YT YouTube2Text

YouTube2Text → Transcripts → 谷歌离职潮：大厂与人才，为什么会相爱相杀？

谷歌离职潮：大厂与人才，为什么会相爱相杀？ — Transcript

by 有点在李 · 2,647 words · 466 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00You're here.
0:00If one word could describe the relationship
0:03between Silicon Valley giants and their talent,
0:05I'd choose...
0:08And in Silicon Valley,
0:08which company has the most talent?
0:10There's only one answer: Google.
0:13It has had a famous nickname for years:
0:17a tribute to how much it values talent,
0:19where brilliant minds shine like stars.
0:21But over the past two months,
0:22this academy has suffered an unprecedented talent quake.
0:26Noam Shazeer, who helped invent
0:27the Transformer behind large language models;
0:29new Nobel laureate John Jumper;
0:31and Jeff Dean, Google's chief scientist of 27 years—
0:34all left Google one after another.
0:36Even its central figure,
0:37AlphaGo creator Demis Hassabis, stepped down as CEO
0:40and handed over day-to-day management.
0:41Yet just last month,
0:43Google posted record profits on paper.
0:46Quarterly revenue surged 24%,
0:48while cloud revenue soared 82%.
0:50So why, when a company is making money,
0:53are its veterans
0:53and defining figures heading for the exits?
0:58Hi, everyone. I'm Xiaoli.
0:59Today, we're not discussing office politics.
1:00We'll treat this exodus as a business case
1:03and ask why tech giants and geniuses have a love-hate relationship.
1:07Behind that tangled relationship,
1:09we can see how the AI arms race
1:12has fundamentally changed.
1:17A love-hate relationship begins, of course,
1:19with love.
1:20And at its peak,
1:21Google's devotion to talent stunned Silicon Valley.
1:24Some even saw it as an absurdly generous pushover.
1:27In 2012, for example,
1:29Google used 1,000 servers
1:30and linked 16,000 CPU cores into one giant machine.
1:34It ran for three straight days
1:35just to teach AI to recognize cats.
1:38At another company,
1:39the proposal alone might have triggered questions:
1:41“What's the business model for watching cat videos?”
1:44“When will we see a return?”
1:45“What growth flywheel could cats possibly create?”
1:48But at Google,
1:49they were determined to find out:
1:50without anyone teaching the AI,
1:52could it learn to understand the world itself?
1:55Talk about doing whatever you want.
1:56Then came the Go showdown of 2016
1:58that stunned the world.
1:59Hassabis had promised a $1 million prize.
2:02But after the match ended,
2:03they realized the budget hadn't been approved.
2:05At another company,
2:06they'd be told the proper procedures weren't followed.
2:08But at Google,
2:09asking forgiveness instead of permission was fine.
2:10Google simply paid up.
2:12That's thinking big.
2:13Playing Go,
2:14folding proteins,
2:15watching cat videos,
2:16or researching things with no profit in sight for a decade—
2:19elsewhere, that was goofing off,
2:20or just wasting research money.
2:21At Google back then,
2:23that was the job.
2:25Google's devotion to talent
2:27was already visible in its 2004 IPO filing.
2:29There, you can see the first clues.
2:30Co-founders Larry Page and Sergey Brin
2:32wrote in their letter to shareholders:
2:34“Google is not a conventional company.”
2:35It would sacrifice short-term results for long-term opportunities.
2:38They wanted employees to trust
2:39that their hard work,
2:40love, and creativity
2:42would be cared for by a company focused on the long term.
2:46Honestly, it sounds a little like corporate hype.
2:48But Google kept making
2:49those promises for over 20 years—
2:50and kept delivering.
2:52Its overall strategy was essentially:
2:57And the tagline?
3:01With that strategy,
3:02how could Google not attract talent?
3:05Around 2017,
3:06Google had two world-class AI dream teams at once.
3:10One was based in London:
3:11DeepMind,
3:12led by Demis Hassabis.
3:13It built AlphaGo, which defeated Lee Sedol in 2016.
3:16That was their creation.
3:17Later, they solved protein structure prediction
3:19and earned Hassabis and Jumper a shared Nobel Prize in Chemistry
3:22with AlphaFold, another DeepMind creation.
3:24The other team was based in Mountain View:
3:27Google Brain,
3:28led by Jeff Dean.
3:28In 2017, Google's research machine
3:30produced the legendary paper
3:30that introduced the Transformer—
3:31another breakthrough from inside Google.
3:33That's the “T” in ChatGPT.
3:35These two powerhouses coexisted
3:36until 2023.
