Behind Google's Comprehensive Integration of AI Power: The Rise and Fall of DeepMind

Jiazi Lightyear

follow

April 22, 2024

0 comments

948 browsing

1 Collection

27 minutes

  B-end product managers need to consider more about product functionality, stability, safety, compliance, etc., while C-end product managers need to consider more about product usability

DeepMind, once seen as the crown of AI, is now rarely mentioned; From "no one left" to "talent loss", DeepMind has now stepped down from the altar and lost its absolute advantage in talent competition. What exactly did it do wrong?

The talent race in Silicon Valley is intensifying.

AI talents with an annual salary of millions of dollars are no longer uncommon; Zuckerberg personally wrote an email to dig up loopholes, even proposing to be admitted without an interview; The celebrity startup company was rejected by the candidate due to being criticized for not having a 10000 Nvidia H100 GPU. No wonder Musk sighed, it is currently the "craziest AI talent battle".

DeepMind, once regarded as the crown jewel of AI, has now fallen from grace and lost its absolute advantage in the talent race.

In recent years, there has been an increasing number of top researchers leaving DeepMind for entrepreneurship. The strongest AI research institution under Google, transformed into a factory that exports AI talents to society on a large scale. Emerging capital darling such as Cohere, Mistral, Inflection, whose founders or chief scientists are all from DeepMind.

If leaving a big company to start a business is a reasonable destination for many talents, then Mustafa Suleyman, co-founder of DeepMind, made a high-profile return to another big company, Microsoft, as a worker in March 2024, two years after founding Inflection, which truly made Gu Ge and DeepMind more embarrassed.

Qi Lili, an international headhunter familiar with AI big model recruitment, told Jiazi Guangnian that in the recent recruitment market, large companies with more data and GPUs are indeed more attractive to candidates than unicorn startups because the former has a much higher probability of making big models. Even after unicorns offered higher salaries, candidates still chose to go to large factories. "In terms of salary, everyone is giving well now." "(Candidate) The first concern is computing power, and the second concern is data volume," said Qi Lili.

Riding the momentum of big companies, Google is also determined to integrate the research and development capabilities of AI big models, taking the culmination Gemini to a new level. Especially in the past week, Google's actions have gradually accelerated.

On April 15th, DeepMind CEO Demis Hassabis stated at the TED conference held in Vancouver that Google will invest over $100 billion in AI development in the future, and made a high-profile statement that Google's computing power is higher than competitors such as Microsoft.

On April 18th, Google CEO Sundar Pichai announced that he would accelerate the development and deployment of AI products through team restructuring. All teams studying AI models will be included in DeepMind led by Hassabis.

"These changes continue the work we have done over the past year, simplifying our structure and improving speed and execution. This helps accelerate the development of our Gemini model, unify our machine learning infrastructure and development team; this enables faster decision-making, smarter computing allocation, and better customer experience," Pichai wrote on the company blog.

It has been a full 365 days since DeepMind and Google Brain officially merged. This larger new department still bears the name DeepMind, indicating a significant increase in Hasabis' power at Google.

But Hasabis is also gradually distancing itself from DeepMind, which was founded 14 years ago.

1、 "A company with no one leaving"

In 2010, Hassabis co founded DeepMind in London with his childhood friend Suleiman and research colleague Shane Legg from New Zealand.

This is a seemingly stable trio with clear style divisions: Hassabis and Legg met at University College London, with doctoral degrees in neuroscience and artificial intelligence respectively. Suleiman focuses on product development and policy aspects.

Perhaps it is related to the study experiences of Hassabis and Suleiman in Cambridge and Oxford, respectively. Within a few years, many researchers from Cambridge and Oxford came to DeepMind, making it a holy land for deep learning. Even if it faced financial difficulties, it could still be the object of competition for Google and Facebook at that time.

Should a product manager choose to work on the B-end?

In recent years, the Internet has experienced the upsurge of spending money on the market, and more and more realize that it is not enough to rely solely on the power of the C end. It needs to make up for the shortcomings of the B end as soon as possible. Should the product manager choose to transform into a B2B platform?

View details>

In January 2014, Google acquired DeepMind for $600 million and granted it great independence. When it made multiple global media headlines in the future, DeepMind remained the "AI company based in London," separated from the bustling Silicon Valley by an entire United States and the Atlantic Ocean.

Hassabis, who was not yet 40 years old at the time, insisted on keeping DeepMind in his hometown. He was born in northern London, with his father being Greek Cypriots and his mother being Chinese Singaporeans. If there is no more introduction, it is easy to mistake this former chess prodigy for a DeepMind intern. But he is also known in the industry as one of the smartest people on Earth.

Hassabis revealed that there is no room for negotiation on whether DeepMind should stay in London. "If you have a doctor's degree in physics from Cambridge University and want to do some technology that can change the world, there are not so many choices here, but there are thousands in Silicon Valley. If you want to focus on some long-term goals, Silicon Valley is like a foam, and everyone wants to create a Snapchat in five minutes. There is too much noise in this system."

