、数据），它已经与竞争对手合作（例如Mistral），并且通过对Inflection AI的收购，现在其AI研究团队得到了大幅加强。

与此同时，OpenAI是否愿意继续在与微软的合作中单线程，而不是部署在其他云上？

鉴于OpenAI的宏大野心，以及微软对全球统治的目标，两家公司何时会得出结论，认为他们更多是竞争对手而不是合作伙伴？

OpenAI continues to fascinate – the $86B valuation, the revenue growth, the palace intrigue, and Sam Altman being the Steve Jobs of this generation:

Sam Altman returning to OpenAI after a day is like Steve Jobs returning to Apple after 12 years, but for the TikTok generation t.co/AHqH7WmVfF

— Matt Turck (@mattturck) November 18, 2023

A couple of interesting questions:

Is OpenAI trying to do too much? Before all the November drama, there was the OpenAI Dev Day, during which OpenAI made it clear that it was going to do everything in AI, both vertically (full stack) and horizontally (across use cases): models + infrastructure + consumer search + enterprise + analytics + dev tools + marketplace, etc. It’s not an unprecedented strategy when a startup is an early leader in a big paradigm shift with de facto unlimited access to capital (Coinbase sort of did it in crypto). But it will be interesting to watch: while it would certainly simplify the MAD Landscape, it’s going to be a formidable execution challenge, particularly in a context where competition has intensified. From ChatGPT laziness issues to the underwhelming performance of its marketplace effort suggest that OpenAI is not immune to the business law of gravity.

Will OpenAI and Microsoft break up? The relationship with Microsoft has been fascinating – obviously Microsoft’s support has been a huge boost for OpenAI in terms of resources (including compute) and distribution (Azure in the enterprise), and the move was widely viewed as a master move by Microsoft in the early days of the Generative AI wave. At the same time, Microsoft has made it clear that it’s not dependent on OpenAI (has all the code, weights, data), it has partnered with competitors (e.g. Mistral), and through the Inflection AI acqui-hire it now has considerably beefed up its AI research team.

Meanwhile, will OpenAI want to continue being single threaded in a partnership with Microsoft, vs being deployed on other clouds?

Given OpenAI’s massive ambitions, and Microsoft aim at global domination, at what point do both companies conclude that they’re more competitors than partners?

20.Will 2024 be the year of AI in the enterprise? 2024年将成为企业界AI之年吗？

正如上文提到的，2023年对企业界来说感觉像是那些关键年份之一，每个人都在争先恐后地拥抱新趋势，但除了一些概念验证之外，实际上并没有太多事情发生。

也许2023年 生成式AI 的最大赢家是世界各地的埃森哲（Accenture），据报道其通过 AI 咨询服务产生了20亿美元的费用。

到目前为止， AI 狂热的大赢家：顾问。

— Matt Turck (@mattturck) 2023年5月12日

不管怎样，人们极其期待2024年将是企业界AI的大年——或者至少是生成式AI的大年，因为传统AI在那里已经有了显著的足迹（见上文）。

但我们在回答全球2000型公司面临的一些关键问题方面还处于早期阶段：

有哪些用例？到目前为止，较低悬挂果实的用例主要是a) 为开发团队生成代码的共同驾驶员，b) 企业知识管理（搜索、文本摘要、翻译等），以及c) 客户服务的AI聊天机器人（一个早于生成式AI的用例）。当然还有其他用例（营销、自动化SDR等），但还有很多问题需要解决（共同驾驶模式与全自动化等）。

我们应该选择哪些工具？正如上所述，未来似乎是混合的，是商业供应商和开源、大型和小型模型、横向和垂直的GenAI工具的组合。但从哪里开始呢？

谁将部署和维护这些工具？在全球2000家公司中明显存在技能短缺。如果你认为招聘软件开发人员很难，那就试试招聘机器学习工程师。

我们如何确保它们不会产生幻觉？是的，围绕RAG和防护栏、评估等正在进行大量工作，但生成式AI工具可能完全错误的可能性，以及我们并不真正知道生成式AI模型如何工作的更广泛问题，是企业中的大问题。