3:38That year, Google Brain and DeepMind
3:39officially merged
3:40to form today's...
3:42Two years later, it launched Gemini
3:43and mounted a comeback against GPT.
3:45But here's something you may not know.
3:47Google's three true cash cows are Search,
3:50YouTube, and Cloud—
3:51with Search by far the biggest.
3:53So Google was effectively
3:56the world's most lavish lab for scientific geniuses,
3:57bankrolled by the money printer of search advertising.
4:00And its geniuses had just one KPI:
4:04For a very long time,
4:06Google took “to love you is to set you free” to the extreme.
4:10But that love also planted a time bomb.
4:13When geniuses raised on freedom
4:15are suddenly told to work the assembly line,
4:17conflict is inevitable.
4:22The AI world
4:22has had several defining moments.
4:24The first was
4:25the ChatGPT moment in late 2022.
4:28For the first time, someone had built
4:29a chatbot that could converse fluently.
4:31But actually,
4:31back in 2020,
4:33Google Brain already had
4:34similar technology.
4:35One of its creators was Noam Shazeer.
4:38But Google's leadership feared it would make things up
4:40and tarnish the reputation of Search,
4:42so they refused to launch it.
4:43Frustrated, Shazeer
4:45left to start a company in 2021.
4:48That was Google's first clash with a genius.
4:50Then, in late 2022,
4:52ChatGPT took the world by storm,
4:53built on the very foundation
4:55Google had invented: the Transformer.
4:56Google was stunned.
4:57“How did someone else turn our recipe into a global hit?”
5:00Even more frustrating,
5:01Google already had a similar product in its pocket.
5:04Then the famous...
5:06...sounded inside Google.
5:08It rushed out Bard to fight back.
5:10But Bard gave a wrong answer in its public debut.
5:12Over $100 billion in market value vanished in a day.
5:14Stung by the setback, Google prepared a major response.
5:16In April 2023,
5:18Sundar Pichai issued an order:
5:19merge the rival dream teams in London and Mountain View
5:21into one,
5:22led by Hassabis.
5:24The name “Gemini,” or twins,
5:25symbolized the union of the two teams.
5:27In 2024, Google spent another $2.7 billion
5:29to bring Shazeer back into a leading role.
5:32Just a modest 20 billion yuan or so.
5:34It may be the costliest reunion in tech history.
5:37Notice that, up to this point,
5:38the love remained.
5:39It was simply expressed through checks instead of freedom.
5:44With money and talent piled in,
5:46Gemini overtook the field by late 2025
5:48and won widespread acclaim.
5:48Google had forced its way back to AI's top table.
5:51It should have been a perfect comeback story.
5:52But before Google could even
5:54settle into its seat,
5:54its talent began leaving one after another.
5:56Why?
5:58We need to pull the camera back
6:00and look at the AI arms race as a whole.
6:03Its center of gravity has shifted four times.
6:08At first, one brilliant mind could build a unicorn.
6:11Ilya Sutskever left Google
6:12and became the soul of OpenAI.
6:13Dario Amodei then left OpenAI
6:15and built a company that could challenge it head-on:
6:17Anthropic.
6:18Wherever the genius went,
6:19capital and attention followed.
6:23Then model parameters began exploding.
6:24Brainpower alone was no longer enough.
6:26What mattered was who had more chips.
6:28The biggest winner was Nvidia, selling the shovels.
6:33As foundation models converged,
6:34the race became about who was closest to users.
6:36Google put Gemini into Search,
6:37YouTube, and its entire productivity suite.
6:39China followed the same playbook:
6:40ByteDance put Doubao into Feishu,
6:41and Alibaba put Qwen into DingTalk.
6:43Claim the position first.
6:44Today,
6:45the battle has entered its fourth round.
6:49In two years, the tech giants have spent
6:51more than $1 trillion combined.
6:53But the cash returns have lagged far behind.
6:55Fear of losing talent gave way to fear of failing to monetize.
6:58Google's Q2 2026 results
7:00made that perfectly clear.
7:01Quarterly capital spending hit $44.9 billion,
7:04double the year before.
7:05That spending produced Google's first quarter
7:07of negative free cash flow in 22 years as a public company.
7:10Facing brutal pressure to monetize,
7:12Google sharply changed direction:
7:13race to scale on one side,
7:15and squeeze costs on the other.
7:16On scale,
7:17the CFO laid it out for shareholders:
7:19Google would raise its full-year
7:20spending budget to roughly $200 billion,
7:23racing to build computing infrastructure.
