YT YouTube2Text

YouTube2Text → Transcripts → You should, unfortunately, be worried about Sam Altman.

You should, unfortunately, be worried about Sam Altman. — Transcript

by AI In Context · 6,813 words · 1,177 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00On November 17th, 2023, Sam Altman,
0:02wildly successful CEO of OpenAI, was
0:05fired.
0:05>> The godfather of ChatGPT kicked out of
0:08the company he founded.
0:10>> Employees were revolting, investors were
0:12freaking out. Nobody seems to understand
0:13why this guy was fired. And then,
0:15>> [music]
0:15>> four days later, Sam Altman is
0:17apparently back as OpenAI CEO.
0:20>> I want to tell you the full story.
0:22[music] Some of it we only know now,
0:23thanks to unsealed court testimony
0:25that's come out just this year. I think
0:26you need to understand what happened
0:28[music] those five days if you want to
0:30understand this strange moment we're
0:32living in. [music]
0:32>> AI is a technology for taking away your
0:34job and maybe eventually displacing
0:37human sovereignty over our own future.
0:39>> Where, frankly, a handful of unelected
0:41tech CEOs have more power over the
0:44future than most heads of state. [music]
0:46>> You've said AI could wipe out half of
0:48all entry-level white-collar jobs and
0:50spike unemployment to 10 to 20% in the
0:52next 1 to 5 years.
0:54>> Yes.
0:54>> And the most powerful of all those
0:56people might be Sam Altman, the CEO of
0:59OpenAI, a company founded on the idea
1:01that no one person should control
1:03artificial intelligence.
1:04>> AI will probably, like most likely sort
1:07of lead to the end of the world, but in
1:08the meantime,
1:10uh there will be great companies
1:11created.
1:12>> So, why was Sam Altman fired? And what
1:14does that tell us about who gets to
1:16decide our future?
1:25Friday, November 17th, 2023.
1:28Sam Altman is in a hotel room in Las
1:30Vegas. He's in town for the Formula 1
1:33race. It is a well-earned break. He's
1:35been spending the better part of a year
1:37doing press, traveling the world meeting
1:39global leaders ever since his company,
1:41OpenAI, released a chatbot, which
1:44you know what? I'm just going to go
1:45ahead and assume you've heard of.
1:46>> ChatGPT.
1:48>> ChatGPT is on pace to pass 100 million
1:51active monthly users after just 2
1:53months.
1:53>> Things are, by any reasonable measure,
1:56going extremely well for Sam. His
1:58company is about to be valued at $86
2:00billion.
2:02People are calling him the Oppenheimer
2:04of his generation, which he doesn't seem
2:06to mind. He sometimes mentions that he
2:07and Oppenheimer share a birthday,
2:09[music] fun fact. At noon, Sam opens his
2:11laptop for a quick video call. On the
2:13other side is OpenAI's board. Let me
2:15introduce them. We've got Adam D'Angelo,
2:18CEO and founder of Quora, you know, that
2:20the website with the
2:22questions, you know the one.
2:24Helen Toner, AI policy researcher, day
2:26job at a Georgetown think tank. We've
2:28got Tasha McCauley, robotics
2:30entrepreneur. And Ilya Sutskever,
2:33OpenAI's chief scientist. Widely
2:35regarded as one of the most brilliant AI
2:37researchers around. Also a co-founder of
2:40the company, also Sam's friend. Sam was
2:43expecting a routine call.
2:45It is not a routine call.
2:47Sutskever reads from a prepared
2:49statement.
2:51"The board has," he says, "lost
2:53confidence in Sam's ability to lead the
2:55company. They are terminating his
2:57employment effective immediately. There
2:59will be a public announcement shortly."
3:02Sam is too stunned to speak. He feels
3:05like he's in a bad dream.
3:07After the call, Sam is locked out of his
3:09laptop.
3:1020 minutes ago, he was CEO of the most
3:12famous AI company on Earth.
3:15And now he can't get into his own Gmail.
3:19Now, minutes later, the board puts out a
3:21public statement. It reads,
3:23"Mr. Altman's departure follows a
3:24deliberative review process by the
3:26board, which concluded that he was not
3:29consistently candid in his
3:30communications with the board, hindering
3:33its ability to exercise its
3:34responsibilities." The board also names
3:36an interim CEO in the announcement, Mira
3:39Murati, OpenAI's chief technology
3:41officer.
3:42Within minutes, this firing is one of
3:44the biggest news stories in the world.
3:47Sam's phone is getting so many texts per
3:49second his iMessage app crashes.
3:51300 miles away at OpenAI's San Francisco
3:54offices,
3:55hundreds of people are standing around
3:57staring at their phones refreshing the
3:59same four tabs. Nobody is working.
4:01Nobody can work. Their CEO was just
4:03fired 20 minutes ago and nobody will
4:05tell them why. Also caught off guard is
4:08Microsoft, OpenAI's most important
4:10investor. Murati had called them mere
4:12minutes before the public announcement
4:14went out. Satya Nadella, the CEO of
4:16Microsoft, [music] is to use the
4:18technical term,
4:19pissed.
4:20Sutskever and Murati [music] hold a
4:22company-wide meeting to try to explain
4:24what's happened.
4:26Someone asks if there was a specific
4:28incident or something and Sutskever
4:30replies, "I wish I could go into
4:31details, but I can't."
4:33Someone asks if they'll ever find out
4:35more. Sutskever says, "Keep your
4:37expectations low."