In 2015, when publishing the above viewpoint to WIRED magazine, Hassabis also stated that Bell Labs or Microsoft Research Institute are more like academic institutions, and the combination of DeepMind and Google is "the fusion of two worlds, and all great progress will come from this - neuroscience and machine learning, academic thinking and entrepreneurial thinking, combined in one big company.".

However, for a long time afterwards, DeepMind also showed a pure academic temperament, without the pressure to help Google increase advertising revenue. "What we are doing is long-term research," Suleiman once said. "If we are only developing solutions for products that can be thought of now, then we are limiting our imagination."

In March 2016, AlphaGo defeated world Go champion Li Shishi, a stunning victory that opened up DeepMind's public awareness and pushed global investors into the wave of AI applications. That month, American media Business Insider published an article praising DeepMind, calling it "a London startup that no one wants to leave."

After releasing AI masterpieces in gaming and research fields such as AlphaZero and AlphaFold, DeepMind has published thousands of highly cited papers by 2020, of which 13 have been published in top journals such as Nature or Science.

Mark Riedl, Associate Professor at the School of Interactive Computing at Georgia Institute of Technology, said when asked about AI company rankings in 2021, "From a reputation perspective, DeepMind, OpenAI, and Facebook's FAIR have good reasons to be in the top three."

AlphaFold simulated schematic diagram of human protein structure, image source: DeepMind official website

As a result, DeepMind embarked on a unique AI product path: using games as a form and papers as results, demonstrating that machines can learn and solve complex problems through repeated experimentation, and then promoting practical applications in medical, scientific research, and other fields.

This is the background laid by the British AI academic community, represented by Oxford and Cambridge, for DeepMind. But in the AIGC era that focuses more on product application after 2022, this has also become its hidden concern.

In 2016, when DeepMind's exposure reached its peak, according to Business Insider citing data from LinkedIn, at least ten deep learning postdoctoral fellows or researchers from Oxford and Cambridge left the academic community to join DeepMind. In the two years following Google's acquisition, DeepMind has also expanded from 100 people to 250 people.

Stephen Cave, Director of the Future Intelligence Research Center at the University of Cambridge, pointed out that DeepMind highlights the problem of academic talent loss to businesses. He pointed out that some research directions in artificial intelligence are difficult to receive the attention they deserve in enterprises, and some philosophical work is difficult to attract most technology companies to invest time in.

After 8 years, Stephen Keff admitted in an email in April 2024 to "Jiazi Lightyear" that there were media exaggerations in the interview content that year.

And the positioning and status of DeepMind have quietly changed.

2、 Running away during the period of labor pain

The decision to acquire DeepMind itself is not well received by Google.

At that time, Google had at least ten AI research teams under its umbrella, among which Google Brain was one of the more well-known, established only one year later than DeepMind. The team's early main projects included improving image recognition on Google Maps, speech recognition on Android systems, etc. Later in 2017, the Transformer architecture was proposed, establishing its own position.

When DeepMind joined Google as a special independent character, the friction between the teams intensified.

Several former Google employees have told The Information that Google Brain has already demonstrated its value, but DeepMind has long been unable to justify the huge costs spent on it. On the other hand, the DeepMind team is also not interested in machine learning algorithms developed by other departments of Google.

Cambridge University computer professor Jon Crowcroft told CNBC, "Investors will focus on where big customers and revenue sources are from in the early stages. (DeepMind) They have developed technology and prototype tools that can achieve this goal, but there is still a long way to go between the laboratory and reality."

Five years after being acquired by Google, DeepMind entered a period of pain in order to better align its academic pursuits with Google's commercial product orientation.

DeepMind's revenue and losses from 2016 to 2019, Image source: VentureBeat

In early 2019, three DeepMind artificial intelligence engineers resigned, such as the renowned security engineer Ben Laurie who returned to his former employer Google. Although the total number is small, the golden body of DeepMind, which has no one to leave, has been broken.

A former DeepMind research scientist told European startup media Sifted that one of the reasons he left DeepMind was that he was unsure if the project he was working on could survive the pressure of converting scientific research results. "What will happen to these basic research projects when we are asked to generate more business impact?"

Even Hassabis himself has shown signs of wavering.

Over time, Hassabis hopes to further separate DeepMind from Google, as he is concerned that Google may apply technology to US military projects. He once had the idea of creating an independent company with DeepMind intellectual property rights.

However, according to insiders, Hassabis told DeepMind employees in 2021 that the spin off project, known as the "Mario Plan" at DeepMind executives, was put on hold after Google CEO Pichai promised to provide more funding.

However, apart from allowing DeepMind to mysteriously turn losses into profits on its financial statements since 2020, Hassabis's once imagined "integration of two worlds" is still not going smoothly. Prior to the release of ChatGPT, DeepMind continued to lose AI talent in the first half of 2022.

In April 2022, Cohere successfully recruited Phil Blunsom, the Chief Researcher who had been working at DeepMind for 7 years, to serve as the Chief Scientist. During the same period, Martin Schmid, along with Rudolf Kadlec and Matej Moravcik, developed the Texas poker AI player DeepStack that could defeat humans. They left DeepMind in March 2022 and founded EquiLibre Technologies, an algorithm company related to stocks and cryptocurrencies. At that time, no one knew for sure which company the person who had run away from DeepMind had founded.

