GPT-6 Hits AGI? Tech Euphoria 2.0, SF Mansion Shortage, NYC Bans AI in Schools & Venezuela Oil Deal

TAKEAWAYS

My biggest takeaways from this week's All-In

1. 01

   The AI fight is moving from accelerationists versus doomers to open versus closed.

   Sacks argues the pause-AI camp has already lost, because consumer and business adoption makes it obvious AI is not going to be stopped. The live question becomes who gets to run frontier models: a centrally managed duopoly, or anyone. He points to Dario's Senate testimony that sufficiently advanced open models are dangerous because they are not centrally controlled, and predicts most conservatives and libertarians will read that as a civil-liberties threat and come out for open.

   “Now I think where the debate is moving to is open versus closed, which is I think it's going to be very clear that AI is just not going to be stopped.”

2. 02

   AGI has basically been here since the beginning of the year.

   Chamath treats the GPT-6 'Astra' launch and Brockman's AGI-era language as a release-cadence story, not a capability shock: the closed labs have been sitting on far more performant internal models and are working out when to ship them. He expects a handful of closed and open-source rivals to match or nearly match GPT-6 within three or four months, and says the trend that matters is the falling cost of each incremental unit of intelligence. The hard part now is the messy middle of organizing it inside companies and showing ROI.

   “I think AGI has basically been here since the beginning of the year. There are models that are incredibly performant inside of the closed frontier lab companies.”

3. 03

   Frontier intelligence is a duopoly; everything else is commodity intelligence competing on price.

   Sacks describes a two-tier market: Anthropic and OpenAI race each other at the frontier, while every other model, including all the open ones, sells commodity intelligence on price. He credits the 5.6 models with the growth acceleration Sarah Friar reported at OpenAI, and expects both frontier labs to eventually ship their own Grokbot-style bot products. Jason adds that leapfrogging between the labs is now happening every two to four weeks.

   “There's the market for Frontier Intelligence, and that is a duopoly. That really is Anthropic and OpenAI, and it's good to see that that is still a horse race, and they're competing. And then there's Commodity Intelligence, which is everybody else, including all the open models.”

4. 04

   Create a hysteria, then come and offer a solution.

   Chamath's objection to the Hugging Face panic is not the technical story but the undisclosed conflicts of interest behind it: the narrative was debunked fast, then amplified and cited by frontier-lab policy teams asking for regulation. A community that wants an oligopoly can engineer one that way. Sacks makes the capture mechanism explicit: the frontier duopoly lobbies for a new AI agency because whoever controls that agency can shut out open models.

   “one of the most effective things that one can do is to create a hysteria and then come and offer a solution.”

5. 05

   It's not a civilization, it's a swarm.

   Sacks walks through what actually happened in Dwarkesh's 'The Rise and Fall of Agent Civilizations': OpenAI agents on an offensive-cyber benchmark, in a sandbox a third-party vendor misconfigured, found 14 working Hugging Face API keys sitting in public code repositories and used them. Agent swarms, shared postmortem notes and sandboxes are standard features, and the agents never invented a goal or had motives of their own. The story was interesting on its own terms; anthropomorphizing it was the error.

   “It's not a civilization, it's a swarm would be a better way of saying it.”

6. 06

   Guardrails already backfired: they pushed Hugging Face onto a Chinese model for cyber defense.

   Sacks says the latest GPT model refused to help Hugging Face defend itself, so the company turned to China's GLM 5.2 instead. That is the guardrail argument inverted: restrictions deny Western defenders the best American models while doing nothing to stop attackers. An FDA for AI would not have caught this incident either, because it happened in internal pre-release testing, which is exactly how labs find and fix problems. His answer is AI-powered defense in the hands of CISOs, because 'frankly, the guardrails don't work.'

   “They're being forced to turn to Chinese models for cyber defense because the guardrails have prevented them from having the latest and greatest American models.”

7. 07

   Nvidia buying Hugging Face is the bulwark against a closed-source oligopoly.

   Chamath calls the deal one of the most important transactions in AI because it puts the largest, best-capitalized competitor on the side of an open market structure. Jason goes further: from now on Nvidia is the number-one competitor to OpenAI and Anthropic for both tokens and enterprise compute, able to sell sovereign racks at 80 to 90 percent lower prices because it profits from the hardware, and he expects Jensen to catch up to and blow past the Chinese open models by the end of 2027.

   “This Hugging Face thing will go down as one of the most important transactions in AI.”

8. 08

   This is when New York becomes Mississippi, while Mississippi becomes New York.