4:39Then, someone asks about the tender
4:41offer.
4:42Now, this wasn't mentioned much in
4:43reporting at the time, but in just a few
4:45days OpenAI was due to close a deal that
4:47would let employees sell their own
4:49shares at that $86 billion valuation.
4:52For a lot of employees, that meant
4:54millions of dollars. Some people had
4:56already put down deposits on houses.
4:59So, Brad Lightcap, the COO, has to take
5:01this question.
5:03And he says,
5:04"The tender, we're um we're going to
5:06see."
5:08None of this, it turns out, is
5:09particularly reassuring.
5:11>> That whole weekend was one of the most
5:12stressful uh periods in my life, I would
5:14say. It was like a civil war within the
5:16company.
5:16>> Around 2:00 p.m., Sam puts out a tweet
5:19in his trademark all lowercase. He says,
5:23"I loved my time at OpenAI. It was
5:25transformative for me personally and
5:27hopefully the world a little bit. Most
5:28of all, I loved working with such
5:30talented people."
5:32Greg Brockman, who is OpenAI's
5:34president, also a co-founder, also a
5:36board member until the same four people
5:38removed him that same day, has been told
5:41that he can stay on as OpenAI's
5:42president.
5:44Instead, he retweets Sam and announces
5:46that he's quitting in solidarity.
5:48Three more senior researchers quit as
5:50well. Employees feel like the sky is
5:53falling.
5:54Sam, for his part, gets on a flight back
5:56to San Francisco.
5:57Late Friday evening, the board takes a
6:00meeting with OpenAI's senior executive
6:02team. So, not the whole company this
6:03time, just the 15 or so most senior
6:05employees.
6:07The board keeps citing confidentiality
6:09and just continues to avoid answering
6:11any questions in detail. [music] And the
6:12execs are furious. They say the board
6:15has no choice but to reinstate Altman,
6:17otherwise they'd be letting the company
6:18die. Helen Toner replies, "That could in
6:21fact be consistent with the mission."
6:25At first glance, that sounds completely
6:26insane.
6:27Why would a board fire a wildly
6:29successful CEO and call it consistent
6:32with the mission?
6:38OpenAI was founded in 2015, not as a
6:41normal tech company, as a nonprofit
6:44research lab. It was funded by donations
6:46from the likes of Elon Musk, also a
6:48co-founder, Peter Thiel, and Sam Altman
6:50himself.
6:51Ilya Sutskever was there from day one.
6:54And the mission was not to make money.
6:56It was very explicitly to save the
6:59world. [music]
7:00Easy to roll your eyes at, but here's
7:02what these people at least claimed to
7:03believe. That within their lifetimes,
7:06possibly within the next decade or two,
7:08humans would build machines smarter than
7:10themselves. Machines that could do new
7:12scientific research, could run a
7:14military strategy, just do any cognitive
7:16work that humans can do.
7:18They had a name for this.
7:20AGI.
7:20>> AGI, artificial general intelligence.
7:24>> And they believed that whoever built AGI
7:26first would have in their hands the most
7:28powerful technology to have ever existed
7:31on Earth.
7:32And they did not think that this was a
7:34safe enterprise.
7:36This is Sam Altman writing on his
7:37personal blog back in 2015.
7:40Development of superhuman machine
7:41intelligence is probably the greatest
7:43threat to the continued existence of
7:46humanity. In that blog post, Sam
7:48recommends this book, Superintelligence.
7:52And the book's argument is basically
7:53this. Humans run this planet not because
7:56we are the strongest species, but
7:57because we're the smartest. You build
7:59something smarter than humans, then the
8:01same cold logic applies.
8:03>> It's smarter than us, right? There's no
8:04way we're going to prevent it getting
8:05rid of us if it wants to. If you want to
8:07know what life's like when you're not
8:09the apex intelligence,
8:11ask a chicken.
8:12>> That's the position that OpenAI's
8:14founders thought humanity was about to
8:16be in. So, they set up this research lab
8:18to build AGI themselves and build it
8:21safely before someone less careful
8:23builds it first. They specifically do
8:24not trust Google DeepMind.
8:26>> The reason I wanted to create OpenAI was
8:28to serve as a counterweight to Google.
8:29The furthest thing from Google would be
8:31like a nonprofit that is fully open cuz
8:34Google was closed, for-profit. We don't
8:36want this to be a sort of profit
8:38maximizing
8:39>> a behemoth from hell, you know?
8:41>> [laughter]
8:41>> That just never stops.
8:43>> OpenAI's original mission, and you can
8:44still read this online, was to advance
8:46digital intelligence in the way most
8:48likely to benefit humanity as a whole,
8:50unconstrained by the need to generate
8:52[music] financial returns. These are not
8:54normal startup commitments. But, here's
8:56the problem.
8:57Training cutting-edge AI turns out to
8:59cost an astronomical amount of money.
9:01>> Like, I think one thing that no one
9:03knows yet is just how much money you
9:04need, you know, billions for sure, tens
9:06of billions, maybe hundreds of billions,
9:07I sure hope not.
9:08>> So, in 2019, Sam Altman comes up with a
9:12fix, one that happens to land him as CEO
9:14of the company.
9:16The idea is this.
9:17A nonprofit board of directors that owns
9:20and controls a for-profit subsidiary.