The financial report released later showed that DeepMind's profits for the same period in 2022 decreased by more than 40%, and employee costs decreased by 39% year-on-year.

For the AI industry, this year is the darkness before a new dawn. But the impending dawn did not illuminate DeepMind's path.

3、 Arranged marriage

Faced with the fierce competition from OpenAI and ChatGPT, Google's earliest response was the chatbot Bard, which was pieced together within a few months. At the March 2023 press conference, Bard was riddled with loopholes, causing Google's market value to evaporate by billions of dollars on the spot.

The only useful aspect of Bard may be to make Google determined to end internal battles and force the merger of the two best AI research teams.

Insiders have stated that DeepMind and Google Brain have rarely collaborated or shared code, but due to both wanting to develop their own models to compete with OpenAI and having massive computing power requirements, they have no choice but to collaborate.

Sid Jayakumar, founder of the generative AI startup Finster AI, worked at DeepMind for 7 years, but chose to leave in August 2023, a few months after the two teams merged. He said, "Focusing more on products will lower the morale of some people in cutting-edge research."

"We hired many excellent engineers and researchers, basically asking them to replicate the academic environment in the industry, which was unique at the time and was also necessary for building products such as AlphaGo and AlphaFold," said Jia Guma.

But this may not be what big model development requires anymore.

Another entrepreneur who has worked at DeepMind, Jonathan Godwin, pointed out in his personal blog that there are relatively few PhDs in OpenAI with a traditional machine learning academic background, and it is a research company that focuses more on engineering applications. More importantly, OpenAI realizes that the competition ground for AGI (General Artificial Intelligence) is not in the academic journals with the highest citation rates, but in the subjective experience of AI users - truly seeing artificial intelligence as a product.

Therefore, even though DeepMind has a high degree of similarity to OpenAI - with the strong support of technology giants and a group of utopian tech elites, it is precisely these reasons that DeepMind did not become the first to make ChatGPT.

In December 2023, Google released the first version of Gemini. With the support of YouTube services, Gemini enjoys unique advantages in data training related to images, videos, and subtitles. Compared to OpenAI, which relies on Microsoft, Google also has its own data center.

But Gemini has always been an arranged marriage. Collaboration did not lead to a handshake between DeepMind and Google Brain, and the brand new AI team sometimes still conflicts due to resource issues.

In some data reviews, the Gemini Ultra has surpassed the GPT-4 Turbo. But in February 2024, Gemini, which received an update, was still ridiculed by the public for showing racist tendencies when replying to users.

To be fair, the new DeepMind team has helped Google narrow the gap with OpenAI, but has not yet fully surpassed it.

According to The Information, Hassabis recently told a colleague that he believes it will be particularly difficult for Google to catch up with its competitors in this field. On the other hand, Google emphasizes that if OpenAI uses YouTube to train Sora, it is a violation of YouTube guidelines.

The report also cited three individuals attending an internal meeting as saying that Hassabis's performance made people feel that he is very indifferent to the merger of Google's AI department. Just a few months ago, Hassabis attempted to limit the collaboration between the two departments to the Gemini project.

On the contrary, Google will close other AI projects to ensure Gemini's progress. At present, the number of people in the Gemini project has increased to 1000. When you open the homepage of the DeepMind website, the first slogan that catches your eye is "Welcome to the Gemini era", rather than products starting with "Alpha -".

Once, every product of DeepMind that started with "Alpha -" was able to win industry recognition. But the newly released AlphaGeometry in January 2014 drew criticism. The new generation of investors and industry professionals are starting to be picky about DeepMind's product development path, which has won countless praises.

The company stated in its blog that the AlphaGeometry system can solve "complex geometric problems at a level close to that of a human Olympic gold medalist.". But New York University professor Ernest Davis believes that the blog hides the limitations of AlphaGeometry, such as its ability to handle only two-dimensional geometric shapes and inability to understand what area is.

AlphaGeometry creative image, image source: DeepMind official website

One of DeepMind's early investors, Frank Meehan, stated that such controversies have been diverting Hassabis' attention from AGI research, which may be the "root cause of ongoing frustration.". He satirizes that OpenAI is "generating incredible videos based on text prompts, while Google is busy with some image issues.".

Meanwhile, personnel turnover continues.

In May 2023, Arthur Mensch, co-founder of Mistral AI, left DeepMind. In less than a year, this company has become OpenAI's "strongest European competitor". He once revealed to the public that the reason for leaving DeepMind is that the company is "not innovative enough".

According to Bloomberg News, in January 2024, Laurent Sifre and Karl Tuyls also chose to leave DeepMind and plan to establish a new AI company in Paris. The new company tentatively named Holistic may raise 200 million euros.

4、 Is it a reservoir or an ocean?

At present, the biggest opportunity for DeepMind to attract engineering and R&D talents is still thanks to Google. Due to computing power and data volume, the attractiveness of large companies in the talent competition is gradually increasing.

In order to more effectively attract top AI talents, Meta is launching an unprecedented strategy, such as Zuckerberg personally writing emails to dig corners and successfully enlisting former DeepMind researcher Michael Valko to join Meta. Sources revealed that Zuckerberg emphasized the importance of artificial intelligence for the development of Meta in his letter and expressed a strong desire to work together with the recipient. In addition, Meta has also proposed a "green channel" without interviews.