什么是ROI？大型科技公司已经率先利用生成式AI满足自己的需求，并且他们展示了有趣的早期数据。在他们的财报电话会议中，Palo Alto Networks提到大约将其差旅费用服务成本减半，ServiceNow提到将我们的开发者创新速度提高了52%，但我们在理解企业中生成式AI的成本/回报方程式方面还处于早期阶段。

对于生成式AI供应商来说，好消息是企业客户有充足的兴趣分配预算（重要的是，不再是“创新”预算，而是实际的运营支出预算，可能从其他地方重新分配）和资源来解决这个问题。但我们可能谈论的是一个3-5年的部署周期，而不是一年。

观看：MAD播客，嘉宾Florian Dou

As mentioned above, 2023 in the enterprise felt like one of those pivotal years where everyone scrambles to embrace a new trend, but nothing much actually happens, beyond some proof-of-concepts.

Perhaps the biggest winners of Generative AI in 2023 were the Accentures of the world, which reportedly generated $2B in fees for AI consulting.

Big winners of the AI craze so far: consultants.

— Matt Turck (@mattturck) May 12, 2023

Regardless, there’s tremendous hope that 2024 is going to be a big year for AI in the enterprise – or at least for Generative AI, as traditional AI already has a significant footprint there already (see above).

But we’re early in answering some of the key questions Global 2000-type companies face:

What are the use cases? The low hanging fruit use cases so far have been mostly a) code generation co-pilots for developer teams, b) enterprise knowledge management (search, text summarization, translation, etc), and c) AI chatbots for customer service (a use case that pre-dates Generative AI). There are certainly others (marketing, automated SDRs etc) but there’s a lot to figure out (co-pilot mode vs full automation etc).

What tools should we pick? As per the above, it feels like the future is hybrid, a combination of commercial vendors and open source, big and small models, horizontal and vertical GenAI tools. But where does one start?

Who will be deploying and maintaining the tools? There is a clear skill shortage in Global 2000 companies. If you thought recruiting software developers was hard, just try to recruit machine learning engineers.

How do we make sure they don’t hallucinate? Yes there’s a tremendous amount of work being done around RAG and guardrails and evaluations etc, but the possibility that a Generative AI tool may be plain wrong, and the broader question that we don’t really know how Generative AI models work, are big problems in the enterprise.

What is the ROI? Large tech companies have been early in leveraging Generative AI for their own needs, and they’re showing interesting early data. In their earnings call, Palo Alto Networks mentioned roughly halving the cost of their T&E servicing, and ServiceNow mentioned increasing our developer innovation speed by 52%, but we’re early in understanding the cost / return equation for Generative AI in the enterprise.

The good news for Generative AI vendors is that there’s plenty of interest from enterprise customers to allocate budget (importantly, no longer “innovation” budgets but actual OpEx budget, possibly re-allocated from other places) and resources to figuring it out. But we’re probably talking about a 3-5 year deployment cycle, rather than one.

WATCH: MAD Podcast with Florian Douetteau, CEO, Dataiku

21.Is AI going to kill SaaS? AI会终结SaaS吗？

这是过去12个月里流行的想法之一。

问题的一个版本是：AI使得编码速度提高了10倍，所以仅凭几个普通开发者，你就能够创建一个定制版的SaaS产品，专门针对你的需求。当你可以自己构建时，为什么要向SaaS提供商支付大量资金。

另一个版本的问题是：未来将由一个AI智能（可能由几个模型组成）运行你的整个公司，通过一系列代理。你不再购买人力资源软件、财务软件或销售软件，因为AI智能能够以完全自动化和无缝的方式完成所有事务。

我们似乎还远未真正以任何成熟的方式实现这两种趋势，但正如我们所知，AI领域的变化非常快。

与此同时，未来的一个可能版本似乎是，随着AI被集成到每一个SaaS产品中，SaaS产品将变得更加强大。

This was one of the trendy ideas of the last 12 months.

One version of the question: AI makes it 10x to code, so with just a few average developers, you’ll be able to create a custom-made version of a SaaS product, tailored to your needs. Why pay a lot of money to a SaaS provider when you can build your own.