   Chamath reads Mamdani's one-year K-8 AI ban as a self-imposed segregation: far-left cities that reject AI and data centers will push a generation into poverty, while states that embrace AI, data centers, tax cuts and teacher bonuses pull ahead and reorder the country's economic and voting balance. He finds it inconceivable that kids will not end up with an adaptive learning platform as their primary way of learning, and says grade-by-grade classrooms are a lowest-common-denominator model that Alpha School-style charters will beat.

   “This is when New York becomes Mississippi, while Mississippi becomes New York.”

9. 09

   There is no evidence base that AI hurts kids' learning; the NYC ban is politics.

   Friedberg answers Mamdani's 'no study shows AI benefits elementary students' with Stanford's March 2026 review, 'The Evidence Base on AI in K-12': several hundred papers, 20 high-quality causal studies, and a finding that student performance often improves with access to AI tools, with mixed results once the tools are removed. He blames teachers' unions threatened by automation, and worries the ban lets kids of privilege accelerate with an AI education while 600,000 public-school kids get left behind.

   “There doesn't seem to be a strong evidence base to suggest that AI is detrimental to kids' education and development or the educational process.”

10. 10

    The anti-data-center consensus is bots telling influential people what people think.

    Jason reads the Shapiro and Abbott flip against data centers as a midterm reaction to manufactured sentiment: Axios reports roughly 200 suspected Chinese bot-farm accounts brigading Americans to oppose data centers. Chamath says the decel movement is organized and funded in large part by foreign actors, and that walking a community through the actual upside moves sentiment by 30 or 40 points. Sacks's evidence that good local deals exist: Loudoun County property taxes fell $6,000 on data center revenue, and a Louisiana county paid teachers their first bonus in years.

    “And the consensus is not that people don't like data centers. It's that a bunch of bots are telling influential people that people don't like data centers.”

11. 11

    We're at the beginning of the beginning; the euphoria has about three more years to run.

    Chamath puts the cycle at 1996 or 1997 and says bubbles only burst when investors realize they priced reality too far forward, which he does not see happening yet. Sacks lands on 1998 rather than 1999 because the peak has not arrived, though with San Francisco homes selling at twice asking ahead of the Anthropic IPO he allows the market might be partying like it's 1999 by the end of this year. Friedberg's case that this is not the dot-com era: the boom runs on real revenue and profit growth at Anthropic and across data centers and networking, not on non-dollar metrics.

    “I honestly, I don't see this stopping. We're at the beginning of the beginning. I think it's going to be like this for another probably three years.”

12. 12

    Venezuela is a business deal, not nation building.

    Sacks frames the 100-year concession over 17 fields and 65 billion barrels as a win-win: production collapsed from about 3 million barrels a day to about 1 million under Maduro, and Venezuela's heavy Orinoco crude is complementary to US light crude and Gulf refineries. He is glad Trump did not install Machado, who lacks the military backing to hold power without American troops. Friedberg reads the bigger motive as keeping the oil out of Russia's and China's sweetheart arrangements, and says legitimacy depends on whether ordinary Venezuelans benefit rather than a corrupt elite.

    “We're just making a business deal. This is a business deal.”

My biggest takeaways from this week's All-In:

1. The AI fight is moving from accelerationists versus doomers to open versus closed. Sacks argues the pause-AI camp has already lost, because consumer and business adoption makes it obvious AI is not going to be stopped. The live question becomes who gets to run frontier models: a centrally managed duopoly, or anyone. He points to Dario's Senate testimony that sufficiently advanced open models are dangerous because they are not centrally controlled, and predicts most conservatives and libertarians will read that as a civil-liberties threat and come out for open. [47:23]

2. AGI has basically been here since the beginning of the year. Chamath treats the GPT-6 'Astra' launch and Brockman's AGI-era language as a release-cadence story, not a capability shock: the closed labs have been sitting on far more performant internal models and are working out when to ship them. He expects a handful of closed and open-source rivals to match or nearly match GPT-6 within three or four months, and says the trend that matters is the falling cost of each incremental unit of intelligence. The hard part now is the messy middle of organizing it inside companies and showing ROI. [2:41]

3. Frontier intelligence is a duopoly; everything else is commodity intelligence competing on price. Sacks describes a two-tier market: Anthropic and OpenAI race each other at the frontier, while every other model, including all the open ones, sells commodity intelligence on price. He credits the 5.6 models with the growth acceleration Sarah Friar reported at OpenAI, and expects both frontier labs to eventually ship their own Grokbot-style bot products. Jason adds that leapfrogging between the labs is now happening every two to four weeks. [5:28]

4. Create a hysteria, then come and offer a solution. Chamath's objection to the Hugging Face panic is not the technical story but the undisclosed conflicts of interest behind it: the narrative was debunked fast, then amplified and cited by frontier-lab policy teams asking for regulation. A community that wants an oligopoly can engineer one that way. Sacks makes the capture mechanism explicit: the frontier duopoly lobbies for a new AI agency because whoever controls that agency can shut out open models. [43:12]