Perplexity CEO Aravind Srinivas stated on a podcast that he cannot poach Meta's top AI talent because his company does not have enough GPUs. "I am trying to hire a very senior researcher from Meta, do you know what he said? 'Come back to me when you have 10000 H100s,'" he said

At the same time, DeepMind founder Suleiman joined Microsoft with a large number of researchers from Inflection, leading the newly established Microsoft AI department and overseeing consumer grade AI products such as Copilot.

Image source: Suleiman's X homepage

This move also made it natural for Microsoft to announce its London office, promising to invest £ 2.5 billion over the next three years, led by former DeepMind and Inflection member Jordan Huffman.

In the struggle for top artificial intelligence talent, Microsoft has stood in front of the moat of DeepMind.

"People are willing to join companies like Microsoft (or Google, Apple) in order to have an impact on the real world," Cambridge University computer professor Jon Crockett told "Jiazi Lightyear." And perhaps these financially strong companies are just hiring them to stop the competition for hiring them, who knows? "

To be fair, Google has helped DeepMind do everything it needs to do.

In January of this year, Google followed the example of OpenAI and utilized a special stock pool to provide millions of dollars in stock incentives for DeepMind talent. On April 9th, Google also announced the launch of Axion, a data center chip based on Arm architecture, eager to break free from its dependence on Nvidia.

On the other hand, in response to investor demand, Google announced two rounds of major layoffs in January and April this year, and made it clear that the reason was not the company's financial difficulties, but rather the difficult decision that Google had to make when prioritizing artificial intelligence. The layoffs in January mainly involved the advertising sales department, as AI technology has been deeply integrated with the advertising business.

Google co-founder Sergey Brin also personally called a DeepMind researcher who was preparing to leave for OpenAI and convinced him to stay.

But the reality facing Google is that top AI talents, whether starting their own businesses or joining other financially strong AI teams, have much more choices than when Google acquired DeepMind a full 10 years ago.

At the same time, the internal chaos within the company has also hindered Google's progress, and the latest AI team reorganization has had a difficult effect beyond the integration of Google Brain and DeepMind over the past year.

"Google is very strong in artificial intelligence, but it has made many different efforts in this field, and so far, the overall results are less than the sum of the various parts." said Pedro Domingos, a machine learning researcher and professor at the University of Oxford. "Google needs someone to make everything coordinated and more profitable."

Ekaterina Almasque, a partner at venture capital firm OpenOcean, believes that DeepMind is no longer the absolute leader in the AI field by far. "All companies are competing for the same AI talent pool, and it's becoming more like a reservoir, not an ocean anymore."

But what can be certain is that DeepMind is facing more open competition, and Silicon Valley is not the only option for AI talents.

Klokoff admitted to "Jia Zi Lightyear" that people are now more willing to work in places they can afford. "I'm not sure how much role Silicon Valley can play in the talent competition, it's too expensive no matter how it looks. It's like people always think that movies must be made in Hollywood, but in fact, the vast majority of movies are not like that."

"The competition for artificial intelligence talent is indeed very fierce, but it is only a normal part of a cycle," added Klokoff.

Everything seems normal, DeepMind has not done anything wrong, and Google has provided the necessary support. Behind the ups and downs, only the giant wheel of the times has sailed through the surging waves.

*Reference materials:

The Deep Mind of Demis Hassabis, WIRED

Oxford and Cambridge are losing AI researchers to DeepMind, Business Insider

Deep Fusion: Tensions Linered Within Google Over DeepMind, The Information

Why top AI talent is leaving Google's DeepMind, Sifted

"Why didn't DeepMind build GPT3?" by Jonathan Godwin

Google's Demis Hassabis Chafes Under New AI Push, The Information

Author: Tian Siqi; Editor: Wang Bo

Original title: Google fully integrates AI power: Behind DeepMind's ups and downs | Jiazi Lightyear

Source official account: Jiazi Guangnian (ID: jazz year), based on the frontier of China's scientific and technological innovation, dynamically tracks the development of leading scientific and technological enterprises and the technological upgrading of traditional industries.

This article is authorized for publication by @ Jiazi Guangnian, a collaborative media platform where everyone is a product manager. Reproduction without permission is prohibited.

The question is from Unsplash and is based on the CC0 protocol

The viewpoint of this article only represents the author himself, and everyone is a product manager. The platform only provides information storage space services.

