[260905] Most AI Tools Die in Months. Which Ones Survive?

TLDR

• AI tools endure by decoupling complex capabilities into simple, modular building blocks that solve lasting problems and serve as infrastructure for future products.

In August 2019, Google open-sourced two machine learning projects that seemed entirely unrelated. One was the speech transcription technology behind Live Transcribe, which initially served deaf and hard-of-hearing users by allowing phones to transcribe sound into text in real time. The other was MediaPipe, aimed at developers, providing computer vision capabilities such as gesture recognition and human pose estimation to help mobile devices process camera information more easily.

One solved accessibility problems, while the other served mobile developers. But looking back seven years later, these two projects shared a common trait: they did not just build a complete application; instead, they opened up a portion of their AI capabilities, enabling more people to continue building.

This raises a question worth pondering: why do some AI tools persist through shifting waves while others disappear so quickly?

From a Feature to a Capability #

Software competition in the past often revolved around complete products. Users opened Photoshop to edit photos, or Zoom to hold meetings—the product itself was the center of value.

However, AI brings a change: more and more complex capabilities are being decoupled from applications. For instance, speech recognition, which used to be merely a feature within a specific app, can now enter various scenarios such as meeting notes, real-time translation, customer service systems, and AI assistants. The same goes for visual understanding; whereas an app previously had to solve image recognition on its own, developers can now directly utilize mature vision capabilities and combine them into their own products.

This is where the value of Live Transcribe and MediaPipe lies. They do not persist long-term because of a single specific feature, but because they package complex capabilities and lower the cost for others to build products.

A Pattern Preceding AI #

Similar things have happened many times in the software industry. Take Stripe in the payments sector: it didn't try to become an e-commerce platform for all merchants, but instead turned payment capabilities into APIs that developers could call. In the past, if a team wanted to integrate payments, they had to handle complex issues such as banking systems, payment workflows, and security compliance. Stripe packaged these complex parts so developers could add payments to their products much faster.

Ultimately, many users don't even know they are using Stripe, yet a vast number of internet products rely on it to complete transactions. The value of such tools lies not in users opening them every day, but in becoming a part of other products.

Even More Pronounced in the AI Era #

In the AI era, the pace of change is faster than in previous software cycles. An image generation product might look stunning today, but could be matched by new model capabilities a few months later; an automated writing tool might be highly valuable today, but could be replaced by a more generic assistant in the future. This is because they primarily provide an end result, whereas the lifecycle of underlying capabilities is typically much longer.

For example, understanding speech, processing visual information, performing reasoning, and having machines execute tasks—these capabilities are not tied to any single application, and they can still be recombined when new product forms emerge. This is why competition among AI tools is not just about who offers more features today, but who can solidify capabilities so that more people can continue to use them.

From Tools to Underlying Infrastructure #

Of course, not all AI tools will end up becoming infrastructure. Many excellent applications will continue to serve users directly, but a tool is more likely to possess long-term value if it has a few characteristics: it solves a long-standing problem rather than a short-term hype, it lowers the barrier for others to use AI rather than just displaying a one-off effect, and it can be integrated, combined, and introduced into more workflows.

Live Transcribe and MediaPipe belong to this category. They initially solved two specific problems: letting machines understand sound, and letting devices comprehend visual information. But the lasting value they left behind was turning these capabilities into modules that others could continue to build upon.

Seven years ago, when Google open-sourced them, it did not predict today's AI Agents, nor was it aiming to become some future gateway. It simply opened up complex capabilities. And many long-lasting tools start right here: when a capability is versatile and easy enough to use, it can gradually evolve from a tool into a foundational layer that the next generation of products relies on.

Please indicate the source and link of this article when reprinting.

Help us make these docs great!

All X-CMD docs are generated from command help and multiple data sources. See something that's wrong or unclear? Feel free to let us know through any of these ways~