Another version of the question: the future is one AI intelligence (possibly made of several models) that runs your whole company with a series of agents. You no longer buy HR software, finance software or sales software because the AI intelligence does everything, in a fully automated and seamless way.

We seem to be somewhat far away from both of those trends actually happening in any kind of full-fledged manner, but as we all know, things change very fast in AI.

In the meantime, it feels like a likely version of the future is that SaaS products are going to become more powerful as AI gets built into every one of them.

22.Is AI going to kill venture capital ? AI会终结风险资本吗？

撇开AI能否在公司选择和投资后增值方面自动化风险资本的（永远有趣的）话题不谈，围绕资产类别是否为AI平台转变正确定位存在一系列有趣的问题：

风险资本规模是否太小？像OpenAI这样的公司需要筹集数十亿美元，并且可能还需要筹集更多亿万资金。这些资金中有很大一部分是由像微软这样的大公司提供的——很大程度上可能是以计算力换股权的形式。当然，许多风投公司已经投资于大型基础模型公司，但至少，这些投资明显偏离了传统的VC软件投资模型。也许AI投资将需要超大型的VC基金——在撰写本文时，沙特阿拉伯似乎正准备与美国风投公司合作推出一个400亿美元的AI基金。

风险资本规模是否太大？如果你相信AI将使我们的生产力提高10倍，包括超级编码者、自动化的SDR代理和自动化的营销创造，那么我们即将见证一整代由骨干团队（或许只有一名独立创业者）运营的全自动化公司的诞生，这些公司理论上能够实现数亿美元的收入（并上市）？

一家由独立创业者运营、年收入达1亿美元的公司在其成长过程中的任何时点需要风险资本吗？

Leaving aside the (ever-amusing) topic of whether AI could automate venture capital, both in terms of company selection, and post-investment value-add, there’s an interesting series of questions around whether the asset class is correctly-sized for the AI platform shift:

Is Venture Capital too small? The OpenAIs of the world have needed to raise billions of dollars, and may need to raise many more billions. A lot of those billions have been provided by big corporations like Microsoft – probably in large part in the form of compute-for-equity deals. Of course, many VCs have invested in big foundational model companies, but at a minimum, those investments are a clear departure from the traditional VC software investing model. Perhaps AI investing is going to require mega-sized VC funds – at the time of writing, Saudi Arabia seems to be about to launch a $40B AI fund in collaboration with US VC firms.

Is Venture Capital too big? If you believe that AI is going to 10x our productivity, including super coders and automated SDR agents and automated marketing creation, then we’re about to witness the birth of a whole generation of fully-automated companies run by skeleton teams (or maybe just one solo-preneur) that could theoretically reach hundreds of millions in revenues (and go public)?

Does a $100M ARR company run by a solo-preneur need venture capital at any point in its journey?

23.Will AI revive consumer?

消费者自社交媒体和移动时代以来一直在寻找下一个机遇。生成式AI很可能就是这样一个机遇。

一些有趣的领域（在众多其他领域中）：

• 搜索：几十年来，谷歌的搜索垄断首次出现了早期但可信的竞争者。一小撮初创公司如Perplexity AI和You.com正在引领从搜索引擎到答案引擎的演变。

• AI伴侣：超越反乌托邦的方面，如果每个人都有一个无限耐心且乐于助人的伴侣，能够满足一个人的特定需求，无论是知识、娱乐还是治疗。

• AI硬件：Humane、Rabbit、VisionPro在消费者硬件中是激动人心的新进入者。

• 超个性化娱乐：随着生成式AI驱动的工具不断变得更好（和更便宜），我们将发明什么新形式的娱乐和艺术？

• 观影体验：

2005年：去电影院

2015年：流式传输Netflix

2025年：请求 LLM + 文本到视频，为今晚创造一个新的《毒枭》季节，但让它发生在叙利亚，由布拉德·皮特、Mr Beast和特拉维斯·凯尔斯主演

— Matt Turck (@mattturck) 2024年2月15日

Consumer has been looking for its next wind since the social media and mobile days. Generative AI may very well be it.