9:23So, the for-profit can take investment,
9:26it can pay [music] competitive salaries
9:27including startup equity, it can build
9:29data centers, it can compete with
9:31Google. But,
9:33it answers to the [music] nonprofit, and
9:35the nonprofit answers to the mission.
9:38There are two guardrails built into this
9:40structure.
9:41Guardrail number one, investor profits
9:44are capped at 100x.
9:46Beyond that point, everything flows back
9:48to the nonprofit [music] to be used for
9:50the benefit of humanity for something
9:52like universal basic income.
9:54The idea was simple. Investors can get
9:56rich, but not infinitely rich.
9:58Past a point, the upside belongs to
10:00everyone.
10:00>> If you wanted to just make money, that
10:02would not be a good way to do that,
10:03right? Like having a nonprofit have that
10:05level of control over your company. They
10:07could have just started over and started
10:09a for-profit company, right? And they
10:10chose not to do that.
10:12>> Guardrail number two, the nonprofit
10:14board can fire the for-profit's CEO.
10:16This is the kill switch.
10:18The idea is you put people on the board
10:21who care about not ending the world. You
10:23let the capitalist machine run
10:25underneath them. It's the only way to
10:26compete. But, if it starts racing
10:28somewhere dangerous, the people upstairs
10:30can pull the plug.
10:32On paper, this board puts humanity above
10:35the company. If the two come into
10:37conflict, humanity is supposed to win.
10:40Sam Altman talks about this constantly.
10:42>> Like no one person should be trusted
10:44here. I don't have super voting shares.
10:47Um,
10:48like I don't want them. The board can
10:50fire me. I think that's important.
10:52>> Now,
10:53this is what OpenAI looked like as of
10:55November 2023 after the release of
10:57ChatGPT, the fastest growing tech
10:59product in history, and $13 worth of
11:02investment from Microsoft.
11:04In theory, the nonprofit is in charge.
11:06In practice, you have a handful of
11:09volunteers overseeing an $86 billion
11:11tech company.
11:18Saturday morning, November 18th, Sam
11:20wakes up in his home in San Francisco.
11:23So,
11:24what do you need if you've just been
11:26fired by one of the most important
11:27companies in tech and you'd ideally like
11:29your job back?
11:30This, to be clear, is not career advice.
11:34Step one, open a line of communication.
11:36On Saturday morning, Sam gets a call
11:38from the board that just fired him.
11:41They want to talk after the disastrous
11:43reaction the previous day from employees
11:45and senior executives. They're right to
11:47be worried, by the way. In a few hours,
11:49the executives are going to give the
11:50board a 5:00 p.m. deadline to reinstate
11:52Sam and then resign.
11:55Step two, assemble your allies. Sam
11:57starts turning his home into a war room,
12:00or to use his words, a government in
12:02exile.
12:03Within hours, dozens of people are
12:05lining up in Sam's home, including
12:07OpenAI employees and senior executives.
12:10Lawyers set up an office outside his
12:12bedroom. Mira Murati is there, which is
12:14pretty weird considering she's just
12:16replaced him as CEO. Sam, for his part,
12:19works the phones. According to his phone
12:21records, he's on calls for 12 hours a
12:23day.
12:24Step three,
12:25stay relaxed.
12:276:00 p.m. rolls around. The 5:00 p.m.
12:28deadline has come and gone. The board
12:30has neither reinstated Sam nor resigned.
12:32Sam walks out to his living room and
12:34starts handing out a round of Negronis.
12:36"You need to relax," he tells people.
12:38"Whatever's going to happen is going to
12:39happen." Step four, make a show of
12:42strength.
12:43On Saturday night, Sam sends out a
12:45single tweet. "I love the OpenAI team so
12:48much."
12:49Within minutes, dozens of OpenAI
12:51employees are retweeting it with heart
12:52emojis.
12:55Step five, set the terms of the fight.
12:58Now, Sam has this guy Chris Lehane on
13:00the line. He's a political operative
13:02whose personal motto is apparently
13:04"Everyone has a plan until you punch
13:06them in the mouth."
13:08Lehane, maybe unsurprisingly, wants Sam
13:10to go scorched earth on the board. Drag
13:13them in the press, destroy them.
13:15Sam, though, plays it differently. He
13:18steps back and lets his supporters and
13:20allies build his case for him.
13:22They start feeding sympathetic
13:24journalists just enough to shape the
13:26narrative in the press. The board
13:27doesn't know what they're doing. The
13:29board is in over their heads. The board
13:30won't even say why they did it. Within
13:3224 hours, the narrative isn't Sam Altman
13:35was fired. The narrative is, what the
13:37hell was the board thinking?
13:39Which
13:40actually is a good question.
13:42What was the board thinking?
13:51In 2003, a teenager shows up at the
13:54University of Toronto and knocks on the
13:56door of Professor Geoffrey Hinton, a
13:58future godfather of AI.
14:00That teenager's name is Ilya Sutskever.
14:03And he's interested in Hinton's work on
14:05artificial neural networks, an out of
14:07fashion approach to AI that's [music]
14:09loosely inspired by biological brains.
14:11So, Hinton gives Ilya an old paper. A
14:14few days later, the teenage Ilya comes
14:16back with an insight it had taken the
14:17field years to produce.
14:20For most of his life, Ilya Sutskever has
14:22seen things in AI before other people
14:24see them.
14:26Fast forward to 2012.
14:27>> [music]
14:28>> Ilya, Hinton, and a colleague have been
14:29researching deep learning, basically
14:31just making neural networks way bigger
14:33than had been tried before.