7:24On costs,
7:26here's the key point.
7:27In Pichai's own words,
7:28intense optimization had pushed
7:29the cost of generating each AI answer in Search
7:31to its lowest level since launch.
7:34Once, Google could use 16,000 CPU cores
7:37to teach AI to recognize cats.
7:39Now, training Gemini, generating AI answers in Search,
7:41and serving cloud customers
7:43all compete for the same TPUs.
7:45Bloomberg reported that
7:46before Shazeer left, computing power for his project
7:48was reassigned
7:50to another DeepMind team in London
7:51to unify and accelerate pretraining.
7:54And since
7:54Google declared “Code Red,”
7:56its scientists have been operating in battle mode.
7:59They once asked which questions were worth exploring.
8:02Now they're asked when the next demo will ship,
8:04and which benchmarks Gemini is still losing.
8:07Is that wrong?
8:08Absolutely not.
8:09A public company must answer to its shareholders.
8:11But consider the disconnect.
8:12When Google recruited you,
8:14it asked you to play Go,
8:15fold proteins,
8:16and promised you freedom.
8:17Now meetings are about
8:18shaving another cent off each query,
8:20or squeezing the model into one more screen.
8:23Geniuses crave the romance of compute and creative freedom.
8:26Earnings reports couldn't care less.
8:28So Jeff Dean left.
8:29His new company is Discovery Loop,
8:30which aims to automate scientific discovery with AI.
8:33Hassabis handed over management of the LLM business
8:36and returned to the lab,
8:37using AI to fight disease and save lives.
8:39Others gradually moved to
8:41companies still willing to fund frontier research:
8:42OpenAI and Anthropic.
8:44Shazeer's story is the most dramatic.
8:46He was held back,
8:47left, and was bought back for $2.7 billion.
8:49Now he has left again.
8:50He alone has lived this love-hate story twice.
8:54What failed to win his heart
8:55was never the $2.7 billion check.
8:57It was one sentence: “You've changed.”
9:01But the story doesn't have a bad ending yet.
9:03Let's step inside Google when Dean announced his departure
9:06and see what it felt like.
9:07At 8:30 that morning,
9:08he told more than 100 close colleagues
9:10that he was leaving.
9:12Some cried on the spot.
9:13The group chat filled with hearts and sad emojis.
9:16At 2:30 p.m.,
9:17Google held a farewell event.
9:19A café built for just 300 people
9:21was packed with 1,500.
9:23At 10:30 that night,
9:24he found more than 400 farewell messages waiting.
9:27So Google's chief scientist of 27 years
9:30replied to them one by one
9:31until 2:30 a.m.
9:33Does this warm scene
9:34look like a bitter breakup?
9:35At 5:20 p.m. on his final day,
9:38he closed his Google laptop.
9:39At 11:59 p.m.,
9:41he changed his status to “unemployed.”
9:43One minute later,
9:44he became a co-founder
9:46of Discovery Loop.
9:49Three of Dean's protégés
9:50left with him.
9:54Put their résumés together,
9:55and you get a condensed history of Google's technology.
9:58From Search infrastructure like MapReduce,
10:00Bigtable,
10:01and Spanner,
10:02to TensorFlow,
10:03TPUs,
10:04AlphaFold,
10:06and Gemini—
10:07their names appear across nearly every project.
10:10And their new company
10:11is called Discovery Loop.
10:13They want AI to propose experiments,
10:16carry them out,
10:17analyze the results,
10:18and design the next experiment based on those results.
10:21First, use the system to build better AI.
10:22Then expand it into drugs,
10:24chips, materials, and other scientific fields.
10:26In short,
10:27they want to build a tireless,
10:29endlessly iterating automated scientist.
10:31Here's the wildest part:
10:32when the company was announced,
10:33they hadn't even leased an office.
10:35They hadn't begun hiring a full team.
10:36A reporter asked, “Which of you is the CEO?”
10:38They looked at one another,
10:39then all pointed to Dean.
10:41Their startup pitch deck was just
10:43a few résumés and a rough blueprint.
10:45Investors still came flooding in.
10:47They said that with a team like this,
10:49they were ready to bet before due diligence even began.
10:52But back to today's question.
10:53These are elite minds investors back almost blindly.
10:56Now they're leaving.
10:56What can Google do?
10:58First, you may wonder:
10:59did Google try to stop the breakup?
11:01According to media reports,
11:02Pichai held more than one round of talks,
11:04hoping the four would keep their Google badges.
11:07In the end, Dean chose the freedom of entrepreneurship.