谷歌全面整合AI力量背后：DeepMind浮沉史

甲子光年关注

2024-04-22

0 评论948 浏览1 收藏27 分钟

曾经被视为AI皇冠的DeepMind，现在已经少有人提及；从“无人离开”到“人才流失”，DeepMind如今走下神坛，在人才争夺中失去绝对优势。它到底做错了什么？

硅谷的人才争夺战愈演愈烈。

年薪百万美元的AI人才已不再稀奇；扎克伯格亲自写邮件挖墙脚，甚至提出免面试录取；明星创业公司遭候选人拒绝，只因被嫌弃没有“10000块英伟达H100 GPU”。怪不得马斯克感叹道，眼下正是“最疯狂的AI人才争夺战”。

曾经被视为AI皇冠明珠的DeepMind，如今却走下神坛，在人才争夺中失去绝对优势。

近年来，从DeepMind出走创业的顶尖研究人员越来越多。谷歌旗下的最强AI研究机构，化身大规模向社会输出AI人才的工厂。Cohere、Mistral、Inflection等新兴资本宠儿，其创始人或首席科学家都出自DeepMind。

如果说离开大公司创业是许多人才合理的去向，那么DeepMind联合创始人穆斯塔法·苏莱曼（Mustafa Suleyman）在创办Inflection两年后，于2024年3月高调回归另一家大厂微软去当打工人的举动，着实让谷歌和DeepMind更尴尬。

熟悉AI大模型招聘的科锐国际猎头綦俐丽对「甲子光年」表示，近期招聘市场中，拥有更多数据量和GPU的大厂，的确要比独角兽创业公司更吸引候选人，因为前者“做成大模型这件事的概率高得多”。甚至在独角兽开出更高的薪资后，候选人还是选择去大厂。“工资的话，现在大家给得都不错。”“（候选人）第一关注算力，第二关注数据量。”綦俐丽说。

乘着大厂起势的东风，谷歌也下定决心整合AI大模型的研发力量，让集大成之作Gemini迈上新的台阶。尤其是最近一周，谷歌的动作逐渐加快。

4月15日，DeepMind首席执行官德米什·哈萨比斯（Demis Hassabis）在温哥华举行的TED大会上表示，谷歌未来将投入超过1000亿美元用于AI开发，并且高调表示谷歌的计算能力高于微软等对手。

4月18日，谷歌CEO桑达尔·皮查伊（Sundar Pichai）宣布将通过团队重组加快AI产品的开发部署。所有研究AI模型的团队都将归入哈萨比斯领导的DeepMind。

“这些变化延续了我们过去一年所做的工作，即简化我们的结构并提高速度和执行力。这有助于加快我们Gemini模型的开发，统一我们的机器学习基础设施和开发团队；这样就能实现更快的决策、更智能的计算分配和更好的客户体验。” 皮查伊在公司博客中写道。

此时距离DeepMind与Google Brain正式合并，已经过去整整365天。这个更庞大的新部门仍然冠以DeepMind的名号，意味着哈萨比斯在谷歌的权力大大增强。

但哈萨比斯也和14年前创立之初的DeepMind渐行渐远。

一、“无人离开的公司”

2010年，哈萨比斯与儿时好友苏莱曼，以及来自新西兰的研究同僚谢恩·莱格（Shane Legg）共同在伦敦创立了DeepMind。

这是一个看起来稳固的三人组，也有着清晰的风格划分：哈萨比斯和莱格结识于伦敦大学学院，两人分别拥有神经科学和人工智能的博士学位。苏莱曼则专注于产品开发和政策方面。

或许和哈萨比斯与苏莱曼分别拥有剑桥和牛津的学习经历有关，几年之内，许多剑桥和牛津的研究人员来到DeepMind，把这里打造为深度学习圣地。即便陷入资金困境，它也能成为当时的谷歌和Facebook竞相追逐的对象。

产品经理到底该不该选择做B端？

近几年互联网经历了砸钱做市场的热潮后，越来越意识到，仅靠C端发力是不行的，需要尽快补齐B端的短板。那产品经理到底该不该选择转型做B端呢？

查看详情 >

2014年1月，谷歌以6亿美元将DeepMind收入囊中，并且赋予对方极大的独立性。在日后多次登上全球媒体头条时，DeepMind仍然是那家“位于伦敦的AI公司”，与熙熙攘攘的硅谷相隔一整个美国外加大西洋。

当时尚未满40岁的哈萨比斯坚持把DeepMind留在了自己的家乡。他出生于伦敦北部，父亲是希腊裔塞浦路斯人，母亲是新加坡华裔。如果不多做介绍，你很容易把这位曾经的国际象棋神童当成DeepMind的实习生。但他也被业内称作“地球上最聪明的人之一”。

哈萨比斯透露，DeepMind是否要留在伦敦是没有商量余地的。“如果你拥有剑桥大学的物理学博士学位，并且想做一些改变世界的技术，那么这里没有那么多选择，但在硅谷有成千上万个。如果你想专注于一些长期的目标，硅谷又像是一个泡沫，人人都想5分钟打造一个Snapchat，这个体系里噪音太多了。”

在2015年面向《WIRED》杂志发表上述观点时，哈萨比斯还表示，贝尔实验室或者微软研究院更像是学术机构，而DeepMind与谷歌的结合是“两个世界的融合，所有伟大的进步都会因此而到来——神经科学和机器学习，学术思维和创业思维，结合在一家大公司内”。

不过在那之后很长一段时间里，DeepMind同样展现出纯粹的学术气质，它没有帮助谷歌增加广告营收的压力。“我们在做的是长期的研究工作，”苏莱曼曾说，“如果我们只是为现在可以想到的产品做解决方案，那我们就是在限制自己的想象力。”