14:35At the time, mainstream AI researchers
14:37have been struggling for years [music]
14:38to get computers to correctly label
14:40images. The best in class programs are
14:42wrong about 25% of the time.
14:44Ilya and Hinton debut a network that
14:47demolishes that world record.
14:50In 2013, a young startup called DeepMind
14:53uses the same type of deep neural
14:55network and unveils an AI system that
14:57can top human experts at multiple
14:59different classic video games. In 2015,
15:02Elon Musk and a charismatic guy named
15:04Sam Altman take Ilya out to dinner.
15:07They don't trust DeepMind.
15:09They want to start a new AI lab. They
15:10want to save the world. They want Ilya
15:13to be a part of it.
15:15Ilya turns down a $6 million a year,
15:17"Please don't leave us" counter offer
15:19from his employer Google
15:21and joins them.
15:23In 2016, DeepMind shocks the world with
15:26an AI system that beats one of the
15:28greatest human masters of the board game
15:30Go.
15:31Next year, they try using 100 times more
15:34computer chips, and instead of teaching
15:36this AI system Go with examples, they
15:38just let it start playing itself. In 8
15:40hours, it is better at the game of Go
15:42than any person or program that has ever
15:44played. By this time, Ilya is convinced
15:47that the path to AGI might be as simple
15:49as finding the right neural network
15:50design and then scaling it up.
15:53And in 2017, 18 months into leading
15:55research at OpenAI,
15:57he reads a paper from Google that makes
16:00him jump out of his chair, run down the
16:02hall, and tell his colleagues, "Everyone
16:04stop what you're doing. This is it."
16:07Over the next 5 [music] years, OpenAI is
16:09going to make a series of exponentially
16:11larger bets that prove Ilya's conviction
16:14right.
16:16They train GPT-2 and discover that it
16:18has abilities they hadn't specifically
16:20trained it for, [music]
16:21like being able to translate text into
16:22new languages.
16:24The text output from GPT-3.5
16:26is so human-like that they wrap it
16:28[music] in a user-friendly interface and
16:30try releasing it as a chatbot.
16:32It hits 100 million users in 2 months.
16:37By this time, Ilya has developed a
16:38slogan.
16:40He says it around the office. A couple
16:41weeks after the ChatGPT [music] release
16:43at the company's holiday party, he gets
16:45his colleagues to chant it back to him.
16:48Feel the AGI.
16:50He wants them to take [music] it
16:51seriously.
16:52To take seriously that the progress
16:54might not stop. That this curve they're
16:57riding, a curve that started
16:58>> [music]
16:58>> way back in 1956 with a few researchers
17:01in New Hampshire dreaming about thinking
17:02machines,
17:03might just keep going up.
17:05Soon, AI systems might match human
17:08intelligence and then surpass it and
17:10then vastly surpass it.
17:12And then,
17:14the person in control of the technology
17:16will have their hands on something
17:18unprecedentedly powerful.
17:20Something they may not be able to
17:21control at all.
17:25>> Scientists have been accused of playing
17:27God for a while.
17:29But,
17:31there is a real sense in which we are
17:33creating something
17:35very different from anything you've
17:37created so far.
17:38>> At the 2022 retreat,
17:40Ilya and his colleagues are hanging
17:42around a campfire and he's got this
17:43wooden effigy in his hand that he
17:45commissioned from a local artist.
17:47The effigy, he explains, represents an
17:49AI that is pretending to be nice, but is
17:52actually been lying to them.
17:54Ilya douses it in lighter fluid and sets
17:57it on fire.
17:59Around the same time, Ilya is starting
18:01to get very worried about Sam.
18:09Sunday morning, November 19th, 2023.
18:12Less than 48 hours since Sam Altman was
18:14fired,
18:15and he's heading into the office.
18:17Interim CEO Mira Murati has invited Sam
18:20to come by for the day. So, Sam goes to
18:21the front desk, gets a visitor's badge,
18:24and because he's Sam, immediately posts
18:26a selfie with the caption, "First and
18:28last time I wear one of these."
18:30Negotiations go on for hour after hour
18:33with the board. Journalists are camped
18:35out outside the OpenAI offices tracking
18:37the steady stream of takeout deliveries.
18:39No one involved has slept much in the
18:41last 48 hours, but Sam feels good. The
18:44board is losing in the court of public
18:46opinion. They've definitely lost with
18:47the employees, [music]
18:48and they're even open, it seems, to
18:50bringing him back as CEO.
18:52There's haggling going on over who would
18:54need to resign, who would make what
18:55statements.
18:56It really seems like the board is out of
18:58cards and about ready to fold.
19:01Or are they?
19:03See, all week at the board has been
19:05putting out feelers to anyone in their
19:06network who might be willing to come in
19:08as interim CEO.
19:10At one point, they reach out to Dario
19:12Amodei, the CEO of OpenAI's bitter
19:14rival, Anthropic, to ask if he wants to
19:17be CEO, maybe even combine the
19:18companies.
19:19He's not interested.
19:21But, Sunday evening, late in the night
19:24with hundreds of employees at the office
19:26expecting an end to the drama, the board
19:28makes a sudden announcement.
19:30Murati has been replaced.
19:33There's a new CEO.
19:34Enter
19:36Emmett Shear.
19:38Co-founder of Twitch, the
19:40website with the videos and the the
19:42game, you know the one. Sam is shocked.
19:45He feels like the board has been
19:46negotiating in bad faith all weekend.
