This is the correct posture for ordinary players to enter AIGC.

Latest update time：2023-06-29

Reads：

Fengse comes from Ao Fei Temple
Qubit | Official account QbitAI

The future is here, it’s just not evenly distributed yet.

This is a classic quote from William Gibson, the "father of cyberpunk", and it is also a summary of the current status of generative AI technology by Matt Wood, vice president of global products at Amazon Cloud Technology .

For the second half of the sentence, we can understand it more straightforwardly as "not evenly distributed yet".

Why do you say that?

Today, generative AI has been used in four fields: creative output (such as writing, programming, design) , functional enhancement (such as writing summaries, searching) , interactive experience (Q&A, chat) and decision support (various assistants) . Shows amazing potential.

We have seen that a large wave of new applications are emerging, and people from academia and industry are rushing to start a wave of entrepreneurship.

However, the benefits it brings are not evenly distributed , as there are barriers to fully embracing this technology.

As Matt Wood says, access to advanced technologies is often only within the “remit” of large tech companies, governments and universities.

Until now, many start-ups, organizations or individuals, as well as non-professional AI companies, either lack infrastructure, resources, or the correct way to open it... they can only watch others take off in the current trend.

what to do? How to break the situation?

This is also an issue that industry giant Amazon Web Services has always been concerned about.

At the just-held Amazon Cloud Technology China Summit, it finally gave a reference answer : providing solutions one by one for basic model construction, private customization, development efficiency, and computing power costs.

Ordinary players who want to seize the opportunity in this revolution and fully utilize the value of generative AI should not miss it.

4 Essential Postures to Unlock Generative AI

Without further ado, let’s go straight to Dr. Matt Wood’s wonderful speech.

To get straight to the point, Dr. Matt Wood said that if ordinary enterprises want to seize this opportunity and fully unlock the value of generative AI, they need to be prepared in four aspects :

First-class basic model, secure and private customized environment, low-cost/low-latency infrastructure guarantee, and professional development tools that can speed up.

How to understand these four points? What are their respective solutions?

First of all , Matt Wood said that in order to narrow the gap of "uneven distribution", our first consideration should be the method that can bridge the gap to the greatest extent.

That is a "best-in-class (best ) " basic model.

As the saying goes, "If you want to do your job well, you must first sharpen your tools." With such a model, we can build the best AI application within our capabilities.

So how do you get the best base model?

This requires moving out of the Amazon Bedrock platform, which was just released by Amazon Cloud Technology.

As a "bedrock" platform, it supports basic models from AI21 Labs, Anthropic, and Stability AI, including "ChatGPT's strongest competitor" Claude, Jurassic-2 that can support multiple small languages, and I don't need to introduce too much. Stable Difussion etc.

In addition, it also includes two advanced large language models, Amazon Titan, exclusively developed by Amazon Cloud Technology:

One is called Amazon Titan Text, which focuses on generative NLP tasks, and the other is called Amazon Titan Embeddings, which is used for search and personalized recommendations. It can translate text input into embedding codes that contain semantics, thereby making search results more relevant and consistent. Context context.

On top of the Amazon Bedrock platform, you can get the simplest experience using these models:

You can quickly use them for your application development by simply accessing them through API without worrying about any infrastructure ;

If you want to customize these base models, just provide a small number (as low as 20) of labeled examples.

Speaking of customized models, we come to our second point : a secure and private customized environment.

The so-called customization is to use relatively "small quantity but fine quality" data to transform the pre-trained basic model into a "vertical" model that is particularly good at a certain type of specific tasks, which is what we usually call fine-tuning.

According to Matt Wood, this process can be called a "game-change" because compared to directly developing a model for a specific task, fine-tuning/customizing a new model requires less data, resulting in a reduction in computing time, allowing it to be completed faster. Construct.

Needless to say, the key to custom models is proprietary data, which is essentially the customer’s valuable private property.

Therefore, to complete this process, it is particularly important to have a secure and private environment that ensures that the data is not used for training the platform's base model and other customers.

Amazon Bedrock has focused on security and privacy from the beginning, and the Amazon SageMaker JumpStart platform for more professional developers has also introduced these advanced basic models and also provides a secure fine-tuning environment.

Then , when the environment and platform are accurately in place, you can speed up development and think about how to complete the goal faster and faster.

答案其实藏在同一处——同样利用生成式AI技术开发的代码AI助手，比如Amazon CodeWhisperer，将显著提升我们的开发速度。

It has been trained on billions of lines of code, supports more than ten common programming languages and programming environments, and:

Be context-aware and provide code suggestions that are not limited to the current document;
Built-in security scanning function can quickly check whether there are vulnerabilities in the code and provide repair suggestions;
Source checking and annotation can be performed so that every line of your code is traceable (especially for open source projects) ;
And a specially added enterprise-level control function can be used to set the company's internal development specifications, policies, etc.

The most important thing is that individual developers can use these "beauties" completely free of charge !

