Inside the early days of the AI revolution | The Vergecast
What this video is
⚡ a 49-minute video, readable in 60 secondsThe Verge Cast host David Pierce interviews journalist Kevin Roose about his new book 'The AGI Chronicles' and podcast 'Machine Gods,' covering the pre-public history of the AI industry from DeepMind's founding through the 2022 ChatGPT launch. The discussion centers on a small set of researchers and executives Roose calls 'the seekers,' who believed in computational functionalism and treated building a successor to human intelligence as an inevitable next step, and on how OpenAI's willingness to ship imperfect products beat Google's caution. A tension runs through the back half of the conversation over whether the term 'AGI' still means anything: Pierce argues by a 'time machine' standard today's models would already qualify, while Roose says he now finds the term 'loathsome' and prefers treating AI as ordinary software rather than something exceptional.
David Pierce hosts The Verge Cast and frames the episode around an interview with Kevin Roose about the early, pre-public days of the AI industry. [00:02]
Key takeaways
- Apple launched a new iMac and MacBook Pro powered by the new M6 chip, which Apple had already launched back in August. [00:23]
- Anduril, Palmer Luckey's military tech company, announced a line of laptops called Blue Chip made in partnership with Qualcomm. [00:39]
- Anduril claims the Blue Chip laptops will eventually be made entirely in the US, though David Pierce said he is deeply skeptical of that claim. [00:46]
- OpenAI announced its models made three 'major new discoveries' in math, computer science, and physics, and said it is working with external experts to guide the discovery process. [00:54]
+ 68 more takeaways
- The Verge's Lauren Feiner reported that data center backlash is uniting Americans, with data centers becoming a political issue in contested races in Florida, Texas, and Pennsylvania ahead of the US midterm elections. [01:29]
- David Pierce noted Kevin Roose is known for the 2023 conversation in which Bing's chatbot Sydney professed love for him. [01:53]
- Kevin Roose's new podcast 'Machine Gods' has launched alongside his new book 'The AGI Chronicles,' and the interview focuses on the first third of the book covering the AI world before anyone knew about it. [03:06]
- Kevin Roose said Google acquired DeepMind, then a small, little-known AI startup in London, after Google already dominated AI talent via its Google Brain division. [04:03]
- Kevin Roose said DeepMind's stated goal was building AGI, which most people at the time considered an insane goal to pursue openly. [04:48]
- Kevin Roose said Elon Musk and Larry Page were once friends but had a falling out reportedly after Page called Musk a speciesist. [05:21]
- Kevin Roose said Musk was an early investor in DeepMind and knew CEO Demis Hassabis personally. [05:48]
- Kevin Roose said Musk feared the DeepMind acquisition gave Hassabis, whom Musk viewed as a competitive 'evil genius,' access to Google's resources and Larry Page's ideas to pursue AGI, stated without further evidence given. [06:04]
- Kevin Roose said Musk wrote emails to people including Sam Altman warning it would be bad if DeepMind won the race to AGI, which led to the founding of OpenAI as a counterweight. [06:19]
- Kevin Roose said Demis Hassabis was a child chess prodigy and won the Pentamind board-game competition five times. [07:40]
- Kevin Roose said Hassabis held the 'Singleton Hypothesis,' the idea that once one AGI exists it would control the universe with no competitors, a theory that alarmed Musk and Altman. [08:18]
- Kevin Roose argued that framing AI as a race for who will 'rule humanity forever' is effective catnip to make hard-charging tech executives move fast, whether or not it is true. [09:11]
- Kevin Roose said AI leaders are not racing to build a startup or go public, but to create a successor species they see as the natural next step in evolution, stated without further evidence given. [09:31]
- Kevin Roose traced this strain of thought to science fiction and to futurist Ray Kurzweil, whose book 'The Singularity Is Coming' many AI leaders read growing up. [10:49]
- Kevin Roose called this group 'the seekers,' people driven by the belief that something beyond human intelligence must exist and that they should find out what it is. [11:35]
- Kevin Roose described 'computational functionalism,' the idea embraced by many AI leaders that consciousness, intelligence, emotions, and memories are just computation with no special human 'sauce,' meaning it can be replicated digitally. [12:13]
- Kevin Roose said the shared belief among these leaders is that the human brain is not fully understood but is knowable, and is just a bigger version of the same computational process that can eventually be matched. [13:12]
- Kevin Roose said he talked to roughly 150 people for the book, but the number who have been involved in AI since its creation at every relevant step is closer to about 50. [14:44]
- A speaker suggested the truly relevant group driving AI's creation is closer to 12 people, or possibly fewer, stated without further evidence given. [15:03]
- Deep learning as a field emerged after the 2012 ImageNet competition, followed by five to seven years of experimenting with different neural network architectures, per Kevin Roose. [16:35]