19:48OpenAI employees flood out of the
19:50building. Sam goes [music] home in
19:52disgust.
19:53But, Sam Altman being Sam Altman, he
19:56already has another move lined up.
19:59The previous day, Microsoft CEO Satya
20:02Nadella had asked Sam if he wanted to
20:03join Microsoft.
20:05Now, Sam tells him he's [music] in. At 7
20:08minutes before midnight on Sunday,
20:11Microsoft CEO Satya Nadella tweets that
20:14Sam, along with his number two, Greg
20:16Brockman, will be joining Microsoft.
20:19Word spreads very quickly through
20:20employees' signal chats. Anyone who
20:22wants to join them will have a
20:23guaranteed job at Microsoft. Now, think
20:27about how this lands with the employees.
20:30They have been banking on this tender
20:31offer that now looks dead in the water.
20:34The board has just fired their leader,
20:36the guy that all the investors love,
20:38who's now been immediately snapped up by
20:40a much larger company, which they are
20:41now being invited to join.
20:43And they still have no idea why this is
20:46happening.
20:47All they've been given is that one
20:48phrase, "Not consistently candid."
20:52I think it's time to tell you what that
20:53means.
20:58A few weeks before the firing, a
21:00terrified Ilya Sutskever sends a
21:02disappearing email to three of his
21:03fellow board members, Tasha McCauley,
21:06Helen Toner, and Adam D'Angelo. [music]
21:08It contains a 70-page two-part memo. The
21:12screenshots inside have been compiled
21:13covertly over the course of months.
21:15There are Slack messages photographed
21:17with personal phones to avoid
21:18screenshots on company devices.
21:21The first line of the memo reads,
21:23>> [music]
21:23>> "Sam exhibits a consistent pattern of
21:25lying, undermining his exacts, and
21:27pitting his exacts against [music] one
21:29another." This forms the centerpiece of
21:31the case that Ilya Sutskever, the man
21:33who spent a decade [music] telling the
21:34world that superintelligence is coming,
21:37is starting to form against his
21:38co-founder and CEO, Sam Altman.
21:41Let's go through it. Exhibit A,
21:44the lying.
21:45Many of those screenshots came from one
21:48person, Mira Murati,
21:50the board's eventual pick as interim
21:52CEO. Back in October, on multiple calls
21:55with the board, Murati had explained in
21:57detail Altman's toxic leadership style,
21:59>> [music]
21:59>> a pattern of behavior going back years.
22:02If someone resisted his decisions, he
22:03would tell them whatever he thought they
22:05wanted to hear. If that didn't work,
22:07then he would undermine their
22:08credibility and turn people against each
22:10other. So, this is why the board was so
22:12cagey during the firing. They'd been
22:14told by outside legal counsel that they
22:16couldn't reveal this information without
22:19revealing Murati as the source and
22:21putting themselves at serious legal
22:23risk. Exhibit B, the safety lapses.
22:26At a board meeting in December 2022, Sam
22:30told the directors that three
22:31controversial changes to GPT-4 had all
22:34been approved by the company's joint
22:35safety review board. Helen Toner asked
22:38for documentation.
22:40It didn't exist. It turns out only one
22:42of the changes had been approved.
22:44Around the same time, Microsoft [music]
22:46started testing the still unreleased
22:48GPT-4 in India without any safety
22:50approval.
22:51The board found out about this by
22:53accident from a conversation in a
22:55hallway.
22:57Later, Sam sends a Slack message to the
22:59company.
23:00He says that the company's lawyer
23:02decided GPT-4 Turbo also doesn't need
23:05joint safety review.
23:08Mira Murati checks with the lawyer.
23:11He says
23:12he never said that.
23:13>> On multiple occasions, he gave us
23:15inaccurate information about the small
23:17number of formal safety processes that
23:19the company did have in place.
23:21>> It really seems like the man who thought
23:23he was building super intelligence
23:26was lying to the board about how he was
23:28keeping it safe.
23:30Exhibit C, the fund.
23:32Sam Altman had also told the board and
23:34the world that he was an independent
23:36director of OpenAI with no financial
23:38stake in the company. It turned out that
23:39the OpenAI Startup Fund, $175 million of
23:43venture capital deployed in OpenAI's
23:44name, was personally owned by Sam
23:47Altman. It does now seem like this was
23:49mostly done to make paperwork easier
23:51rather than for Sam's personal financial
23:53gain.
23:54But it took months for the full
23:55situation to be clarified to the board
23:57after they raised the issue.
23:59Exhibit D, the paper.
24:02Helen Toner authors an academic paper
24:04buried [music] within it some very mild
24:06criticism of OpenAI.
24:08Altman is furious.
24:10He goes to Sutskever personally, says
24:12that Tasha McCauley agrees Helen Toner
24:15needs to be removed from the board.
24:17McCauley had said no such thing.
24:20>> Not trusting the word of of the CEO who
24:23is your main conduit to the company,
24:24your main source of information about
24:25the company, it's just like totally
24:27totally impossible.
24:28>> It turned out that the board had been
24:29deadlocked for months over filling
24:31multiple vacancies. Sam wanted his
24:33picks. Interviews were on the agenda for
24:35the November board meeting just [music]
24:37days away.
24:38If Sam aligned directors joined the
24:40board, then these four would lose their
24:43majority and they'd never be able to
24:45fire him. Their window was closing.