Some interesting areas (among many others):

Search: for the first time in decades, Google’s search monopoly has some early, but credible competitors. A handful of startups like Perplexity AI and You.com are leading the evolution from search engine to answer engine.

AI companions: beyond the dystopian aspects, what if every human had an infinitely patient and helpful companion attuned to one’s specific needs, whether for knowledge, entertainment or therapy

AI hardware: Humane, Rabbit, VisionPro are exciting entries in consumer hardware

Hyper-personalized entertainment: what new forms of entertainment and art will we invent as Generative AI powered tools keep getting better (and cheaper)?

Movie watching experience: 2005: go to a movie theater 2015: stream Netflix 2025: ask LLM + text-to-video to create a new season of Narcos to watch tonight, but have it take place in Syria with Brad Pitt, Mr Beast and Travis Kelce in the leading roles

— Matt Turck (@mattturck) February 15, 2024

Watch:

• MAD Podcast with Aravind Srinivas, CEO, Perplexity AI (Apple, Spotify)

• MAD Podcast with Richard Socher, CEO, You.com (Apple, Spotify)

• MAD Podcast with Cris Valenzuela, CEO, Runway (Apple, Spotify)

24.AI and blockchain : BS , or exciting? AI与区块链：胡说八道，还是令人兴奋？

我知道，我知道。AI与加密货币的交叉感觉像是X/Twitter笑话的完美素材。

然而，不可否认的是，AI正在集中到少数几家拥有最多计算资源、数据和AI人才的公司手中——从大型科技公司到那个著名的“并非开放”的OpenAI。与此同时，区块链提议的核心正是使创建去中心化网络成为可能，允许参与者共享资源和资产。那里有探索的肥沃土壤，这是我们几年前开始探讨的话题（演讲）。

一些与AI相关的加密项目经历了明显的加速，包括Bittensor*（去中心化机器智能平台）、Render（去中心化GPU渲染平台）、Arweave（去中心化数据平台）。

虽然我们今年的MAD生态图中没有包含加密货币部分，这是一个值得关注的领域。

现在，一如既往的问题是，加密货币行业能否自我帮助，而不是沦为数百个与AI相关的迷因币、抽水排放计划和骗局。

I know, I know. The intersection of AI and crypto feels like perfect fodder for X/Twitter jokes.

However, it is an undeniable concern that AI is getting centralized in a handful of companies that have the most compute, data and AI talent – from Big Tech to the famously-not-open OpenAI. Meanwhile, the very core of the blockchain proposition is to enable the creation of decentralized networks that allow participants to share resources and assets. There is fertile ground for exploration there, a topic we started exploring years ago (presentation).

A number of AI-related crypto projects have experienced noticeable acceleration, including Bittensor* (decentralized machine intelligence platform), Render (decentralized GPU rendering platform), Arweave (decentralized data platform).

While we didn’t include a crypto section in this year’s MAD Landscape, this is an interesting area to watch.

Now, as always, the question is whether the crypto industry will be able to help itself, and not devolve into hundreds of AI-related memecoins, pump-and-dump schemes and scams.

PART III: FINANCINGS, M&A AND IPOS 第三部分：融资、并购及IPO

融资

当前的融资环境是“两极市场”的情形之一，有A类市场，还有其他。

总体资金继续下滑，2023年下降42%至2484亿美元。2024年前几个月显示了一些可能的复苏迹象，但到目前为止趋势或多或少保持不变。

数据基础设施，出于上述所有原因，获得的资金活动非常少，Sigma Computing和Databricks是少数例外。

显然，AI则是另一个故事。

AI融资市场不可避免的特点有：

• 大量资本集中在少数几家初创公司，特别是OpenAI、Anthropic、Inflection AI、Mistral等。
• 来自大公司投资者的活动程度不成比例。2023年最活跃的3家AI投资者是微软、谷歌和英伟达。
• 上述大公司交易中有些关于实际现金与“计算力换股权”金额的不明确之处。