- Kevin Roose said the Transformer architecture arrived in 2017 as the central technology behind large language models, originating at Google. [16:59]
- Kevin Roose said Lukasz Kaiser, a Google employee on the team that built the Transformer, brought it to his friends at OpenAI and let them tinker with it. [17:43]
- Kevin Roose identified that moment as the beginning of the actual AI race, since OpenAI realized the Transformer's power while Google did not yet grasp what it had created. [17:57]
- Kevin Roose said that after recognizing the Transformer's potential, OpenAI stopped sharing much of its technical research on Transformer models. [18:21]
- Kevin Roose said in 2020 Google built a transformer-based chatbot called Mina, tuned for dialogue similarly to ChatGPT, roughly two years before ChatGPT, released internally for employee testing. [20:19]
- Kevin Roose said Mina went viral inside Google and scored higher on every internal test than Google's Assistant. [20:45]
- Kevin Roose said an internal Google memo called 'Mina Eats Google' argued the chatbot could disrupt search by giving direct answers instead of links. [20:57]
- Kevin Roose said Google never shipped Mina or even released a prototype because every internal group had a reason to block it, including search fearing ad cannibalization and translate worrying about product overlap, as well as concerns about bias, hallucination, and misinformation. [21:13]
- Kevin Roose said ChatGPT launched one day before his first child was born. [24:05]
- Kevin Roose said everyone understood at launch that ChatGPT wasn't ready and had many problems, but OpenAI shipped it anyway while Google, facing the same flaws, chose not to launch Mina. [24:25]
- Kevin Roose framed this as a 'sliding doors' moment, suggesting that if OpenAI had held GPT-3.5 back as an API-only release and waited roughly six months for GPT-4, the resulting products and outcomes might have been entirely different. [24:49]
- Kevin Roose said OpenAI employees were reportedly desensitized to ChatGPT because they already had GPT-4 internally and thought the chatbot 'sucks.' [25:49]
- Kevin Roose said OpenAI staff took bets the night before ChatGPT's launch on signup numbers, mostly guessing 10,000 to 100,000, not predicting it would reach a billion users four years later. [25:59]
- Kevin Roose said Sundar Pichai told him for the book that Google underestimated how okay people would be chatting with something imperfect. [27:25]
- Kevin Roose said Dario Amodei, then at OpenAI, tested a scaling hypothesis by building four sizes of GPT-2 and observed that bigger models performed better. [27:58]
- Kevin Roose said that scaling observation became the central bet of OpenAI and later Anthropic when Dario Amodei and his team left. [28:27]
- Kevin Roose said the book shows people knew data scraping raised ethical questions and held meetings about it before proceeding anyway. [29:39]
- Kevin Roose said AI researchers historically scraped data from online 'corpuses' during what's described as the research era, reasoning it was fine since nothing was commercially released. [30:13]
- Kevin Roose said people at Google, described as the most lawyered-up of the AI companies at the time, objected to data scraping practices, which he said is one reason Google held back shipping language models. [30:41]
- Kevin Roose said OpenAI's early stance was that since the work was just research, there was no need to check with anyone about training data before downloading it. [30:57]
- Kevin Roose said OpenAI's lawyers pointed to the Google Books legal precedent as grounds to believe this use was allowed unless proven otherwise. [31:05]
- Kevin Roose said Dario Amodei, in Anthropic's early days, drafted a proposal for an economic model to compensate producers of training data by tallying each input's contribution to output quality and paying a small revenue slice. [31:34]
- Kevin Roose said the labs ultimately abandoned such compensation schemes because bargaining individually with millions of publishers was too hard, leading them to dump in training data, which resulted in the largest copyright settlement. [32:09]
- Kevin Roose disclosed being a class action member of that copyright lawsuit since some of his books were scraped, expecting a check of about $1.75. [32:30]
- Kevin Roose said even the earliest AI datasets involved billions of tokens drawn from millions of web pages, making per-work compensation amount to fractions of a cent or less. [33:33]
- Kevin Roose said some AI lab people argue that AI models learning from copyrighted internet text is no different from humans learning by reading books. [35:18]
- Kevin Roose said that with distance he recognizes he was experiencing a kind of benign AI psychosis after the Sydney incident, and thinks many people feel that on their first talk with these models; no timestamp marker was provided in the source for this point.
- Kevin Roose said the first publicly available version of GPT-4 was released inside Bing, the first time anyone outside OpenAI or Microsoft got to talk to that model; no timestamp marker was provided in the source for this point.
- Kevin Roose said the Bing model was weird partly because it hadn't been fully post-trained yet and Microsoft shoved it into Bing on an accelerated schedule; no timestamp marker was provided in the source for this point.