11:10Once Google failed to keep them,
11:11the usual script
11:12offered only two endings.
11:13The good one: part on friendly terms.
11:15The ugly one:
11:16enforce the noncompete
11:17and meet in court.
11:18But Google wiped away its tears
11:19and offered a very different breakup agreement.
11:22Consider these terms.
11:23First:
11:24you don't want Google's paycheck anymore.
11:26You can leave,
11:27but you must let Google invest in you.
11:29Was Google simply afraid its old friends would go hungry?
11:31What it feared more was this:
11:33what if they really built a machine that could invent drugs,
11:35chips, or even the next generation of AI?
11:37Equity would still give Google a share of the future upside.
11:40Second:
11:42during Discovery Loop's first year,
11:43Google would provide its computing resources.
11:46Third:
11:47Google and the startup would establish a joint research framework
11:49for machine-learning systems and related infrastructure.
11:52In other words,
11:53the reporting line ended,
11:54but the technical relationship did not.
11:55The only real difference may be the missing Google badges.
11:58See how subtle this relationship is?
12:00It's like going from intimate roommates
12:01to strategic partners.
12:03Of course, this doesn't mean Google suddenly went soft.
12:06Google no longer insists on keeping every elite mind
12:09and every leading model locked inside the company.
12:13By building Gemini, it competes on the track.
12:15By selling TPUs
12:16and cloud services, it builds the track.
12:18By investing in teams that spin out of Google,
12:21it owns a stake in the race itself.
12:22Even if Gemini doesn't win,
12:24as long as the winner runs on Google's track,
12:26Google can still make money.
12:28Anthropic is the clearest example.
12:30Its Claude model is a direct
12:32competitor to Gemini.
12:33Yet in 2025,
12:34Anthropic announced
12:35an expanded use of Google Cloud,
12:37planning to deploy up to one million TPUs.
12:39The compute expansion was worth tens of billions of dollars.
12:41Amazon remained Anthropic's
12:43primary cloud provider.
12:44But Google could still profit from its rival's growth
12:47by selling chips and cloud capacity.
12:48If Gemini wins,
12:49Google wins the model race.
12:50If Claude wins,
12:51Google sells the compute.
12:52If Discovery Loop really
12:53builds an automated scientist,
12:54Google still has equity and a research partnership.
12:55This is the new relationship
12:57between tech giants and geniuses in industrialized AI.
12:59They no longer need to own you completely.
13:01They just want to appear on your cap table,
13:04your cloud bill, and your research partnerships.
13:06Of course, this relationship isn't as romantic as it sounds.
13:08One U.S. agency issued a sharp assessment
13:10of ties between tech giants and startups:
13:11combining equity investment with cloud services
13:13could raise a startup's switching costs.
13:15It could also give cloud providers access
13:16to technical and business information unavailable to others.
13:19In plain English:
13:20these relationships are never acts of selfless generosity.
13:27So let's return to our opening question.
13:28Why do tech giants and geniuses always love and fight?
13:31Because they never loved the same vision of the future.
13:34Geniuses want to invent it.
13:35Tech giants want to manufacture it.
13:37The lab asks, “Is it possible?”
13:39The assembly line asks, “How much per unit?”
13:41From Bell Labs to Google, this divide
13:45has never truly been bridged.
13:46But Google's breakup agreement
13:48may be the most graceful ending available today.
13:51The first time they fell in love,
13:52Google said:
13:54The second time, when they clashed,
13:55Google said:
13:57“We still need to calculate the cost of every query.”
14:00By the third time,
14:01Google had finally learned a new line:
14:02“You can leave this place.
14:06I'll provide the compute.
14:07I'll invest the money.
14:08And we'll keep doing research together.”
14:10So who stole Google's talent?
14:12No one, really.
14:13Faced with the irreversible tide of industrialized AI,
14:16the giants made the choices business empires must make.
14:19The geniuses who left may still be inventing humanity's future—
14:22just in a different way,
14:24and on a different balance sheet.
14:25Meanwhile, giants like Google, Microsoft,
14:26and Amazon still stand behind them,
14:29smiling as they offer compute and checks.
14:32The end of a love-hate relationship isn't a breakup.
14:34It's saying, “If I can't keep you,
14:35or keep your heart,
14:36then let's not burn every bridge.
14:38Our paths may cross again.”

About this transcript

This page contains the full transcript of 谷歌离职潮：大厂与人才，为什么会相爱相杀？ by 有点在李, generated from the public captions YouTube serves with the video. The transcript has 2,647 words across 466 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.