24:47They decided they had to act. As Ilya
24:50told one board member at the time,
24:52"I don't think Sam is the guy who should
24:54have his finger on the button.
25:00The early hours of Monday, November
25:0220th, 2023. 61 hours since Sam Altman
25:05was fired.
25:06Around 1:00 a.m., a letter starts to
25:08circulate on various signal chats. It's
25:10never been totally clear who drafted
25:12this letter, but it seems likely that it
25:14came from senior executives.
25:16Recent evidence suggests the original
25:18idea came from Sam himself.
25:21Let me read you some of it. It's
25:22addressed to the board of directors.
25:25The process through which you terminated
25:27Sam Altman and removed Greg Brockman
25:29from the board has jeopardized all of
25:31this work and undermined our mission and
25:33company.
25:34Your conduct has made it clear you did
25:36not have the competence to oversee
25:38OpenAI. Despite many requests for
25:40specific facts for your allegations, you
25:42have never provided any written
25:44evidence.
25:45You also informed the leadership team
25:47that allowing the company to be
25:48destroyed, quote, would be consistent
25:50with the mission.
25:52That mission, if you recall from
25:54earlier, being to ensure that AGI
25:56benefits all of humanity.
25:59The letter demands that OpenAI reinstate
26:01both Sam Altman and Greg Brockman.
26:04The first signature
26:06is Mira Murati's.
26:08This is one of those moments that we can
26:10only fully appreciate now that we know
26:11the full details of why the board fired
26:14Sam. Murati is essentially accusing the
26:16board of not having evidence when she
26:19herself provided many of the screenshots
26:20that formed the basis for their case.
26:23Of course, no one else knew that at the
26:25time. The letter spreads around it
26:26quickly, and within a few hours most of
26:28OpenAI's employees have signed. Along
26:31with the letter, many employees start
26:32posting the same phrase on X.
26:35OpenAI is nothing without its people.
26:38Now, there are a few possible readings
26:39for why so many employees signed this
26:41letter so quickly. One is that they were
26:44genuinely shocked and unhappy about
26:46Altman's firing.
26:47Another is that they were worried about
26:48the value of their shares. For a lot of
26:50them, the majority of their net worth.
26:53As one employee who chose to remain
26:54anonymous told a journalist a few weeks
26:56later, Sam Altman is not the best CEO,
26:59but millions and millions of dollars and
27:01equity are at stake.
27:04Another reading backed up by some
27:06reporting at the time is that once the
27:08letter started to snowball, people felt
27:10pressure to jump on the bandwagon,
27:11whether for fear of Altman's retribution
27:13or simple peer pressure.
27:14>> I think there was a lot of like phone
27:16banks and like people like getting
27:18really organized to try to like really
27:19get everyone to sign this thing.
27:20>> In the wee hours of Monday morning,
27:22another name appears on the growing list
27:24of signatures.
27:26Ilya Sutskever.
27:37There's this great line from season 1 of
27:39The Wire.
27:41If you come at the king, you best not
27:43miss.
27:45It's pretty timeless advice, whether
27:46you're talking about drug lords or
27:48ancient emperors. [music]
27:50If you challenge someone extremely
27:51powerful,
27:52you have to succeed immediately and
27:54totally,
27:55or you're screwed. Turns out that OpenAI
27:58was not the first time Sam Altman had
27:59faced this kind [music] of challenge.
28:02First, Loopt, Sam Altman's first
28:04company.
28:05>> Loopt is about connecting with people on
28:06the go, which is after all the main
28:08reason you have a phone.
28:09>> It's a location-sharing app, raises
28:11something like 30 million, doesn't
28:13really [music] ever take off, sells
28:15several years later for 43 million.
28:17Fine.
28:18But it ends badly.
28:20Senior executives twice tried to get Sam
28:22fired for deceptive and chaotic
28:24behavior. Still, he moves on to his next
28:26job, becoming president of Y Combinator
28:28age 28.
28:30He's really good at this. Incredible
28:33fundraiser, brilliant at building
28:34relationships. This is where his star
28:36really starts to rise in Silicon [music]
28:37Valley.
28:38But YC's partners and executives grow
28:42increasingly frustrated at Sam's
28:44distractibility, at the attention that
28:46he's paying to his own personal
28:47projects, including the newly founded
28:49OpenAI, and to his own funds.
28:51Eventually, YC's beloved founder Paul
28:54Graham and his wife Jessica Livingston
28:56had to intervene to get Sam removed.
28:58Graham has this incredible quote about
29:01Sam, written when Sam was just 23
29:02[music] and intended to be
29:04complimentary.
29:06You could parachute him into an island
29:07full of cannibals,
29:08>> [music]
29:08>> come back in 5 years, and he'd be the
29:10king.
29:11Murati, Sutskever,
29:14the board, they know this is who they're
29:15going up against.
29:16His experience, his network, his allies
29:19in the press. The board knows their only
29:21hope is to act quickly and without
29:23warning. This is why they couldn't risk
29:25getting more senior execs involved, why
29:27the firing comes out of the blue, why
29:29they couldn't warn Microsoft in advance.
29:31But this strategy also ultimately proved
29:34their undoing. Within the first day,
29:36Murati had seen enough to know that
29:38things weren't going to go the board's
29:39way.
29:40>> [music]
29:40>> When I realized that their decision was
29:43like potentially catastrophic for the
29:45company, I felt like I had to act very
29:48quickly.