2016年3月，AlphaGo战胜围棋世界冠军李世石，这场惊天胜利打开了DeepMind的公众知名度，也把全球的投资者推向AI应用的浪潮。当月，美国媒体Business Insider发文吹捧DeepMind，将它称作“一家无人愿意离开的伦敦初创公司。”

接连发布AlphaZero、AlphaFold等游戏和科研领域的AI力作后，截至2020年，DeepMind已发布上千篇引用量极高的论文，其中13篇发表于《自然》或《科学》这两大顶级期刊。

佐治亚理工学院交互式计算学院副教授马克·利德（Mark Riedl）在2021年被问及AI公司排名时表示：“从声誉上看，DeepMind、OpenAI 和 Facebook的FAIR有充分理由当选前三名。”

AlphaFold模拟的人类蛋白质结构示意图，图片来源：DeepMind官网

由此，DeepMind走上一条独特的AI产品路线：以游戏作为形式，以论文作为成果，论证机器可以通过反复试验来学习并解决复杂的问题，然后推动在医疗、科研等领域的实际应用。

这是以牛津和剑桥为代表的英国AI学术界为DeepMind打下的底色。但在2022年后更侧重产品应用层面的AIGC时代，这也成为它的隐忧。

在DeepMind的曝光度达到顶峰的2016年，据Business Insider援引领英数据统计，至少十余位牛津和剑桥的深度学习博士后或研究人员离开学术界加入DeepMind。在谷歌收购后的这两年里，DeepMind也从100人扩大到250人。

剑桥大学未来智能研究中心主任史蒂芬·凯夫（Stephen Cave）指出，DeepMind彰显了学术界人才流失到企业的问题。他指出部分人工智能研究方向很难在企业得到应有的关注，一些富有哲学性的工作很难吸引大多数科技公司在上面投入时间。

时隔8年，史蒂芬·凯夫于2024年4月在邮件中对「甲子光年」坦言，当年的采访内容存在媒体夸大的成分。

而DeepMind的定位和地位，也悄然发生了改变。

二、阵痛期的出走

收购DeepMind的决定本身，在谷歌就不是众望所归。

当时谷歌麾下至少有十多个AI研究团队，其中Google Brain是知名度较高的一个，仅比DeepMind晚成立一年。该团队早期的主要项目包括改进谷歌地图的图像识别，安卓系统的语音识别等，后来在2017年提出Transformer架构，奠定了自身的地位。

当DeepMind以一种特殊的独立角色加入谷歌后，各团队的摩擦加剧。

谷歌多位前员工对The Information表示，Google Brain早已展示了自己的价值，但DeepMind长期无法证明花费在它身上的巨额成本是合理的。另一方面，DeepMind团队对谷歌其他部门开发的机器学习算法也不感冒。

剑桥大学计算机教授乔恩·克洛克夫特（Jon Crowcroft）对CNBC表示：“投资者会在更初期的阶段就关注大客户和收入来源到底在哪儿。（DeepMind）他们开发了可以实现这一目标的技术和原型工具，但实验室与现实之间还有很长的路要走。”

被谷歌收购五年后，为了更好地磨合学术追求和谷歌的商业产品导向，DeepMind来到一段阵痛期。

DeepMind从2016年到2019年的营收与亏损，图片来源：VentureBeat

2019年初，三位DeepMind人工智能工程师离职，比如著名安全工程师本·劳里（Ben Laurie）回到前雇主谷歌麾下。总数虽小，但DeepMind“无人离开”的“金身”已破。

一位DeepMind前研究科学家对欧洲创业媒体Sifted表示，他离开DeepMind的原因之一是，他不确定自己正在从事的项目能否在科研成果转化的压力中幸存。“当我们被要求产生更多商业影响时，这些基础研究项目会怎么办？”

甚至哈萨比斯本人也出现了摇摆。

随着时间的推移，哈萨比斯希望将DeepMind与谷歌更加彻底地分离，因为他担心谷歌会把技术应用于美国军方的项目。他一度萌生了创建一家拥有DeepMind知识产权的独立公司的想法。

不过知情人士称，哈萨比斯在2021年告诉DeepMind员工，在谷歌CEO皮查伊承诺提供更多资金后，在DeepMind高层被称作“马里奥计划”的分离大业就此搁置。

然而除了从2020年起让DeepMind“神秘地”在报表上扭亏转盈以外，哈萨比斯曾畅想过的“两个世界的磨合”仍然不顺利。在ChatGPT发布前夕，DeepMind在2022年上半年持续流失AI人才。

2022年4月，Cohere顺利把在DeepMind工作了7年之久的首席研究员Phil Blunsom招募来担任首席科学家。同一时期，Martin Schmid与Rudolf Kadlec、Matej Moravcik开发了可以击败人类的德州扑克AI玩家DeepStack，他们于2022年3月离开 DeepMind，创立了EquiLibre Technologies，这是一家有关股票和加密货币的算法公司。当时已没人数得清，这是当年从DeepMind出走的人创立第几家公司。

