Shulou(Shulou.com)06/02 Report--

Lai can compile

Qubit report | official account QbitAI

What difficulties will be encountered in the landing of AI technology? The gross profit margin is low, it is difficult to have a general solution, and it is difficult to form technical barriers.

These challenges are analyzed one by one in an article by Silicon Valley's top VC companies. According to the existing experience, some suggestions are provided.

This article, from A16Z, has become one of Silicon Valley's top VC companies with investments in Instagram, Twitter and Skype. In 2019, it announced the transformation into an investment advisory firm and made the field of artificial intelligence one of its five ongoing concerns.

According to this article, AI is essentially a new type of business. Here are the main things that qubits compile for you:

New business type: software + service = AI?

Because of its firm belief that AI can drive business transformation, A16Z will continue to invest heavily in AI applications and AI infrastructure. However, in many cases, AI's economic structure is fundamentally different from that of its software business. It even looks more like a traditional service company.

The AI industry is a bit like the combination of traditional software industry and service industry. AI applications, like ordinary software, can be sold many times, but each time it requires a large number of professionals to provide services. Therefore, AI essentially creates a new type of business.

Challenge 1: it is difficult to reduce the cost of arithmetic and labor

In AI's financial data, A16Z notes an unexpectedly consistent pattern: gross margins are usually between 50 and 60 per cent. It is well below the gross margin of 60-80 per cent for SaaS.

Early private capital can cover up these inefficiencies in the short term. However, can long-term product or market optimization completely solve the problem? This is not clear.

Overall, there are two reasons for AI's low gross profit margin:

1. Infrastructure brings huge costs

AI companies often spend 25% or more of their revenue on cloud resources.

Training an AI model may require hundreds of thousands of dollars or more of computing resources. And because the data that feeds the AI model will change over time (that is, data drift), retraining will also bring continuous costs.

Model inference is also more complex in calculation. AI often involves data such as image, audio or video, which requires higher storage resources and processing costs.

For some companies, AI models have to be transferred between regional clouds, making cloud computing operations more complex and costly.

In many areas, multiplying processing and data are needed to ensure improved accuracy. As a result, the complexity of the model is growing rapidly, and the processor is unlikely to keep up. Distributed computing can solve the speed problem, but not the cost problem.

2. In order to achieve high accuracy, manual work is indispensable in the cycle process.

Training most advanced AI models requires manual cleaning and tagging of large datasets. This process is labor-intensive and expensive, and has become one of the biggest obstacles to the widespread adoption of AI.

After the model is deployed, in order to maintain the accuracy, it is necessary to constantly capture, mark and feedback the new training data into the system. Many companies spend as much as 10-15% of their revenue in the process.

For tasks that require more cognitive reasoning, manual work needs to be embedded in the AI system in real time.

For example, social media companies employ thousands of human auditors to enhance their AI-based audit systems. Many self-driving car systems have remote operators, and most AI medical devices interact with doctors as co-decision makers.

With the improvement of the performance of the AI model, the need for human intervention may decrease, but it cannot be completely eliminated.

Because of the necessity of labor, many AI companies that plan to sell pure software products begin to introduce internal service functions and cost budgets.

Challenge 2: it is difficult to have a universal solution

It is more difficult for AI to know when to find the right product for the market than the traditional software industry. Startups also often spend more time and resources than expected to deploy their products.

Customers may enter anything into the AI model, so new data may be needed when deploying each new customer.

Even customers with similar appearance-for example, both carmakers are conducting defect testing-may need substantially different training data.

One founder called this the "time cost" of AI products. Her company conducts special data collection and model fine-tuning at the beginning of each new customer engagement. The company looks at the distribution of customer data to eliminate some of the extremes before deployment.

This also brings new human and financial costs: the company's team and financial resources are tied together until the accuracy of the model is acceptable.

The timing of the training period is usually unknown. Because, no matter how hard you try, there is often no choice to generate training data more quickly.

Not only is the process of deploying the product time-consuming, it doesn't necessarily disappear over time.

Identifying requirements in advance is difficult because AI models need to take into account edge situations, while traditional prototyping tools (such as models, prototypes, or Beta tests) tend to cover only the most common paths.

