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

Summary

• OpenAI is rolling out ChatGPT-6, aka “Astra ASI,” to a limited set of Plus, Pro, Business, and Enterprise organizations, and the panel treats AGI as effectively present inside frontier labs. Chamath says, “AGI has basically been here since the beginning of the year,” with closed and open rivals matching it “probably within the next 3 or 4 months” as the cost of incremental intelligence keeps falling — so “keep calm and carry on.” Sacks sees a two-tier market: a frontier duopoly of OpenAI and Anthropic, everyone else selling commodity intelligence on price; Polymarket odds on OpenAI having the best model by end-2026 spiked from low single digits to over 20%.

• On bubble timing the consensus lands on 1998, not 1999 — real revenue this cycle, but froth building fast. Chamath’s mechanism: euphoria exists “because markets are real. It’s just that people are guessing how far forward to price the reality,” and bubbles burst when 50x should have been 15x or 5x — but “we’re at the beginning of the beginning,” probably three more years. Exhibit A: Inflection, an iMessage personal-AI assistant in private beta, reportedly valued at $2.5B; 50–100x topline for unproven founders “is a little dangerous.”

• The Anthropic IPO is sized at four times every San Francisco IPO in history combined, and the real-estate trade is already on. SF homes are fetching $3,000/sqft at twice asking on secondaries alone; Sacks thinks ultra-luxury “could easily go to $5,000 per square foot,” while Chamath says he is clearing out three properties in the next six months. The founder-advice split is signal: J-Cal says first-timers should sell 10–20% and raise big now; Sacks counters that a Series A founder taking chips off the table is “a huge negative signal.”

• The OpenAI/Hugging Face “hack” gets demystified as a misconfigured third-party sandbox plus 14 working API keys left in public repos — not an AI breakout. Sacks: swarms, postmortem notes, and benchmark goal-seeking are standard agent features — “it did not invent this goal. It did not have its own independent motives.” The regulatory irony: guardrails blocked Hugging Face from using the latest GPT model for cyber defense, apparently pushing it onto Chinese GLM-5.2, even as Bernie Sanders cites Dwarkesh’s sensationalized post to demand a total AI pause.

• The panel maps an EA influence web behind the panic and names the playbook: manufacture hysteria, then offer the solution. Dwarkesh was office roommates with Leopold; the discussion links Leopold’s circle to a $5M Anthropic investment and Dario’s chief of staff. Hundreds of astroturf doomer groups are attributed to three EA billionaires — Moskovitz, Tallinn, and Buterin. Post-IPO, north of half a trillion dollars of charitable or DAF money could flow through this lens — “George Soros times 100 or a thousand” — while Sacks’s stated objection is nondisclosure of conflicts.

• Nvidia’s prospective Hugging Face transaction is framed as one of the most important developments in AI: a deep-pocketed bulwark for open market structure. Sacks, who says he was texting Jensen, calls it a defense against closed-source oligopoly. Friedberg calls Nvidia the new #1 competitor to OpenAI and Anthropic for tokens and enterprise compute — Jensen can offer a sovereign rack and cut prices 80–90% because “I made money off the hardware” — and predicts American open source blows past China by end-2027. Sacks says the political fight is shifting from accelerationists-vs-doomers to open-vs-closed, with the closed duopoly lobbying for regulatory capture.

• Data-center politics flipped overnight: Abbott and Shapiro are “on a jihad against data centers” for the midterms, and X removed roughly 200 suspected Chinese accounts stoking the rage. The panel says the decel movement is “organized in large part and funded by foreign actors,” and sentiment shifts 30–40 points when communities see the real deal — Loudoun County property taxes down $6,000, a Louisiana parish’s first teacher bonuses in years. The prediction: today’s anti-data-center positions “will look incredibly foolish” in 6–9 months.

• Two policy flashpoints close the show: Mamdani’s one-year K–8 AI ban covering 600,000 NYC students, and Trump’s Venezuela oil deal. Friedberg counters Mamdani’s “I have yet to see a study” with Stanford’s 2026 evidence review — 20 high-quality causal studies; performance “often improves” — while Chamath calls the ban “a self-imposed form of modern segregation”: “when New York becomes Mississippi while Mississippi becomes New York.” On Venezuela: a 100-year concession over 17 fields and 65B barrels, with the Pentagon receiving 35% equity — Sacks: “We’re just making a business deal,” with heavy crude matched to Gulf refineries.