自2023年MAD以来一些值得注意的交易（按大致时间顺序，非详尽列表）：

OpenAI，（或许是）基础模型开发商，跨两轮筹集了103亿美元，目前估值860亿美元；Adept，另一家基础模型开发商，筹集了3.5亿美元，估值10亿美元；AlphaSense，金融服务市场研究平台，在两轮中筹集了4.75亿美元，现估值25亿美元；Anthropic，又一家基础模型开发商，三轮共筹集了64.5亿美元，估值184亿美元；Pinecone，向量数据库平台，筹集了1亿美元，估值7.5亿美元；Celestial AI，光互连技术平台，两轮共筹集了2.75亿美元；CoreWeave，GPU云提供商，筹集了4.21亿美元，估值25亿美元；Lightmatter，光驱动芯片开发商，两轮共筹集了3.08亿美元，现估值12亿美元；Sigma Computing，云托管数据分析平台，筹集了3.4亿美元，估值11亿美元；Inflection，又一基础模型开发商，筹集了13亿美元，估值40亿美元；Mistral，基础模型开发商，两轮共筹集了5.28亿美元，现估值20亿美元；Cohere，（意外）基础模型开发商，筹集了2.7亿美元，估值20亿美元；Runway，生成视频模型开发商，筹集了1.91亿美元，估值15亿美元；Synthesia*，企业视频生成平台，筹集了9000万美元，估值10亿美元；Hugging Face，机器学习和数据科学平台，筹集了2.35亿美元，估值45亿美元；Poolside，专注于代码生成和软件开发的基础模型开发商，筹集了1.26亿美元；Modular，AI开发平台，筹集了1亿美元，估值6亿美元；Imbue，AI代理开发商，筹集了2.12亿美元；Databricks，数据、分析和AI解决方案提供商，筹集了6.84亿美元，估值432亿美元；Aleph Alpha，又一基础模型开发商，筹集了4.86亿美元；AI21 Labs，基础模型开发商，筹集了2.08亿美元，估值14亿美元；Together，生成式AI开发的云平台，两轮共筹集了2.085亿美元，现估值12.5亿美元；VAST Data，深度学习的数据平台，筹集了1.18亿美元，估值91亿美元；Shield AI，航空航天和国防行业的AI驾驶员开发商，筹集了5亿美元，估值28亿美元；01.ai，基础模型开发商，筹集了2亿美元，估值10亿美元；Hadrian，航空航天和国防的精密组件工厂制造商，筹集了1.17亿美元；Sierra AI，客户服务/体验的AI聊天机器人开发商，两轮共筹集了1.1亿美元；Glean，AI驱动的企业搜索平台，筹集了2亿美元，估值22亿美元；Lambda Labs，GPU云提供商，筹集了3.2亿美元，估值15亿美元；Magic，代码生成和软件开发的基础模型开发商，筹集了1.17亿美元，估值5亿美元。

并购、私有化

自2023年MAD以来，并购市场相对平静。

许多传统软件收购者专注于自身股价和整体业务，而不是积极寻找收购机会。

特别严格的反垄断环境使潜在收购者的情况变得更加棘手——导致交易减少和一些像Inflection AI被微软收购这样的杂技般的操作。

私募股权公司在艰难的市场中寻找低价机会保持相对活跃。

涉及MAD生态图中多年出现的公司的一些值得注意的交易（按规模顺序）：

半导体制造商博通收购云计算公司VMWare，价值690亿美元；网络和安全基础设施公司思科收购监控和可观察性平台Splunk，价值280亿美元；客户体验管理公司Qualtrics被Silver Lake和CPP投资私有化，价值125亿美元；支出管理平台Coupa被Thoma Bravo私有化，价值80亿美元；监控和可观察性平台New Relic被Francisco Partners和TPG收购，价值65亿美元；数据分析平台Alteryx被Clearlake Capital和Insight Partners私有化，价值44亿美元；收入编排平台Salesloft被Vista Equity收购，价值23亿美元，随后Vista还收购了客户体验的AI聊天机器人开发商Drift；数据湖仓库提供商Databricks收购AI开发平台MosaicML，价值13亿美元（以及其他公司，如Arcion和Okera，金额较低）；数据分析平台Thoughtspot收购商业智能初创公司Mode Analytics，价值2亿美元；数据仓库提供商Snowflake收购消费者AI搜索引擎Neeva，价值1.5亿美元；云托管提供商DigitalOcean收购云计算和AI开发初创公司Paperspace，价值1.11亿美元；云计算芯片制造商英伟达收购AI/ML优化平台OmniML，用于边缘计算。