29:49>> In her deposition in the Musk v. Open
29:51trial this May,
29:52Helen Toner said this about Murati.
29:55She was waiting to see which way the
29:56wind would blow, and she didn't realize
29:59that she was the wind.
30:01And as for Ilya Sutskever,
30:04he thought he would be greeted as a
30:05hero, that he was freeing people from
30:07what he saw as Sam Altman's gaslighting.
30:10Instead, people were blaming him for
30:12destroying the company.
30:14On that Sunday night, in the office
30:16lobby as employees were streaming out,
30:19Anna Brockman, wife of OpenAI's
30:21president Greg Brockman, confronts Ilya.
30:24They were old friends. Ilya had
30:26officiated the Brockmans' wedding.
30:28In tears, she pleads with Ilya to change
30:30his mind.
30:33Ultimately,
30:35he blinked.
30:42By Tuesday morning, it's all over. More
30:44than 700 of OpenAI's 770 employees have
30:47signed the letter. With over 90% of the
30:50company threatening to walk, the board
30:52know they've lost.
30:54A deal is reached.
30:55Sutskever, Toner, and McCauley are out.
30:58Bret Taylor, CEO of Salesforce and
31:01someone that McCauley had previously
31:02been worried would be too deferential to
31:04Sam, and also Larry Summers, former
31:07Treasury Secretary, president of
31:08Harvard, agree to join the board.
31:11D'Angelo stays. He's actually still on
31:13the board to this day.
31:15By holding out for as long as they did,
31:17the board do win a few concessions.
31:18[music]
31:19Brockman and Altman won't get their
31:20board seats back, not right away, and
31:23there'll be an investigation by an
31:24outside law firm into what happened.
31:27But, it's not much, really.
31:29At 10:00 p.m. Pacific, the announcement
31:32goes out.
31:33Sam Altman will return as CEO of OpenAI.
31:37There's an impromptu party at the
31:39office. Everyone's celebrating. Greg
31:41Brockman posts a selfie with the
31:43caption, "We are so back."
31:46There's a great detail from the
31:48reporting.
31:50At some point, someone turned on a smoke
31:52machine. It set off the fire alarm.
31:55Everyone kept partying.
32:03About 2 and 1/2 years have passed since
32:05that weekend in November. Many experts
32:07now think that AGI could arrive in as
32:10little as 5 years. Even skeptics these
32:12days rarely give it more than 20.
32:15Earlier this year, Claude 3 Opus,
32:17an AI system capable of finding
32:19vulnerabilities in software that experts
32:21had missed for years or even decades.
32:23Enough of a jump up in capabilities that
32:25its makers decided not to release it to
32:27the general [music] public. Over the
32:29last couple of years, a lot of changes
32:30have happened at OpenAI as well.
32:32Remember that structure, nonprofit on
32:34top, profit caps for investors, the
32:37board in total control, keeping to the
32:39mission, the benefit of humanity. In
32:41December 2024, OpenAI tried to
32:44restructure the company.
32:45Under the new plan, the nonprofit would
32:47essentially lose control.
32:49The plan didn't go through. After months
32:51of pressure, OpenAI backed down.
32:55Sort of.
32:56Under the new new structure, the
32:57nonprofit is on top in name.
33:00There's a safety committee that in
33:01theory could halt model releases, and it
33:03owns about a quarter of the equity in
33:05the company.
33:06But, the profit caps that were supposed
33:08to keep OpenAI focused on its mission
33:10and not on investors,
33:12those are gone.
33:13And in practice, the board's authority
33:15is massively weakened.
33:18Now, it should be said that OpenAI still
33:21gets its models tested by third-party
33:23safety evaluators, and they're doing
33:25meaningfully important safety work on
33:27their current models.
33:29But, what's left is functionally much
33:31closer to a normal tech company.
33:34A normal tech company that was built on
33:36purpose by people who thought they might
33:38be making a civilization-ending
33:39technology not to be a normal tech
33:42company.
33:44Sam Altman built that structure. He
33:46talked publicly about how it was crucial
33:49that the board could fire him, how that
33:50was part of why OpenAI was responsible.
33:53He recruited some of the world's top AI
33:55researchers under [music] the idea that
33:57AGI should not be built by a
33:59mega-corporation.
34:00And then,
34:01when the mechanism was actually
34:03triggered, when the board actually did
34:04the thing it was designed to do, within
34:0672 hours,
34:08Sam had announced plans to go build AGI
34:10at Microsoft [music]
34:11and refused to come back until two of
34:13the board members who had challenged him
34:14were gone.
34:16There are fewer people left at OpenAI
34:17who might challenge Sam these days.
34:20Within a year of his return, both Ilya
34:22Sutskever and Mira Murati had left. A
34:24lot of the senior people at OpenAI who
34:26were most worried about AI safety have
34:28left, too.
34:29Just a few months before he was fired,
34:31Sam signed a one-sentence open letter
34:34that reads,
34:35"Mitigating the risk of extinction from
34:37AI should be a global priority alongside
34:40other societal scale risks such as
34:42pandemics and nuclear war."
34:44Since the board drama, which is referred
34:46to internally as the blip,
34:48Sam talks more about a gentle
34:50singularity.
34:51>> Someday, we will make something that
34:53qualifies as an AGI by whatever fuzzy
34:55definition you want. The world will have
34:56a 2-week freakout, and then people will
34:58go on with their lives.
34:59>> He recently told the New Yorker, "My
35:01vibes don't really match a lot of the
35:03traditional AI safety stuff."
