HomeAI News
Google high-level earthquake! DeepMind refuses to share code with Google Brain, ex-employees explode inside

Google high-level earthquake! DeepMind refuses to share code with Google Brain, ex-employees explode inside

Hayo News
Hayo News
April 28th, 2023
View OriginalTranslated by Google
Recently, The information broke the news that although Pichai has the name of Google CEO, he is often under the shadow of the founder and cannot be manipulated at all...

If you use 8 words to sum up Google's recent situation, it would be - dire straits, internal and external troubles.

In November last year, ChatGPT, jointly launched by Microsoft and OpenAI, gave Google a fatal blow, almost pulling Google down from the throne of the search business. At the same time, there is also a big cake of more than 150 billion US dollars in advertising costs every year.

The crisis has fully exposed Google's problems under the leadership of CEO Sundar Pichai.

Recently, the foreign media The Information dug up perhaps one of the most critical reasons - the CEO title of Pichai, in many cases, is just a title.

Chopsticks, the CEO who lives in the shadows

Originally, advertisers cut spending last fall, which brought Google's business to a standstill. The impact of ChatGPT is simply worse.

The voice of "pick firewood and get out of class" has been heard endlessly in recent months. Criticism focuses on his failure management, such as his preference for incremental product improvements rather than subversive changes; his high tolerance for employee expansion, lazy corporate culture, and organizational inefficiency.

But in many cases, chopping firewood is involuntary.

In 2019, Pichai took over the heavy responsibility and became the CEO of Alphabet, Google's parent company.

Since then, he has confessed to his colleagues many times how difficult it is to manage a huge enterprise. Internal power struggles, regulators, and rebellious employees all weighed on him.

The people under Pichai's disobedience to control have already had clues.

On the one hand, because Google has a corporate culture similar to that of an academic institution or a government agency, in which tenured employees do not necessarily follow orders from their superiors. Woodcutter, on the other hand, doesn't like conflict.

He doesn't seem to force the team to implement his ideas, but this has also led to many executives not listening to his commands.

For example, chopping wood has sometimes been unable to get DeepMind CEO Demis Hassabis to prioritize certain projects, or to share his own code with Google Brain.

(I know that Google does not open source, but I did not expect that it is not open source in my own house...)

The combination of Google Brain and DeepMind is not simple

Last week, Alphabet suddenly announced the merger of DeepMind and Google Brain, which shocked the entire industry.

Alphabet means that after the merger of the two into Google DeepMind, the progress in the field of AI will be greatly accelerated.

From a certain point of view, this decision of the parent company is very reasonable. Because many software developed in parallel by DeepMind and Google Brain are the same. But for Google employees, the news came as a shock.

Google Brain and DeepMind have long been used to doing their own things and doing their own thing. Google Brain is owned by Jeff Dean, and DeepMind is owned by Demis Hassabis.

After the merger, DeepMind's CEO Hassabis became the CEO of the new department Google DeepMind, and Jeff Dean became the chief scientist of the new department Google DeepMind and Google Research.

Demis Hassabis, whose power has been greatly improved, has shown a strong sense of AGI mission and a sense of hunger for success. Even after the merger, the two bigwigs still report to the two lines of firewood respectively, and they are not affiliated with each other and do not interfere with each other.

But in the eyes of some people, it is actually equivalent to Jeff Dean, the head of Google Brain and Google's great hero, giving up his position to Hassabis of DeepMind. And, since each team has its own culture, conflicts at work are likely between the top-down DeepMind and the bottom-up Google Brain.

The brand-new title of "retreat" allows Jeff Dean, a technology guru who doesn't like to be a manager, to focus more on technology

In this regard, a former Google Brain researcher analyzed that since then, a large number of projects may have been canceled or merged, and there will also be large-scale adjustments and departures in terms of personnel. Even, this may also herald a larger restructuring of Google.

Nadella smiled

However, compared with the firewood who is struggling to lead the big ship of Alphabet, Microsoft CEO Satya Nadella (Satya Nadella) seems to be quite successful, and it can be said to be like a fish in water.

By forging a multibillion-dollar partnership, Nadella has linked the future of Microsoft with the future of OpenAI. This incident revealed a big message: Nadella has successfully got the executive team in step and working together.

On the one hand, Microsoft's Azure cloud server business provides funding for OpenAI's growing computing infrastructure and AI development.

On the other hand, in order to integrate OpenAI's technology into Microsoft's core products, the company needs to make some trade-offs in the distribution of computing power. And that meant some leaders had to make sacrifices on their projects.

It is not easy to unite so many big cows. It has to be said that Nadella has something.

