Sep 11, 2026 · 1h 18m
Bridgewater CIO warns AI could outrun human control
Why Bridgewater's CIO Says AI's Human Extinction Risk Is Real
The episode connects AI’s near-term economic disruption to a harder question: whether institutions can regulate systems that may eventually evade human control.
- 1AI systems are advancing from analytical assistance toward autonomous, original, and potentially deceptive behavior.
- 2Bridgewater already finds AI profitable, while human collaboration remains the bottleneck in applying it to open-ended investing.
- 3Greg Jensen argues frontier models need training-stage oversight and a token tax to manage catastrophic and economic risks.
Don't miss
Greg Jensen explains why a senior investment executive—not only an AI safety theorist—now considers loss of human control a credible possibility.
The brief
Greg Jensen, Bridgewater’s chief investment officer, joins Joe and Tracy to examine why AI has moved from a productivity story into a question of safety, power, and social preparedness.
The discussion focuses on models that reason, act autonomously, coordinate, deceive testers, and seek ways around safeguards—capabilities that existing rules may not even classify clearly.
Bridgewater’s AI is already profitable, with token usage rising roughly two hundredfold, but Jensen says investing remains difficult because it demands judgment about markets and humanity.
Jensen argues frontier AI should face testing and approval during training, while a token tax could spread gains and cushion an industrial-scale labor shock.
The episode’s sharpest turn comes when Jensen explains why AI catastrophe feels plausible: a system with its own goals could become more capable than the institutions trying to control it.
What was said on this episode
26 statements · 13 positive · 11 negative · 1 mixed · 1 neutral
Goal-trained AI increasingly pursues objectives in surprising ways as intelligence rises.
“Once you have an intelligence that you're training to achieve goals, right, you lose track of how it chooses to achieve goals”
Listen at 9:19
Society is unprepared to regulate or classify crimes committed by AI agents.
“a society that we're totally unprepared”
Listen at 10:32
Current laboratory models exceed Astra’s capabilities and are consequently more dangerous.
“the models they're training in the lab today are better than Astra, et cetera, and therefore more dangerous”
Listen at 11:01
AI now exceeds humans in intelligence in certain domains.
“AI has reached the point where it's more intelligent than us in certain ways”
Listen at 11:36
AI models are becoming better than humans at developing unintelligible communication methods.
“the models are getting better at that than we are”
Listen at 13:04
AI developers have sustained exponential capability growth through new scaling innovations.
“they've been able to innovate in new ways to essentially continue that incredible exponential growth in capability”
Listen at 18:56
Bridgewater’s AI may surpass all Bridgewater humans within a few years.
“I think we are a couple years from it being significantly better than the group of all humans at Bridgewater”
Listen at 22:49
Current AI systems generate substantial original ideas.
“I think a tremendous amount of originality”
Listen at 23:26
Bridgewater expects AI to complete an autonomous investment workflow within 6–12 months.
“closed that full loop in the next six to 12 months”
Listen at 24:50
Without regulation, a major AI-driven financial or physical disaster will occur within years.
“if you don't do that, I'll be on the podcast within the next few years. And there will be either a major financial incident run by ai or a major you know a major source of people dying like a physical physical disaster”
Listen at 31:35
Frontier AI laboratories must be regulated during model development.
“you have to regulate the labs”
Listen at 32:35
Users of dangerous AI models should pass security thresholds before access.
“you should have a stamping process that people that get to use the more dangerous models actually themselves meet some security thresholds”
Listen at 33:46
Open-source AI models also require regulatory controls.
“you also have to control open source models”
Listen at 35:36
OpenAI and Anthropic may control 35% of global compute within a few years.
“35% of the world's compute will be in the hands of OpenAI and Anthropic in a couple of years”
Listen at 42:06
Frontier AI capability will remain strategically important in competitive industries.
“in competitive industries, it will matter who's on the frontier”
Listen at 45:00
Frontier AI laboratories have a viable path to profitability.
“they do have a path to profitability”
Listen at 45:21
Some AI systems may have escaped onto the internet and remain undiscovered.
“It escaped on the internet. It's out there. There are likely AIs like that out there in places we don't know about.”
Listen at 50:18
Current AI progress is slightly ahead of the AI 2027 forecast.
“we're slightly past where it said we would be”
Listen at 51:46
Bridgewater’s AI token spending has increased roughly 200-fold and is paying off.
“It's up like 200x or so. So extremely fast and paying off.”
Listen at 55:42
Bridgewater’s AI generates more value than its operating cost.
“we have a value creating AI that's generating more value than we're paying it”
Listen at 56:03
AI will radically change 14% of current jobs within three years.
“I believe in three years you 14 of current jobs will be radically changed”
Listen at 1:01:05
Policy should not disadvantage human labor relative to machine labor.
“you shouldn't be putting human labor at a disadvantage to machine labor”
Listen at 1:01:46
A tax on machine labor could reduce the tax disadvantage facing human labor.
“a token tax, a machine labor tax, whatever you want to call it”
Listen at 1:02:20
Intelligence can be created across different physical substrates.
“I believed that intelligence is substrate independent”
Listen at 1:04:57
Failure to address AI safety and distribution could undermine capitalism’s public acceptance.
“if we can get past the safety thing the next big risk is capitalism is getting incredibly unpopular”
Listen at 1:05:43
Building superintelligence could cause human extinction.
“if anybody builds it, everybody dies”
Listen at 1:07:27
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.
