Analysis: Wake up, Brazil! Digital transformation is not just for companies

RIO DE JANEIRO, BRAZIL – Recently a study about the digital strategy of a developed and advanced country like Japan was deeply analyzed. Despite being the third largest economy on the planet, the country ranks 27th in the 2020 index of digital competitiveness by IMD (International Institute for Management Development).

This study assesses economic performance, infrastructure, and government and business efficiency. In all, 63 countries were analyzed. Brazil, for comparison purposes, ranks 56th, a position that places it among the seven worst countries.

What did Japan do? It designed a digital strategy for 2030, with a strong focus on digital talent training, industry transformation, digital government and economic renewal, modernization of the business environment, and strong incentives for innovation and startups.

It is clear to them that the country’s digital transformation is not an alternative, but rather imperative. Each and every country that wants to have a solid position in the 21st century economy must undertake its digital transformation. That is what Brazil needs. It is falling further and further behind in a digital world. Digital transformation is not only for companies, but for the country as a whole.

One striking point in the study is the emphasis on AI. They clearly recognize that AI is a transformative technology and will significantly impact every aspect of Japanese society. AI is slowly moving out of prototypes and into production, including in extremely regulated and cautious sectors, such as in medicine.

The example of a hospital in the UK, described in “The world’s first large scale medical AI in Production – eye diagnosis by Deepmind” shows that AI is already the present and not the future.

The British hospital article shows something that most of the cases in the specialized media do not. Typically we see interviews with startups and reports on cases of companies using AI in prototype stages or in innovation labs. As Linus Torvalds, creator of Linux, once said, “Talk is easy. Show me the code. That’s right: most of the time we don’t see AI systems in production, only in labs.

Here in Brazil, the application of AI faces many challenges. A report published in NeoFeed, “Brazil advances, but is still in the ‘relegation zone’ of artificial intelligence,” shows that in the Latin American scenario, Brazil has a prominent role. The country has 206, or 42% of the total number of companies identified. Mexico comes second, with 97 companies, followed by Chile, with 57.

But despite the development, Latin America, which has less than 0.5% of global private investment in artificial intelligence, is still far behind more mature ecosystems, such as the United States and China. According to According to Stanford University’s 2019 Artificial Intelligence Index Report, only 0.2% of citations in AI-related patents worldwide come from the region, compared to 60.4% from North America and 22.1% from West Asia and the Pacific.

One of the greatest obstacles is the shortage of talent. It must be noted that Brazil does not have deep enough pockets to afford a company like OpenAI and its GPT-3, which is estimated to cost between US$5 million and US$10 million.

It is difficult to estimate the cost of developing GPT-3, since there is not much transparency in the process. But one thing is known: training large neural networks is very expensive. GPT-3 is a very large Transformer model, a neural network architecture that is particularly good at processing and generating sequential data.

GPT-3 consists of 96 layers and 175 billion parameters, the largest model language to date. To put this in perspective, Microsoft’s Turing-NLG, the previous record holder, had 17 billion parameters, and GPT-3’s predecessor, GPT-2, had 1.5 billion parameters.

Lambda Labs estimated the computing power needed to train GPT-3 based on projections from GPT-2. According to their estimate, training a neural network of 175 billion parameters requires 3,114E23 FLOPS. And operating GPT-3 requires an investment in hardware (and electricity, cooling, backup, etc.) that is estimated to be around US$150,000.

The key words in AI are “computational power and talent”. So, it is not available to everyone, and here in Brazil we don’t have Big Techs and investors in sufficient volume to invest billions of dollars in AI research and development. But, the question is: do we need to train many talents in AI that are capable of creating a GPT-3, or do we not need talents capable of generating solutions that use platforms like GPT-3 or others as a base?

We should train talents that are able to create business solutions. And to do this we won’t need as many PhDs, but rather professionals with enough background to use models and frameworks already in place. Big Techs have created them and they are available in the open source model. Moreover, we are beginning to see more and more AutoML solutions, which although they do not allow us to create something very sophisticated, they solve the vast majority of business problems.

