Oct 14, 2026 · 1h 4m
Listen from 40:10
Listen at 40:10
AI’s leaders race ahead while safety struggles to keep pace
AI Leaders Can't Get Along. So How Will They Keep Us Safe? — With Kevin Roose
The people directing frontier AI disagree about its risks and goals even as increasingly capable systems may reshape work, power, and public safety.
- 1AI capabilities and research automation may be advancing faster than safety methods and governance can respond.
- 2Rivalries among OpenAI, Anthropic, Google DeepMind, and Meta complicate efforts to manage increasingly powerful systems.
- 3Broader public awareness of AI risks offers a cautious reason for optimism, despite unresolved dangers around recursive improvement and autonomy.
Don't miss
Roose explains how reasoning traces and test-time computation transformed AI systems from limited language tools into technologies capable of tackling extremely difficult problems.
The brief
Kevin Roose argues that AI’s capabilities are advancing faster than safety research, while rivalries among leading labs make coordinated restraint difficult.
The conversation profiles Sam Altman, Dario Amodei, Demis Hassabis, Sergey Brin, and Satya Nadella as competing personalities shape the direction of frontier AI.
A key turning point came when reasoning traces and test-time computation showed that models could tackle difficult problems far beyond their earlier apparent limits.
The discussion weighs recursive self-improvement, automated AI research, and the possibility that a major incident could abruptly change public and regulatory attitudes.
Roose ends on a cautious note: wider public awareness of AI risk may be reassuring because safety is no longer confined to a small specialist circle.
What was said on this episode
34 statements · 16 positive · 13 negative · 3 mixed · 2 neutral
AI leaders’ motives cannot be assumed to be pure
“i don't think we can trust that their motives are pure”
Listen at 4:59
A small group cannot be trusted to govern advanced AI
“I don't trust any group of people that small to create and govern technology like this”
Listen at 7:25
AI leaders should cooperate on safety solutions
“I want them to be able to like get together in a room and figure out some solutions on things like safety”
Listen at 7:36
AI development should be slowed
“I think we should slow down”
Listen at 8:39
AI capabilities are advancing faster than safety research and control methods
“the capabilities are outpacing the safety research. and the control methods.”
Listen at 8:57
AI will provide major benefits in health and medicine
“I think there will be incredible benefits to this technology, especially in, like, health and medicine.”
Listen at 9:20
Reporting on AI labs left him less reassured about safe development
“less reassured”
Listen at 10:55
The AI race resembles a Red Queen competition
“that's a very good description of where the ai race is now”
Listen at 12:18
A catastrophic incident could rapidly change AI policy
“things can happen quickly when there's some catastrophic incident”
Listen at 14:32
Google remains third but could still become competitive
“they're still probably in third place right now, but like I wouldn't count them out.”
Listen at 18:56
The first successful AGI builder could become the world’s most powerful entity
“whoever gets it will be the most powerful person in the world, the most powerful entity in the world.”
Listen at 20:41
There is no clear technical barrier to building superintelligence
“I don't see any real technical reason that they can't.”
Listen at 21:36
AI skeptics underestimate the technology’s rate of progress
“the people who are looking at this technology and saying, like, what's the big deal are just not understanding the rate of progress.”
Listen at 23:42
Skepticism about AI companies’ financial models and economic impacts is warranted
“I buy the skepticism about the financial models of these companies, the economic impacts.”
Listen at 26:01
Many AI companies may be wrong about AI’s effects on jobs
“a lot of these companies are... are wrong or at least could be wrong about the impact of AI on jobs.”
Listen at 26:09
Sam Altman’s interpersonal adaptability is both an asset and liability
“Sam's superpower, is also his weakness”
Listen at 29:04
Dario Amodei is exceptionally honest
“Dario is a guy who is honest to a fault.”
Listen at 32:16
AI leaders are highly intelligent but significantly flawed
“I would put all of these guys in the category of extremely smart and obviously flawed in many ways.”
Listen at 35:29
Reasoning models evaluate and revise their answers iteratively
“the reasoning developments were of course like, the bots will evaluate their answers and then decide whether it's on the right path or not, and then keep going.”
Listen at 37:42
Additional inference compute can improve models without retraining
“we now have a way of making models smarter without retraining them.”
Listen at 39:06
AI researchers repeatedly recognized that breakthroughs would scale transformatively
“they have understood at basically every point in this journey. that what they have just seen is going to scale and is going to have these transformative effects”
Listen at 39:32
Microsoft will become more of a commodity cloud than a frontier AI lab
“I think they're going to be something more like a sort of commodity neocloud or something like that than they are going to be like a frontier AI lab.”
Listen at 49:10
AI models should have more distinct personalities
“I think that models should have more divergent personalities.”
Listen at 50:53
Muse lost $41 trading a Kalshi account
“It lost $41.”
Listen at 51:23
Tarbell has no editorial control over Machine Gods
“they have no editorial control over our show.”
Listen at 53:37
OpenAI’s belief in AGI helps explain its continued lead
“this explains a lot of why OpenAI is still in the lead.”
Listen at 56:25
OpenAI’s early compute investment contributes to its leading position
“they made the bet before it was obvious. So I think that, I think belief, conviction, whatever you want to call it, is like the central explanation for why they're still in the lead or close to it.”
Listen at 57:17
Recursive self-improvement by AI is possible
“I believe in recursive self-improvement. I believe that it is possible.”
Listen at 58:45
AI labs are racing toward dangerous recursive self-improvement and should stop
“I believe that the labs are racing to do it. And I think it's dangerous. I think they should stop.”
Listen at 58:48
AI could eventually lead most or all of Anthropic’s research projects
“I don't see any reason why that number couldn't go up to 50 or 75% or maybe someday 100%.”
Listen at 59:26
AI labs want to automate their researchers’ jobs
“they want to automate their researchers out of jobs.”
Listen at 1:00:31
Society could solve AI safety and control problems
“I think we actually could solve this.”
Listen at 1:02:18
Slowing AI development could create time for better control methods
“We could slow down. We could buy ourselves more time to develop better ways of controlling these systems.”
Listen at 1:02:21
AI should be developed to improve health and longevity
“I want AI to be built. AI to cure cancer and disease and to help us all live longer and healthier lives.”
Listen at 1:02:27
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.