Just open your IDE search plug-in and install it, then complete the registration with an email address. You don’t even need an Amazon Cloud Technology account.

According to statistics, after Amazon's internal developers used Amazon CodeWhisperer, the development speed was directly 57% faster than before, and the success rate of project completion was also increased by 27%.

Finally , after code development is complete, the cost and performance of training and inference become our top considerations.

Here, Amazon Cloud Technology has developed training and inference chips specifically for large models (including LLM and diffusion models) : Amazon Inferentia, Amazon Trainium, and Amazon Inferentia2.

How powerful are they?

The Amazon Inferentia chip alone, relying on the Amazon EC2 Inf1 instance developed by it, can directly reduce the inference cost of the model by 70% compared with ordinary GPUs .

With them, ordinary players can easily achieve cost-effective computing power.

After reading this, we can find that Amazon’s solutions in basic model construction, private customization, development efficiency and computing power costs have been prepared for a long time (all products have been released before) , and today “have not yet been evenly distributed” "Under the current technology status, it can be directly packaged into a mature solution and benefit everyone immediately.

The good news doesn’t stop there. In addition to the above four parts, Amazon Cloud Technology can also provide comprehensive support in terms of the most basic data .

As Dr. Matt Wood says:

Data is the genesis of generative Al .

How to fully tap the value? This is a problem that practitioners cannot avoid.

According to Dr. Matt Wood, what we need is not just transformative technology and infrastructure , but also an end-to-end data strategy , which can be summarized in three keywords: comprehensive, integrated and governed.

Amazon Cloud Technology has also been prepared for this:

Amazon Aurora and Amazon RDS can provide comprehensive and complete relational database services;
Seven tools, including Amazon Athena and Amazon EMR, can handle all analysis tasks you can think of, including interactive query, big data processing, warehousing, and integration;
Amazon Aurora and Amazon Redshift have opened up connections and seamless integration to further realize the vision of "zero ETL" and reduce users' work of manually migrating or converting data between different services;
Finally, there is the new product Amazon DataZone, which helps you realize transparent cross-department sharing of data (now open for preview) . You don’t have to worry about how to control the “degree” of data governance.

Someone has already experienced it for you

Talk is cheap, show me the code.

At this summit, Chen Xiaojian, general manager of the product department of Amazon Cloud Technology Greater China , also brought two types of practical cases to show us how it can help companies solve the problems encountered when innovating in today’s generative AI context. Computing resources and globalization issues.

Let’s look at the first one first.

Due to the explosion of generative AI technology, the overall computing power demand in the industry has exploded, which has brought about three major challenges :

The shortage of computing resources itself, confusion about the flexible supply of computing resources (not knowing when how much computing resources are needed), and the threshold and complexity of cloud operation and maintenance are not low enough.

How does Amazon Cloud Technology solve these problems?

First , they have launched various self-developed chips, such as Amazon Nitro, a dedicated cloud computing chip that can reduce the load on the CPU, Amazon Graviton, a general-purpose processor chip based on ARM architecture, and Amazon Inferentia and Amazon Traininum series chips specially developed for machine learning training and inference. , to help enterprises achieve ultra-high computing power cost performance.

Take F1 racing as an example. By using Amazon Graviton 3E chip of Amazon Cloud Technology for aerodynamic simulation, the organizers were able to develop a new generation of racing cars 70% faster than before, and reduced the pressure loss of the racing cars from 50% to 15%.

This means that car owners can more easily overtake when driving, giving fans more exciting racing battles.

Secondly , Amazon Cloud Technology helps users cope with sudden computing power needs and achieve a high degree of flexibility through a variety of rich computing network storage (which can be selected on demand or directly intelligently graded for you) and product portfolio (more than 600 different computing instances) oriented computing resource supply.

Take Weta Digital Cloud Studio as an example. It has won six Oscars for Best Special Effects, and its works include "Avatar", "Lord of the Rings", and "Rise of the Planet of the Apes".

Thanks to the infrastructure computing resources, cloud production stack, and machine learning stack provided by Amazon Cloud Technology, the studio completed the 3.3 billion rendering thread hours of "Avatar 2" on the cloud in just 8 months. Special effects production; looking back at "Avatar 1", "it took a full 14 months to see the first frame."

Finally , Amazon Cloud Technology also uses Serverless technology to simplify cloud operation and maintenance management, helping the game company Zhaoxiguangnian to focus on the development of the game "Marvel Snap" itself, without having to worry about system upgrades and expansion when the number of visits increases dramatically.

Chen Xiaojian said that after the game was finally launched, no back-end error occurred, which is unheard of in the gaming industry.

In addition to computing resources, many companies also have three major demands in today’s globalization trend :

All services must be managed uniformly on the cloud to ensure high reliability and low latency. The process of building a network must be fast and efficient, and must comply with local regulatory requirements to ensure smooth business expansion overseas.

In this regard, Amazon Cloud Technology has also responded one by one.