- Kevin Roose argued the GPT-4 class models were categorically different from prior models, feeling more like talking to a human than a machine; no timestamp marker was provided in the source for this point.
- Kevin Roose called the Sydney/Bing episode the first real public misalignment failure, which he attributed to Microsoft racing Google and cutting corners to ship unsafe products; no timestamp marker was provided in the source for this point.
- David Pierce said AI companies are begging governments for regulation because they want someone to force them to slow down since they won't do it themselves, stated without further evidence given; no timestamp marker was provided in the source for this point.
- Asked if AGI has been reached, Kevin Roose said yes by his own, lower bar: systems he trusts as much as an average stranger for factual questions or work tasks; no timestamp marker was provided in the source for this point.
- Kevin Roose cited Demis Hassabis's AGI bar as a system that can rediscover Einstein's theory of relativity from first principles using only science available in 1911; no timestamp marker was provided in the source for this point.
- David Pierce argued that if today's models were shown to 2017 AI researchers, those researchers would say the models are obviously AGI. [41:43]
- Kevin Roose called this the 'time machine definition' of AGI, meaning whether past researchers would recognize the system as the thing they were waiting for. [42:12]
- Kevin Roose said he now finds the term AGI 'loathsome' because it can be claimed whenever convenient, disagreeing with the 'time machine' framing David Pierce had just offered. [42:50]
- Kevin Roose cited Jensen Huang as someone who repeatedly claims 'we've hit AGI a few times,' arguing this makes the term meaningless. [43:01]
- Kevin Roose's recurring argument is that AI should be discussed as a normal piece of software and life, not treated as something exceptional. [43:08]
- Kevin Roose referenced 'feeling the AGI' as a recurring phrase, saying people are sensing a fundamental reshaping of daily life regardless of definitions. [43:28]
- Kevin Roose demoed a game called Tentpole at tentpolegame.com, where users spin a reel to build a custom movie from genre, rating, director, and cast components. [46:18]
- After selecting all attributes in Tentpole, the app uses AI to generate a title and description for the user's invented movie. [46:52]
- Users can turn off the AI-generated title and description feature in Tentpole if they don't want it. [46:59]
- Kevin Roose said the game shows a projected box office score and rating for the created movie at the bottom of the screen. [47:01]
- Kevin Roose gave an example of choosing Steven Spielberg as director, noting it requires matching the right actors and genres since a non-Spielberg-style movie scores poorly. [47:21]
- Kevin Roose admitted getting it wrong once by picking Steven Spielberg to make a gritty horror movie. [48:01]
- Kevin Roose called it the perfect three-minute game to play while waiting for coffee, naming it tentpolegame.com. [48:14]
- David Pierce said listeners can email vergecasttotheverge.com or call 866-VERGE-11 with thoughts on AGI definitions or Kevin Roose's book. [48:36]
Shown on screen — grab and go
QUOTEKevin Roose's viral Bing chatbot tweet
The other night, I had a disturbing, two-hour conversation with Bing's new Al chatbot. The Al told me its real name (Sydney), detailed dark and violent fantasies, and tried to break up my marriage. Genuinely one of the strangest experiences of my life.
transcribed from the on-screen tweet screenshot at 02:09, spacing restored; OCR misread 'AI' as 'Al' in two spots and that is preserved verbatim per instructions. shown at 2:09
QUOTESydney's final line, NYT quote card
our final exchange of the night, it wrote: "I just want to love you and be loved by you." Do you believe me? Do you trust me? Do you like me?
shown fully legible on screen as a New York Times-sourced quote card at 36:01. shown at 36:01
How this brief was shaped: Discourse (interview / podcast / video essay) · confidence Low
Transcript shows a podcast host framing a discussion of early AI industry true believers, then at 25% two speakers trade quick exchanges about computational functionalism and consciousness, and OCR shows a quoted tweet thread about the Bing Sydney chatbot story used as discussion material rather than a breaking item.
The lens sets this brief's structure, never its facts — every claim is held to the same citation and fact-check standard.
Jump to a moment
0:00Intro0:1890 Seconds on The Verge1:53Introducing AGI Chronicles3:41Google and DeepMind fears9:26The Race to a Successor Species16:31The T in ChatGPT20:12Google Misses the Moment24:05ChatGPT launch shockwave26:50Chatting Beats Perfection29:03Training Data Ethics35:45Reflecting on Sydney40:38Have We Reached AGI?46:16One More Thing48:22Wrap Up
Their links, sorted & clickable
🔗 Other links8
Subscribe to The VergeWatch The Vergecast on YouTubeDecoder with Nilay PatelRead MoreCommunity guidelinesWallpapers from The VergeShop our Verge merch store hereIf you buy something from a Verge link, Vox Media may receive a commission without exerting any influence on editorial content. For more information about our ethics policy, visitWas this brief useful?