The solution is not to do away with doctoral and masters courses in AI. On the contrary: we should encourage their creation, but in parallel we should massively create undergraduate courses and specialized training to form the professionals we need to have more AI applications in companies and governments.

When talking about training AI talents, we must separate researchers and scientists from the professionals who produce business solutions. Consider computing. There are several roles in computing, ranging from computer science researcher to application developer, front-end specialist, back-end specialist, and even infrastructure analysts. Quite a wide variety of professional activities.

In AI, one should also look at it this way: an AI solution does not live in isolation. AI algorithms are part of a business solution and must interact with other systems. So the first thing is to train executives and managers to understand AI and identify business opportunities for their companies.

Without the engagement of the companies’ managers there will hardly be funding for AI projects. By managers is meant company board members (most of them do not have much knowledge about AI and, therefore, do not take the subject to board meetings), the CEO and other C-levels (who often dump the problem of adopting AI in the hands of the CIO, who alone has little power to act), and middle managers, who are on the operations front line.

Any undergraduate course (medicine, law, management, marketing, education, etc.) should embed algorithms and AI as an essential subject. AI should not be only for ML scientists.

In the specific case of AI, there is life beyond the data scientist. Yes, there are the well-known data scientists, who use various techniques in statistics and ML to process and analyze data. They are the ones responsible for building models to investigate what can be learned from some data source, though often in a prototype, and far less at the production stage.

But we also have the data engineers, who develop a robust and scalable set of tools and platforms for processing data. One should be comfortable with the tasks of preparing SQL / NoSQL databases and building / maintaining ETL pipelines.

But beyond that. We have the machine learning (ML) engineers, who are responsible for training models and producing them. They need to be familiar with some high-level ML framework, and also know how to create scalable training, inference, and deployment pipelines for the models. And of course there are the ML scientists, the researchers working on cutting edge research and exploring new ideas that can be published at academic conferences.

Simplistically we can classify into (1) scientists who focus on Big Techs and companies with lots of money, the ones who create the GPT-3, and (2) operational professionals, who need to know how to use the AI tooling to create business solutions. They are the ones who use GPT-3 and the like to solve business problems with AI. This is the group that needs to be trained in quantity. In their training, aspects that today are not covered in adequate depth must be emphasized, because most master’s and doctoral courses are geared towards the training of scientists.

Training courses for operational AI professionals, either undergraduate or short term (a set of nanocourses could be a solution) should emphasize the operational aspects of identifying business problems and whether or not to validate the use of AI (does the AI project have value? Does it really need AI? Will the solution really involve AI? Are the conditions in place to implement the solution in AI? Is there any data? Is the data adequate?), training of the algorithms (how to deal with under and overfitting, long tail and eliminating biases), tuning for commissioning, performance analysis, ethics, security and privacy, adherence to LGPD, how to minimize the drift effect, interfacing with other systems and so on.

In short, these people need to be trained in AutoML tools, frameworks and environments such as AWS, Azure and Watson, etc., and above all to be trained in how to implement AI projects in production.

These professionals have to know how to deal with the real world, which is different from the research environment where training data sets don’t include Amazonian biotype faces and the system works well in the lab, where the vehicle is trained with American and European traffic signs and road signs, and not with those that exist in the Brazilian countryside, and where medical imaging data of diseases only shows them in advanced stages.

As the study “Japan Digital: Agenda 2030” proposes a digital strategy for Japan, we need a Brazilian AI strategy, not only for big and wealthy corporations, but for medium and small-sized companies. AI will be the new electricity and all companies will need to use it to remain competitive.

Our talent training is focused on training masters and PhD students, and virtually not at all on implementing and operating AI as business solutions to improve Brazil’s competitiveness. This should be our main attention focus in the next few years. Or we will miss another opportunity!

Source: Neofeed