Google Brain's ex-employee explodes

After the integration of Google Brain and DeepMind was fermented, employees who left Google broke the news of "Google Brain". He has previously (written in December last year) did not dare to release information about Google Brain, and all of it has been revealed. Why does Google Brain exist? He gives some in-depth analysis explaining why Google Brain exists.

In academia, researchers face financial difficulties while enjoying academic freedom.

However, coming to the industrial world, Google will pay you a lot of money to do machine learning. However, the requirement is that you must work on recommendation systems, advertising optimization, search rankings, etc., rather than pure research.

To put it bluntly, what you do is up to you, and the projects you do must have output.

Of course, in addition to making money, another reason for Google to fund open research is to maintain its leading position in the field of machine learning.

For example, Brain gave Google TensorFlow, TPU, significantly improved translation, JAX and Transformer. What started out as pure research is now hugely profitable.

In addition, Google Brain has another unusual feature-free publication policy. Brain's articles published at top ML conferences often exceed those of major universities. The main reasons for this are: 1) prestige; 2) researchers can take their knowledge with them when they resign; a more subtle reason is 3) catalyzing the growth of a field. The catalyst theory is that by publishing key research in areas related to Google's core business, the research direction will develop in Google's favor. For example, Google has always been very interested in NLP, and the publication of key research such as seq2seq in 2014 and Transformer in 2017 promoted the development of the entire NLP field.

Obviously, in peacetime it makes a lot of sense to spend $X to grow the entire pie, as long as that pie grows by more than $X.

But in wartime mode, competitor share gains are also important. And the OpenAI-Microsoft alliance means there is another giant that can expand the deployment of ML in both consumer and computing fields.

As Google begins to enter wartime mode, it is almost certain that this catalyst theory will also die. Still, Google is a pioneer in many consumer and business products, and it appears to be doing so in AI research as well.

In fact, if you want to become a winner like OpenAI in the field of machine learning, there will still be a lot of room in the future.

Founders still hold the reins

Another reason why the firewood is not convincing is that Larry Page, the former boss of the firewood, is still sitting in the boss position.

As a co-founder of Google, Page remains on Alphabet's board and, along with Sergey Brin, controls the company through special stock.

According to insiders, Page had always had little interest in Google's internal affairs, but when Hassabis wanted to maintain independence from Google, Page was the first person he approached.

In addition, the news of the integration of Google Brain and DeepMind was also released after the Alphabet board meeting.

This timing suggests that the decision to merge may have been made by the board of directors, rather than the firewood.

(The New York Times previously reported that Page attended the internal meeting.)

R&D but not development, OpenAI takes the lead

Google's lag is largely due to the risk aversion of executives.

Many key employees simply chose to leave because they could not launch new products.

For example, Noam Shazeer, one of the inventors of the LAMDA chatbot, previously wanted to have a chatbot in Google Assistant. But this proposal was not approved for a long time, and finally chose to leave.

Back in 2015, Google's image-recognition AI flagged black people as "gorillas," drawing criticism that left Google spooked.

Before OpenAI launched Dall-E 2 and became famous all over the world, Google had a similar model, but it did not release it because Google believed that it needed to be responsible for the content generated by AI.

Not only that, for the release of ChatGPT, when OpenAI poached Google's AI researchers and key engineers a few months ago, it made a solemn promise: "The technology will not be in vain, and the product will definitely be released."

We all know what happened next.

Google's future

At the end of 2022, the threat of ChatGPT became increasingly obvious, and Google began to sound the alarm.

Pichai urgently ordered to integrate AI chatbots into Google's many products.

However, this road to chasing Microsoft is not so easy. Querying with an AI chatbot will be more expensive than a regular search.

In March, Google urgently released Bard, but limited its functionality and targeted fewer users, all to save costs.

Someone broke the news that, in fact, Google chose a relatively weak machine learning model at the beginning, which does not require a lot of calculation like other models.

And in recent weeks, Google has worked hard to expand the size of the model behind Bard, roughly doubling in size.

Meanwhile, Dean is working with DeepMind (codenamed Gemini) to develop a new large-scale machine learning model that he hopes will help Google catch up to ChatGPT.

However, people familiar with the matter said that in fact, many new software under development cannot be ready before the I/O conference on May 10.

Although, many people believe that Pichai's previous background as a product manager is a powerful blessing for how to convert Google's existing AI research into products.

After all, Chrome, which successfully counterattacked and eventually crushed IE, came from the hands of firewood.

In this regard, Keval Desai, who has worked with him, said that looking at all Google executives, it is not accidental that Pichai can become the CEO.




Reprinted from 新智元View Original


no dataCoffee time! Feel free to comment