Take OPPO as an example. Their mobile phone business is spread all over the world, and there are hundreds of cloud VPCs and local resources that need to be connected, which greatly tests the difficulty of global networking.

Relying on the Amazon Cloud WAN service of Amazon Cloud Technology, a global network can be quickly built in just minutes. It can not only maintain the independent operation of OPPO's global business, regional compliance and autonomy, but also provide unified management.

Secondly, for the capital trading market Nasdaq , which has very high requirements for network reliability and latency , Amazon Cloud Technology used the Amazon Outposts service to help it build the first private local area in the history of the capital market industry and achieve ultra-low Delayed edge computing capabilities.

Finally, Pax , a global electronic payment terminal supplier , also relies on a series of services such as Amazon CloudHSM and Amazon Security Hub from Amazon Cloud Technology to meet the security service standards provided by multiple suppliers, ensure user payment security, and shorten the delivery cycle. 40%, cost reduction by 20%, and smooth expansion of international business.

……

The above successful cooperation cases well illustrate the effectiveness of Amazon Cloud Technology in helping enterprises get rid of infrastructure constraints and focus on innovation .

Adhere to "build a base for the base"

In fact, regarding this wave of generative AI, Amazon Cloud Technology made its positioning clear to the public as early as April :

Everyone goes to roll the big model itself, and we have to build the base for the base.

Now the speeches of two executives at the Amazon Cloud Technology China Summit once again released the same message:

Provide basic support and base support for the public to develop generative AI applications.

How do you understand the concept of "base"?

Amazon Bedrock, a fully managed platform that can quickly obtain the world's most advanced basic model capabilities, is the base;

High-performance infrastructure for training and running your own models, such as Amazon EC2 Inf1 instances powered by Amazon Inferentia chips, Amazon EC2 Trn1 instances powered by Amazon Trainium, and Amazon EC2 P5 instances powered by NVIDIA H100 Tensor Core GPUs, are the base;

Amazon SageMaker, which builds, trains, and deploys your own models from scratch, is the base;

Amazon SageMaker Jumpstart supports one-click deployment and fine-tuning of more than 150 popular open source models, and is even more of a base;

……

They are basic tools that serve our large models and generative AI application development. As a base, the purpose is to lower the threshold of use and turn large models and generative AI technology into a ready-to-use resource, allowing Reaching more businesses, organizations and individuals.

In today's era of fully embracing generative AI, the market does need companies to provide such services.

According to estimates by Grand View Research, the market size of generative AI may approach US$110 billion by 2030 alone.

Now, it is only a few months since this technology exploded, and the market is far from saturated.

In this blue ocean, various companies are full of expectations for the potential of generative AI to transform their products and business operations. Countless people want to seize the opportunity and occupy the high ground, but not everyone has the resources, methods, and sufficient the cost of.

So it is obviously not a wise move to explore from scratch on your own . Ordinary players are eager for a method that can quickly get on the car without spending too much effort on building the bottom support.

Of course, for most players, the lower the cost, the better.

In addition, with the development of technology, the division of labor in society has become more and more detailed. We pay attention to efficiency and convenience. We do not need everyone to reinvent the wheel. We only need to see who can use the wheel to run and reach their respective destinations faster.

Therefore, the introduction of these base services is also in line with the general trend .

Secondly, Amazon Cloud Technology has the ability to meet this urgent market demand and trend.

Needless to say, as a well-established technology giant, Amazon Cloud Technology has more than 25 years of AI experience , and more than 100,000 customers use Amazon Cloud Technology’s AI and machine learning services to help their businesses.

In the previous part, everyone also followed Chen Xiaojian to witness its achievements in helping large and small enterprises in all walks of life get rid of infrastructure constraints.

Then, if the market has demand and Amazon Cloud Technology has the capability, such a "two-way rush" can allow the public to focus most of their energy on real innovation, thus accelerating the technological progress of the entire human society——

Speaking of innovation, there’s good news.

Recently, Amazon Cloud Technology announced:

Invested US$100 million to establish a generative AI innovation center.

This move means that it will further provide flexible and cost-effective generative AI services to help every enterprise or organization utilize AI and unleash the huge potential of generative AI.

We can expect that with the implementation and deepening of this major initiative, we will see more innovative ideas and products become reality faster and more quickly.

One More Thing

It is worth mentioning that at the end of Dr. Matt Wood’s speech at this summit , the official ended with Bob Dylan ’s song “ The Times They Are A-Changin ”.

Just like the title of the song, times have changed, and the future of generative AI has arrived .

In this torrent, thanks to the existence of base services such as Amazon Cloud Technology, every enterprise and everyone has the opportunity to obtain, utilize and enjoy the benefits brought by new technologies.

We believe that no one will be left behind in this transformation, the uneven distribution will eventually be improved, and the looming future will become extremely clear.

Click to read the original text

Learn more about the 2023 Amazon Cloud Technology China Summit