之后公布的财报显示，DeepMind在2022年全年同期利润下降40%以上，员工成本同比削减39%。

对于AI行业来说，这一年是崭新黎明前的黑暗。但即将到来的黎明，并没有照亮DeepMind的前路。

三、包办婚姻，貌合神离

面对来势汹汹的OpenAI和ChatGPT，谷歌最早给出的回应是在几个月内拼凑而成的聊天机器人Bard。2023年3月发布会上，Bard漏洞百出，当场让谷歌市值蒸发上千亿美元。

Bard唯一有用的地方，或许是让谷歌下定决心终结内部厮杀，让最优秀的两个AI研究团队强行合并。

知情人士表示，DeepMind和Google Brain原本就很少合作或者共享代码，但由于双方都想开发自己的模型来与OpenAI竞争，都有海量算力需求，所以他们别无选择，只能合作。

生成式AI初创公司Finster AI的创始人希德·贾古玛（Sid Jayakumar）在DeepMind工作了7年，但在2023年8月，也就是两个团队合并几个月后选择出走。他表示：“更加注重产品会让一些人在前沿研究方面士气低落。”

“我们聘请了很多非常优秀的工程师和研究人员，基本上是在要求他们在业界复刻学术环境，这在当时是独一无二的，也是构建AlphaGo和AlphaFold等产品所需要的，” 贾古玛表示。

但这或许已经不是大模型开发所需要的。

另一位曾在DeepMind任职的创业者乔纳森·古德温（Jonathan Godwin）在个人博客中指出，OpenAI中出身传统机器学习学术背景的博士相对较少，它是一家更注重工程应用的研究公司。更重要的是，OpenAI意识到，AGI（通用人工智能）的竞赛场不在引用率最高的学术期刊上，而是存在于AI用户的主观体验中——把人工智能真正看作一个产品。

因此，即便DeepMind与OpenAI相似度极高——拥有科技巨头的鼎力支持，以及一群乌托邦中的技术精英，但正是上述原因导致DeepMind没有成为率先做出ChatGPT的角色。

2023年12月，谷歌发布了Gemini的第一个版本。在YouTube服务的支持下，Gemini有关图片、视频和字幕的数据训练享有独特的优势。与依赖微软的OpenAI相比，谷歌也有自己的数据中心。

但Gemini始终是一场“包办婚姻”。合作并没有让DeepMind与Google Brain握手言和，全新的AI团队有时仍会因资源问题发生冲突。

在某些数据评测中，Gemini Ultra已经可以超越GPT-4 Turbo。但2024年2月，迎来更新的Gemini因回复用户时出现种族主义倾向仍然受到公众嘲笑。

公允地说，全新的DeepMind团队已经帮助谷歌缩小了与OpenAI的差距，但还没有全面超越。

据The Information报道，哈萨比斯最近告诉一位同事，他认为谷歌要赶上该领域的竞争对手是特别困难的事。另一方面，谷歌强调，如果OpenAI使用Youtube来训练 Sora，属于违反YouTube准则的行为。

该报道还援引参加某内部会议的三位人士称，哈萨比斯的表现让人觉得他对于谷歌AI部门的合并十分淡漠。就在几个月前，哈萨比斯还曾试图把将两个部门的合作限制在Gemini项目上。

相反，谷歌会为了确保Gemini的进展而关闭其他AI项目。目前Gemini项目的人数已经增至1000人。而打开DeepMind网站主页，首先映入眼帘的标语是“欢迎来到Gemini的时代”，而不是以“Alpha-”开头的产品。

曾经，DeepMind每一项以“Alpha-”开头的产品都能赢得业界的赞赏。但2014年1月新发布的AlphaGeometry却招致批评。新一代投资者和行业人士对于DeepMind赢得无数赞誉的产品研发路线，开始百般挑剔。

该公司在博客中表示，AlphaGeometry系统可以解决“复杂的几何问题，其水平接近人类奥林匹克金牌得主”。但纽约大学教授欧内斯特·戴维斯（Ernest Davis）认为，该博客隐瞒了AlphaGeometry的局限性，比如它只能处理二维几何形状，无法理解什么是面积。

AlphaGeometry创意图，图片来源：DeepMind官网

DeepMind早期投资者之一弗兰克·米汉（Frank Meehan）表示，此类争议一直在分散哈萨比斯对AGI研究的注意力，这可能是“挫败感持续的根源”。他讽刺道，OpenAI正在“根据文本提示生成令人难以置信的视频，而谷歌则在一些图像问题上忙得团团转”。

同时，人员流失还在继续。

2023年5月，Mistral AI联合创始人亚瑟· 门施（Arthur Mensch）离开DeepMind。不到一年时间里，这家公司已成为了OpenAI“最强大的欧洲对手”。他曾对外透露，离开DeepMind的原因是这家公司“不够创新”。

据彭博新闻社报道，2024年1月，劳伦特·西弗雷（Laurent Sifre）和卡尔· 图伊斯（Karl Tuyls）也选择离开DeepMind，计划在巴黎成立全新的AI公司。暂命名为Holistic的新公司可能将融资2亿欧元。