IPO？

在公共市场，AI是一个热门趋势。"Magnificent Seven"股票（英伟达、Meta、亚马逊、微软、Alphabet、苹果和特斯拉）在2023年至少增长了49%，推动了整个股市的上涨。

总的来说，公共市场上纯粹的AI股票仍然非常缺乏。可用的少数几个受到了丰厚的回报——Palantir股票在2023年跳涨了167%。

这应该为一大批与AI相关的即将上市的初创公司预示着好兆头。MAD空间中有许多公司在规模上达到了显著的水平——首先是Databricks，但也包括Celonis、Scale AI、Dataiku*或Fivetran等其他几家。

然后是OpenAI和Anthropic如何看待公共市场的有趣问题。

与此同时，2023年在IPO方面是非常糟糕的一年。只有少数与MAD相关的公司上市：营销自动化平台Klaviyo于2023年9月以92亿美元的估值上市（参见我们对Klaviyo S-1的拆解）；论坛式社交网络平台Reddit（为AI玩家提供其内容的许可），于2024年3月以64亿美元的估值上市；为AI和云基础设施提供智能连接的半导体公司Astera Labs，于2024年3月以55亿美元的估值上市。

Financings

The current financing environment is one of the “tale of two markets” situations, where there’s A, and everything else.

The overall funding continued to falter, declining 42% to $248.4B in 2023. The first few months of 2024 are showing some possible green shoots, but as of now the trend has been more or less the same.

Data infrastructure, for all the reasons described above, saw very little funding activity, with Sigma Computing and Databricks being some of the rare exceptions.

Obviously, AI was a whole different story.

The inescapable characteristics of the AI funding market have been:

• A large concentration of capital in a handful of startups, in particular OpenAI, Anthropic, Inflection AI, Mistral, etc.
• A disproportionate level of activity from corporate investors. The 3 most active AI investors in 2023 were Microsoft, Google and NVIDIA
• Some murkiness in the above corporate deals about what amount is actual cash, vs “compute for equity”

Some noteworthy deals since our 2023 MAD, in rough chronological order (not an exhaustive list!):

OpenAI, a (or the?) foundational model developer, raised $10.3B across two rounds, now valued at$86B; Adept, another foundational model developer, raised $350M at a$1B valuation; AlphaSense, a market research platform for financial services, raised $475M across two rounds, now valued at$2.5B, Anthropic, yet another foundational model developer, raised $6.45B over three rounds, at a$18.4B valuation; Pinecone, a vector database platform, raised $100M at a$750M valuation; Celestial AI, an optical interconnect technology platform for memory and compute, raised $275M across two rounds; **CoreWeave**, a GPU Cloud provider, raised$421M at a $2.5B valuation; **Lightmatter**, developer of a light-powered chip for computing, raised$308M across two rounds, now valued at $1.2B; **Sigma** **Computing**, a cloud-hosted data analytics platform, raised$340M at a $1.1B valuation; **Inflection**, another foundational model developer, raised$1.3B at a $4B valuation; **Mistral**, a foundational model developer, raised$528M across two rounds, now valued at $2B; **Cohere**, (surprise) a foundational model developer, raised$270M at a $2B valuation; **Runway**, a generative video model developer, raised$191M at a $1.5B valuation; **Synthesia*** , a video generation platform for enterprise, raised$90M at a $1B valuation; **Hugging Face**, a machine learning and data science platform for working with open source models, raised$235M at a $4.5B valuation; **Poolside**, a foundational model developer specifically for code generation and software development, raised$126M; Modular, an AI development platform, raised $100M at a$600M valuation; Imbue, an AI agent developer, raised $212M; **Databricks**, provider of data, analytics and AI solutions, raised$684M at a $43.2B valuation; **Aleph Alpha**, another foundational model developer, raised$486M; AI21 Labs, a foundational model developer, raised $208M at a$1.4B valuation; Together, a cloud platform for generative AI development, raised $208.5M across two rounds, now valued at$1.25B; VAST Data, a data platform for deep learning, raised $118M at a$9.1B valuation; Shield AI, an AI pilot developer for the aerospace and defense industry, raised $500M at a$2.8B valuation; 01.ai, a foundational model developer, raised $200M at a$1B valuation; Hadrian, a manufacturer of precision component factories for aerospace and defense, raised $117M; **Sierra AI**, an AI chatbot developer for customer service / experience, raised$110M across two rounds; Glean, an AI-powered enterprise search platform, raised $200M at a$2.2B valuation; Lambda Labs, a GPU Cloud provider, raised $320M at a$1.5B valuation; Magic, a foundational model developer for code generation and software development, raised $117M at a$500M valuation.