35:06So,
35:072 and 1/2 years on,
35:09AI is more powerful than it was in 2023
35:12by a lot.
35:13OpenAI is more powerful, too. It is now
35:16worth $852 billion.
35:19The people that tried to put the brakes
35:21on one of the most powerful AI companies
35:23in the world are gone.
35:25The structure that was supposed to act
35:27as a brake is functionally gone, too.
35:30So,
35:31where does that leave us?
35:37I've actually met Sam Altman.
35:39In 2022, I was volunteering at a
35:41conference. I was the speaker liaison,
35:43and he was one of the speakers. So, I
35:46showed him to the green room, gave him a
35:48bottle of water, made small talk.
35:51I get it.
35:53I get why so many talented people have
35:55wanted to work with him over the years.
35:57I get why he's so successful.
35:59You talk to the guy, and
36:01he makes you feel like the most
36:02interesting person in the world.
36:05I think Sam Altman is not that
36:07mysterious.
36:08I think he's a guy who's really good at
36:10finding winning moves on whatever game
36:12board he happens to find himself on.
36:15I think he'll continue to do that.
36:17And I think it's
36:18sad and more than a little scary that
36:21the moves he makes might matter so much
36:24to the development of one of the most
36:25powerful and dangerous technologies in
36:27history.
36:28Sam is not the only one playing this
36:31game. We're doing things our way
36:32>> [music]
36:33>> and they're doing things their way.
36:35We'll see who wins in the market and
36:36we'll see who wins in the court of
36:38public opinion.
36:39>> It's unbelievably competitive, that
36:40whole AI space.
36:42>> They all want to win and they all want
36:43to beat each other and they all want to
36:44go fast.
36:45>> Right now, the future of AI basically
36:47depends on the moral compass of like
36:49five people.
36:50That's the situation we're in. Most
36:52powerful technology that humanity might
36:54ever build and how it goes depends on
36:56how a few CEOs are feeling.
36:58That is not a stable situation. I don't
37:01want to live in that world and I'm
37:02pretty sure that you don't, either.
37:04AI is going to change all of our lives.
37:07I think it's going to change our entire
37:08civilization.
37:10And we should all have a seat at the
37:12table where [music] how that goes gets
37:14decided.
37:15Because I think what this story really
37:17tells us is that self-regulation
37:20is not going to work.
37:22The OpenAI nonprofit structure was a
37:24close to best-case scenario of corporate
37:26self-regulation of AI.
37:29You had a nonprofit board of volunteers
37:32explicitly committed to humanity over
37:35profit with the power to fire the CEO.
37:38And it was defeated in 72 hours by
37:41market pressures, employee revolt, and
37:43one very powerful man's network.
37:45>> Although the board had the
37:47legal
37:48ability to fire me, in practice it
37:50didn't quite work. And that is its own
37:53kind of governance failure. It does make
37:56it harder for me to like look you in the
37:57eye and say, "Hey, the board can just
37:58fire me."
37:59>> So,
38:01here's the deal.
38:02We are not going to fix this from our
38:03bedrooms, but none of this is set in
38:05stone. People are realizing that these
38:07AGI companies are building something
38:09extraordinarily powerful and
38:11consequential and risky. I'm talking
38:13about people like Bernie Sanders and
38:14Mitt Romney and the freaking Pope.
38:17When OpenAI tried to restructure into a
38:18normal tech company without nonprofit
38:20governance,
38:21former OpenAI employees, Nobel
38:23laureates, and elected officials stepped
38:25up and made them back down.
38:28When OpenAI and other big tech companies
38:30lobbied to ban US states [music] from
38:32passing laws regulating AI,
38:34non-profits went to battle and got the
38:36Senate to vote 99 to 1 against the US
38:40cannot agree on anything Senate.
38:42>> [music]
38:42>> And then there's the super pack funded
38:44by Greg Brockman, OpenAI's president,
38:46with advice from our friend Chris
38:47punch-them-in-the-mouth Lehane.
38:49As more and more of its bad behavior was
38:52exposed, individual researchers inside
38:54of OpenAI demanded answers.
38:57OpenAI ended up distancing themselves.
38:59You really do have more leverage, more
39:01opportunities to influence how things go
39:04than you think.
39:06And more now than you may have in a few
39:07years.
39:08We've put a link in the description of
39:10this video. You can't miss it. If you're
39:11a US citizen, they'll tell you how to
39:13contact your representative in less than
39:15a minute. If you work at one of these AI
39:17companies, then please
39:20hold your leaders accountable. Ask them
39:21questions. We'd love to help you think
39:23about how to do that.
39:24If you work in policy, you can sign up
39:26for one-on-one advising with us. We
39:27would love to help you figure out how to
39:29pivot into AI governance and regulation.
39:31>> We're in a pivotal moment right now.
39:33I've got a 2-year-old. You know, I want
39:35to be able to tell him when he's grown
39:36up what I did and have him be proud of
39:37me. Just like make the jump. Redirect
39:39all of your focus onto trying to make
39:42your career something that you'll be
39:43proud of.
39:44>> And let's do it. Let's make sure that
39:46Sam Altman is not the only person who
39:48decides our future.
39:53>> [music]
39:58[music]

About this transcript

This page contains the full transcript of You should, unfortunately, be worried about Sam Altman. by AI In Context, generated from the public captions YouTube serves with the video. The transcript has 6,813 words across 1,177 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.