四、是蓄水池，还是海洋？

目前来看，DeepMind吸引工程研发类人才的最大机会，还要靠谷歌所赐。出于算力和数据量的原因，大公司在人才争夺战中的吸引力正在逐步加强。

为了更有效地吸引顶尖AI人才，Meta正推出史无前例的大招，比如扎克伯格亲自写邮件挖墙脚，成功让DeepMind前研究员迈克·瓦尔科（Michael Valko）加入Meta。消息人士透露，扎克伯格在信中强调了人工智能对Meta发展的重要性，并表达了强烈的与收信者共同努力的愿望。此外Meta还提出了免面试的“绿色通道”。

Perplexity首席执行官阿拉温德·斯里尼瓦斯（Aravind Srinivas）在一档播客节目上表示，他无法挖走Meta的顶尖AI人才，因为他的公司还没有足够的GPU，“我试图从 Meta 聘请一位非常资深的研究员，你知道他说什么吗？‘当你拥有 10000 个 H100时再来找我吧。’”

与此同时，DeepMind创始人苏莱曼带着Inflection的大批研究人员加入微软，他将领导新成立的微软AI部门，主管Copilot等消费级AI产品。

图片来源：苏莱曼的X主页

此举也让微软顺理成章地官宣了伦敦办事处，承诺未来三年内投资25亿英镑，由前DeepMind和Inflection成员乔丹·赫夫曼（Jordan Huffman）主导。

在争夺顶尖人工智能人才的斗争中，微软已经站到了DeepMind的护城河前。

“人们愿意加入微软（或谷歌、苹果）这样的公司是为了对现实世界产生影响。”剑桥大学计算机教授乔恩·克洛克夫特告诉「甲子光年」，“而且，也许这些财力雄厚的大公司雇佣他们只是为了停止这场雇佣他们的竞争，谁知道呢？”

平心而论，谷歌帮助DeepMind做到了自己该做的一切。

今年1月，谷歌已效仿OpenAI，动用特殊的股票池，为DeepMind人才提供上百万美元的股票激励。4月9日，谷歌也宣布推出基于Arm架构的数据中心芯片Axion，渴望摆脱对英伟达的依赖。

另一方面，应投资者需求，谷歌今年已经在1月和4月官宣两轮大裁员，而且明确表明原因不是公司的财务困境，而是谷歌在将人工智能置于优先地位时“不得不做出的艰难决策”。1月的裁员主要涉及广告销售部门，因为AI技术已经与广告业务深度融合。

谷歌联合创始人谢尔盖·布林（Sergey Brin）也亲自打电话给一位准备离开去 OpenAI的DeepMind研究员，并说服他留了下来。

但谷歌面临的现实环境是，顶尖AI人才无论是自己创业，还是加入其他资金雄厚的AI团队，他们的选择要比谷歌在整整10年前收购DeepMind时多得多。

同时，公司内部的混乱也还拦住了谷歌前行的脚步，最新的AI团队整顿，效果难以超出过去一年Google Brain与DeepMind的整合。

“谷歌在人工智能方面非常强大，但它在该领域有很多不同的努力，到目前为止，整体成果小于各个部分的总和。”机器学习研究员、牛津大学教授佩德罗·多明戈斯（Pedro Domingos）表示，“谷歌需要有人让一切协调一致，更好地盈利。”

风险投资公司OpenOcean的合伙人叶卡捷琳娜·阿尔马斯克（Ekaterina Almasque）认为，DeepMind已不再是AI领域遥遥领先的绝对领导者，“所有公司都在争夺同样的AI人才库，这越来越像一个蓄水池，而不再是一片海洋。”

但可以肯定的是，DeepMind面临的是更加开放的竞争，硅谷不是AI人才们唯一的选项。

克洛克夫特对「甲子光年」坦言，现在人们更乐意在他们能负担得起的地方工作，“我不确定硅谷在人才竞争中能发挥多大作用，那里无论怎么看都太贵了。这就好比人们总是认为电影必须得在好莱坞制作，但其实绝大多数电影都并非如此。”

“人工智能人才的竞争确实非常激烈，但这只是一个周期中正常的部分。” 克洛克夫特补充道。

一切都似乎很正常，DeepMind并没有做错什么，谷歌也提供了应有的支持，浮沉背后，唯有时代巨轮驶过卷起的浪花。

*参考资料：

《The Deep Mind of Demis Hassabis》，《WIRED》

《Oxford and Cambridge are losing AI researchers to DeepMind》，《Business Insider》

《Deep Confusion: Tensions Lingered Within Google Over DeepMind》，《The Information》

《Why top AI talent is leaving Google’s DeepMind》，《Sifted》

《Why didn’t DeepMind build GPT3?》，Jonathan Godwin

《Google’s Demis Hassabis Chafes Under New AI Push》，《The Information》

作者：田思奇；编辑：王博

原文标题：谷歌全面整合AI力量背后：DeepMind浮沉史｜甲子光年

来源公众号：甲子光年（ID：jazzyear），立足中国科技创新前沿阵地，动态跟踪头部科技企业发展和传统产业技术升级案例。

本文由人人都是产品经理合作媒体 @甲子光年 授权发布，未经许可，禁止转载。

题图来自Unsplash，基于CC0协议

该文观点仅代表作者本人，人人都是产品经理平台仅提供信息存储空间服务。