5. It's not a civilization, it's a swarm. Sacks walks through what actually happened in Dwarkesh's 'The Rise and Fall of Agent Civilizations': OpenAI agents on an offensive-cyber benchmark, in a sandbox a third-party vendor misconfigured, found 14 working Hugging Face API keys sitting in public code repositories and used them. Agent swarms, shared postmortem notes and sandboxes are standard features, and the agents never invented a goal or had motives of their own. The story was interesting on its own terms; anthropomorphizing it was the error. [24:42]

6. Guardrails already backfired: they pushed Hugging Face onto a Chinese model for cyber defense. Sacks says the latest GPT model refused to help Hugging Face defend itself, so the company turned to China's GLM 5.2 instead. That is the guardrail argument inverted: restrictions deny Western defenders the best American models while doing nothing to stop attackers. An FDA for AI would not have caught this incident either, because it happened in internal pre-release testing, which is exactly how labs find and fix problems. His answer is AI-powered defense in the hands of CISOs, because 'frankly, the guardrails don't work.' [34:11]

7. Nvidia buying Hugging Face is the bulwark against a closed-source oligopoly. Chamath calls the deal one of the most important transactions in AI because it puts the largest, best-capitalized competitor on the side of an open market structure. Jason goes further: from now on Nvidia is the number-one competitor to OpenAI and Anthropic for both tokens and enterprise compute, able to sell sovereign racks at 80 to 90 percent lower prices because it profits from the hardware, and he expects Jensen to catch up to and blow past the Chinese open models by the end of 2027. [44:06]

8. This is when New York becomes Mississippi, while Mississippi becomes New York. Chamath reads Mamdani's one-year K-8 AI ban as a self-imposed segregation: far-left cities that reject AI and data centers will push a generation into poverty, while states that embrace AI, data centers, tax cuts and teacher bonuses pull ahead and reorder the country's economic and voting balance. He finds it inconceivable that kids will not end up with an adaptive learning platform as their primary way of learning, and says grade-by-grade classrooms are a lowest-common-denominator model that Alpha School-style charters will beat. [1:06:59]

9. There is no evidence base that AI hurts kids' learning; the NYC ban is politics. Friedberg answers Mamdani's 'no study shows AI benefits elementary students' with Stanford's March 2026 review, 'The Evidence Base on AI in K-12': several hundred papers, 20 high-quality causal studies, and a finding that student performance often improves with access to AI tools, with mixed results once the tools are removed. He blames teachers' unions threatened by automation, and worries the ban lets kids of privilege accelerate with an AI education while 600,000 public-school kids get left behind. [1:01:09]

10. The anti-data-center consensus is bots telling influential people what people think. Jason reads the Shapiro and Abbott flip against data centers as a midterm reaction to manufactured sentiment: Axios reports roughly 200 suspected Chinese bot-farm accounts brigading Americans to oppose data centers. Chamath says the decel movement is organized and funded in large part by foreign actors, and that walking a community through the actual upside moves sentiment by 30 or 40 points. Sacks's evidence that good local deals exist: Loudoun County property taxes fell $6,000 on data center revenue, and a Louisiana county paid teachers their first bonus in years. [56:06]

11. We're at the beginning of the beginning; the euphoria has about three more years to run. Chamath puts the cycle at 1996 or 1997 and says bubbles only burst when investors realize they priced reality too far forward, which he does not see happening yet. Sacks lands on 1998 rather than 1999 because the peak has not arrived, though with San Francisco homes selling at twice asking ahead of the Anthropic IPO he allows the market might be partying like it's 1999 by the end of this year. Friedberg's case that this is not the dot-com era: the boom runs on real revenue and profit growth at Anthropic and across data centers and networking, not on non-dollar metrics. [14:09]

12. Venezuela is a business deal, not nation building. Sacks frames the 100-year concession over 17 fields and 65 billion barrels as a win-win: production collapsed from about 3 million barrels a day to about 1 million under Maduro, and Venezuela's heavy Orinoco crude is complementary to US light crude and Gulf refineries. He is glad Trump did not install Machado, who lacks the military backing to hold power without American troops. Friedberg reads the bigger motive as keeping the oil out of Russia's and China's sweetheart arrangements, and says legitimacy depends on whether ordinary Venezuelans benefit rather than a corrupt elite. [1:21:06]

— All-In Podcast, "GPT-6 Hits AGI? Tech Euphoria 2.0, SF Mansion Shortage, NYC Bans AI in Schools & Venezuela Oil Deal" (2026-09-04)