Chris Lattner, the father of LLVM: Why we need to rebuild AI infrastructure software

The vision of AI that people once imagined was very beautiful, but the current situation is not satisfactory. AI has not realized its original predictions in daily applications such as autonomous driving and new drug research and development. A common complaint is that the global technology giants have brought together a large number of the smartest brains, but they still focus more on precise advertising and credit scoring. And on "smart" speakers that are not very smart.

Theoretically, as long as there are correct algorithms and sufficient computing resources, AI can solve all problems represented by any available data. Now that data, algorithms and hardware resources are abundant enough, AI can be used to benefit society. All conditions are met. We have seen the broad application and initial effects of AI, but in fact, the technology is not applied in depth, and it is far from realizing the full potential of existing machine learning research.

Why does this happen? The facts are more profound than the latest research updates on AI by the world’s technology giants and media. ​​Compiler expert Chris Lattner once pointed out that the singleness and fragmentation of AI systems and tools are the root cause of this problem. In order to solve this problem, in January 2022, compiler expert Chris Lattner announced that he would start a business overseas and co-founded Modular AI with Tim Davis. The goal is to rebuild the global ML infrastructure, including compilers and runtimes. , pay equal attention to heterogeneous computing, edge to data center, and focus on availability to improve developer efficiency. Currently, the Modular AI team has participated in building most of the world's production machine learning infrastructure from TensorFlow, TF Lite, XLA, TPU, Android ML, Apple ML, MLIR, etc., and has deployed production workloads to billions of users and equipment.

Recently, Modular AI announced the completion of a US$30 million seed round of financing, led by Google Venture. In the latest official blog post published by Chris Lattner and others, they issued "Three Questions for the Soul": AI is so important, why is the software so miserable? Why haven’t tech giants solved the AI ​​puzzle? How to solve this problem? Of course, they also gave answers. The OneFlow community compiled and organized the original text.

1 AI is so important, why is the software so miserable?

AI software was originally designed for full-stack researchers, engineers, and architects building AI technology. It was never defined as a product. Therefore,

AI software has defect.

This kind of software is built by big tech companies to solve their own problems, and other enterprises use it on a "trickle down" infrastructure. The result is that only the largest and most commercially impactful AI applications are built and deployed in practice, and even then only if the needs of the enterprise are highly aligned with the internal needs of large technology companies. accomplish.

Why is this? Because the current AI software is very simple and has heavy research attributes, it is mainly used to meet the development plans of technology giants (the developers of these software). These softwares were created by researchers to do research, and the rapid development of AI leaves researchers no time to stop and rebuild.

Instead, over time we have added more and more complexity, making it difficult for the industry to maintain and scale fragmented custom tool chains that are used in research and production, training and There are differences between deployments, servers and edges.

Artificial intelligence systems have now become a vast ocean of incompatible technologies, and only those comprehensive technology giants have the ability to use AI to achieve their goals.

2 Why haven’t technology giants solved AI problems?

AI research and developers work together to make deploying AI a success, and tech giants use their vast computing and financial resources to advance their products and core business priorities, including their own clouds, phones, Social networking and artificial intelligence hardware.

Although they have made outstanding contributions to the field, from a business perspective, it is impossible for them to promote AI to the whole world (covering all hardware, cloud and ML frameworks), and the rest of the world cannot Expect them to do so. However, this unfortunate fact limits the rest of the world’s ability to use this technology to solve problems outside of the focus areas of big tech companies, including some of the biggest socioeconomic and environmental issues facing the world. But this is not the future we want.

Although giants have made huge contributions to the development of artificial intelligence, for artificial intelligence to fully realize its potential, an independent company is needed. This company does not have to prioritize its own hardware, cloud infrastructure, Mobile phone development or own research; at the same time, we need a neutral company to do what is best for the interests of global users and businesses. We need to incorporate what we learn from the rapid growth of AI software into next-generation technologies to provide usable solutions and common standards for the types of problems faced by all organizations.

Today, the most pressing issue facing small and medium-sized technology companies is how to break through the limitations of capacity, cost, time and talent to put AI into production.

Due to opportunity cost considerations, it is difficult for their innovative technologies to be promoted to the market, and the product experience is poor, which will ultimately have a negative impact on their development. For society as a whole, this means that we still have to wait quite a while before AI can solve some of the world's biggest problems.

We don’t have time to wait for the tech giants to roll out trickle-down AI software. AI can change the world, but only if fragmentation is resolved and the global AI developer community doesn’t have to struggle with high-quality infrastructure.

3 Who will solve this problem? How to solve?

Modular is building a next-generation AI developer platform that will be more practical, faster and more flexible.

Our platform unifies the front-ends of popular AI frameworks through common interfaces, and enhances access and portability to various hardware back-ends and cloud environments. We're rebuilding our core developer workflow tools to be more expressive, usable, debuggable, reliable, scalable, and performant. Our tools can be easily deployed into existing workflows, allowing users to seamlessly complete their work without refactoring or rewriting code, and achieve improvements in productivity and performance at a lower cost. We will accelerate the exploration of the value of AI and bring it to the market as soon as possible to benefit the majority of users.

When AI can penetrate into various applications in a more subtle way, its potential will be fully demonstrated - by then, you will no longer have to define your application around AI. Our platform is built from modular, composable infrastructure components that can be rearranged and extended to implement a variety of use cases. At the same time, experts in various fields can innovate through our platform even without understanding how the entire system works. We’ve seen firsthand how a modular approach can unlock new use cases that we hadn’t thought of in the past.

In order to truly repair AI infrastructure, we must not only solve "hard technical" problems (such as compilers for heterogeneous computing technologies), but also establish a seamless end-to-end developer workflow.

4 From the "AI Research Era" to the "AI Production Era"

Our success means that developers around the world will have access to truly usable, portable and scalable AI software.

In the new world, developers who lack sufficient budgets or top talent can also work as efficiently as global technology giants; the efficiency and total cost of ownership (TCO) of AI hardware will be improved Optimization; enterprises can easily plug in custom ASICs to suit their use cases; deploy to the edge as easily as deploying to servers; enterprises can use whichever AI framework best suits their needs; AI programs can seamlessly scale on hardware, Deploying the latest AI research into production couldn't be easier.

We will see: the development of the AI ​​industry is no longer limited by the timetable determined by the technology giants according to their own needs; the development of the AI ​​industry will be faster and more concentrated; innovation will be at all levels of the stack Booming, with developers focused on bringing new innovations to market in their areas of expertise and building a better future for all of us; the industry is developing at a rapid pace, leading us from the "AI Research Era" to the "AI Production Era" ".

The above is the detailed content of Chris Lattner, the father of LLVM: Why we need to rebuild AI infrastructure software. For more information, please follow other related articles on the PHP Chinese website!