Challenge 3: technological differentiation is difficult to achieve

Superior technology can bring differentiated competitive advantage to the company. But this is more difficult for AI to achieve.

The new model architecture is mainly developed in the open academic environment, and the pre-training model can be obtained from the open source library, and the model parameters can be optimized automatically. Data is the core of AI system, but it is in the hands of customers, either in the public domain, or as a commodity over time.

With the maturity of the market, the value of data gradually decreases, and the network effect is also relatively weak. As the model becomes mature, the cost of solving each new marginal case becomes higher and higher, and the value provided is less and less.

This does not necessarily mean that AI products are less defensive than pure software products. However, for AI, the competitive advantage seems to be less than many expected.

Advice to the founder of AI

To achieve long-term success, AI artificial intelligence companies need to make full use of the advantages of services and software. A16Z gives the following recommendations:

Eliminate model complexity

A16Z notes that there is a huge difference in the cost of sales between startups that train a unique model per customer and startups that share a unified model among all customers.

The "single model" strategy is easier to maintain, rolls out to new customers faster, and supports simpler, more efficient engineering organizations.

One key is to learn as much as possible about customers and customer data before closing a deal. Sometimes, new customers place a huge burden on ML projects, and in most cases only a few unique models or fine-tuning are involved.

Weighing long-term economic health against recent growth is one of the most important tasks facing AI founders.

Pick problems in narrow areas and reduce data complexity

It is difficult to automate manual labor. AI, a company in the field of CRM, has found that valuable areas can be found for AI based on updated records.

It involves large-scale, low-complexity tasks. For example, auditing, data entry / coding, transcription and other issues are difficult for humans to implement, while AI is relatively easy.

Focusing on these areas can simplify the data feed of the development process.

Planned high variable cost

As a founder, you should provide a reliable and intuitive thinking framework for the business model. A16Z recommends the establishment of a business model and a GTM strategy, taking into account lower gross margins.

There are some suggestions from the founders: deeply understand the distribution of model data; make model maintenance and personnel failover a top priority; track and measure actual variable costs so that they are not hidden in research and development; make conservative unit economic assumptions in financial models, especially during fund-raising; don't wait for scale expansion or external technological advances to solve the problem.

Plan changes in the technology stack

Tightly coupling the application to the way it works today may lead to architectural disadvantages in the future.

Modern AI is still in its infancy. Over the next few years, tools are likely to be widely used to automate model training, improve reasoning efficiency, standardize developer workflows, and monitor and protect production AI models. As a cost issue, cloud computing has attracted more and more attention.

Embrace the service.

This may mean providing full-stack translation services rather than translation software and taxi services rather than selling self-driving cars.

Building a hybrid business is more difficult than pure software, but this approach can provide deep insight into customer needs and generate companies that define the market. Services can also be the engine to kick-start a company's entry into the market, especially when selling complex and / or new technologies.

The key is to pursue a strategy in a determined way, rather than supporting software and serving customers at the same time.

Build the defense force in the old way and consolidate the market position

Although it is not clear whether the AI model itself or the underlying data can provide defense, good products and proprietary data almost always build a good business.

AI provides founders with a new perspective on solving old problems. For example, AI technology has generated new value in the relatively tired malware detection market through better performance.

In terms of the initial unique product capabilities, the opportunity to build sticky products and lasting business is eternal.

A16Z also found that several AI companies have consolidated their market position through effective cloud strategies, similar to the latest generation of open source companies.

Summary: AI is different from traditional software services.

Today's AI business is not exactly like software business. They involve ongoing human support and significant variable costs. It is usually not as easy to expand as people think.

Powerful advantages are not realized naturally with the mastery of technology, which is essential for the "one-time / multiple-sell" software model.

To some extent, these features make AI look like a service industry. In other words, the service company can be changed, but the service cannot be completely replaced.

Variable costs, scale dynamics and defense capabilities will ultimately be determined by the market, not by individual companies. The fact that the company data we are seeing now shows an unfamiliar pattern shows that AI is really new.

Original text link:

Https://a16z.com/2020/02/16/the-new-business-of-ai-and-how-its-different-from-traditional-software/

-end-

Https://www.toutiao.com/i6802425187846324748/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.