M&A , Take Privates

The M&A market has been fairly quiet since the 2023 MAD.

A lot of traditional software acquirers were focused on their own stock price and overall business, rather than actively looking for acquisition opportunities.

And the particularly strict antitrust environment has made things trickier for potential acquirers – leading to less deals and some contortionist exercises like the Inflection AI acquisition by Microsoft.

Private equity firms have been reasonably active, seeking lower price opportunities in the tougher market.

Some noteworthy transactions involving companies that have appeared over the years on the MAD landscape (in order of scale):

Broadcom, a semiconductor manufacturer, acquired VMWare, a cloud computing company, for $69B; **Cisco**, a networking and security infrastructure company, acquired **Splunk**, a monitoring and observability platform, for$28B; Qualtrics, a customer experience management company, was taken private by Silver Lake and CPP Investments for $12.5B; **Coupa**, a spend management platform, was taken private by **Thoma Bravo** for$8B; New Relic, a monitoring and observability platform, was acquired by Francisco Partners and TPG for $6.5B; **Alteryx**, a data analytics platform, was taken private by **Clearlake Capital and Insight Partners** for$4.4B; Salesloft, a revenue orchestration platform, was acquired by Vista Equity for $2.3B, which then also acquired **Drift**, an AI chatbot developer for customer experience; **Databricks**, a provider of data lakehouses, acquired **MosaicML**, an AI development platform, for$1.3B (and several other companies, for lower amounts like Arcion and Okera); Thoughtspot, a data analytics platform, acquired Mode Analytics, a business intelligence startup, for $200M; **Snowflake**, a provider of data warehouses, acquired **Neeva**, a consumer AI search engine, for$150M; DigitalOcean, a cloud hosting provider, acquired Paperspace, a cloud computing and AI development startup, for $111M; NVIDIA, a chip manufacturer for cloud computing, acquired OmniML, an AI/ML optimization platform for the edge.

IPOs ?

In public markets, AI has been a hot trend. The “Magnificent Seven” stocks (Nvidia, Meta, Amazon, Microsoft, Alphabet, Apple and Tesla) gained at least 49% in 2023 and powered the overall stock market higher.

Overall, there is still a severe dearth of pure-play AI stocks in public markets. The few that are available are richly rewarded – Palantir stock jumped 167% in 2023.

This should bode well for a whole group of AI-related pre-IPO startups. There are a lot of companies at significant amounts of scale in the MAD space – first and foremost Databricks, but also a number of others including Celonis, Scale AI, Dataiku* or Fivetran.

Then there’s the intriguing question of how OpenAI and Anthropic will think about public markets.

In the meantime, 2023 was a very poor year in terms of IPOs. Only a handful of MAD-related companies went public: Klaviyo, a marketing automation platform, went public at a $9.2B valuation in September 2023 (see our [Klaviyo S-1teardown](https://mattturck.com/klaviyos1/)); **Reddit**, a forum-style social networking platform (which licenses its content to AI players) , went public at a$6.4B valuation in March 2024; Astera Labs, a semiconductor company providing intelligent connectivity for AI and cloud infrastructure, went public at a $5.5B valuation in March 2024.