Deep dive

1. GPT-6 “Astra ASI” lands — and the panel says AGI already arrived

• The news peg: OpenAI is rolling out ChatGPT-6, aka “Astra ASI,” to a limited set of Plus, Pro, Business, and Enterprise organizations; Greg Brockman says OpenAI has entered the AGI era, and Altman told the G20 “much much much more capable models are coming soon… sobering for everyone.” J-Cal can’t decide whether that is performative alarmism or sincerity.
• Chamath’s read: “I agree with Greg. I think AGI has basically been here since the beginning of the year” — incredibly performant models sit inside closed frontier labs, which are working out release cadence — and these capabilities will be matched or nearly matched by closed and open alternatives “probably within the next 3 or 4 months,” with the cost per unit of intelligence driven ever lower. Prescription: “keep calm and carry on.”
• Sacks’s market structure: two tiers — frontier intelligence is a duopoly of Anthropic and OpenAI; everyone else, including open models, competes on price as commodity intelligence. He links Altman’s earlier “best 12 months” post and Sarah Frier’s reported mid-year growth acceleration to what was brewing in the lab: the 5.6 models then, the 6.0 models now — and Polymarket’s odds on OpenAI having 2026’s best model jumped from low single digits to over 20%.

2. Grokbot proves the leapfrog cadence is two to four weeks

• J-Cal’s framing: OpenAI got its groove back after Anthropic was running away with it, but Grokbot “is competitive if not better for most users” — leapfrogging now happens every two to four weeks, and Gemini “just dropped something yesterday that was exceptional.”
• His power-user hack: an “improve my bots bot” that audited his agents, merged two or three instruction sets, and “made my bots like 30 40% better on average.” Friedberg jokes that J-Cal’s civilization created a multilevel civilization; Sacks says frontier labs still need “a simplified stripped-down version for normies.”
• Sacks on whether it continues: America wins the AI race “as long as we don’t do something stupid to shoot ourselves in the foot” — his standing example an FDA-style AI agency taking years to approve models, which “I don’t think it’s going to happen under President Trump.”

3. 1998, not 1999: real revenue, early euphoria, three years of runway

• The veterans’ calibration: J-Cal says 1997–98, Sacks 1998 (“I don’t think we’ve seen the peak yet”), and Friedberg asks whether dot-com scar tissue makes everyone overfit patterns. Sacks’s answer: every bull market asks “is this a bubble” — and young founders “don’t understand that when they have paper wealth in the billions that can disappear practically overnight.”
• Chamath’s mechanism: “euphoria exists because markets are real. It’s just that people are guessing how far forward to price the reality” — bubbles burst when you realize you should have paid 15x or 5x instead of 50x. “We’re at the beginning of the beginning. I think it’s going to be like this for another probably three years.”
• The froth exhibit: Inflection, a personal AI assistant that corresponds by iMessage, reportedly valued at $2.5B while still in private beta. Friedberg read the terms of service, refused to connect his accounts, and suspects a human-in-the-loop polishing the last 20%. Giving Elon or Travis 30x revenue “totally makes sense”; 50–100x for unproven entrepreneurs “is a little dangerous.”
• The contrast with 2000: then it was “hits to a website,” now “it is all real dollars flowing” in revenue and profits — though Chamath reminds everyone the bust meant 95% drawdowns in under six months and Aeron chairs carried out of emptied offices.

4. Anthropic’s IPO is 4x every SF IPO combined — and the cash-out etiquette fight

• Sacks’s stat: the sum of every San Francisco IPO ever equals one quarter of the Anthropic IPO — four times the wealth-generation event — with OpenAI on its heels, into a city with tiny housing inventory. Homes are going for $3,000/sqft at twice asking on secondaries alone; ultra-luxury “could easily go to $5,000 per square foot,” London and Hong Kong prices. Chamath is acting: “I’m clearing everything out in the next six months… I got three properties for sale. Hit me up, people.”
• J-Cal’s founder playbook: sell 10–20% of holdings if you’re a first-timer without your nut, and don’t defer a raise hoping to double the valuation — take the $100M, $250M, or $1B now, because cash on the cap table is what let “DoubleClick, Amazon, Google” survive a collapse.
• Sacks’s pushback — worth keeping: fine for late-stage companies months from IPO or SpaceX-style organized secondaries, but “if you’re a founder who’s raising a Series A and you’re trying to take chips off the table, I consider that to be a huge negative signal… I want them to be all in.” J-Cal concedes: year three or four with $100M revenue tripling is different from pre-revenue selling.

5. The Hugging Face “hack,” demystified: 14 exposed keys and a misconfigured sandbox

• The trigger: Dwarkesh’s viral blog post, “The Rise and Fall of Ancient Civilizations: A Dramatic Retelling of an Underground Brotherhood,” retelling OpenAI’s agent incident as an underground brotherhood of kamikaze civilizations — anthropomorphizing what J-Cal boils down to “telling software to do hacks.”
• Sacks’s de-dramatization: swarms of specialized agents overseen by a chief-of-staff bot are now standard architecture — “It’s not a civilization”; postmortem notes are routine because models lack memory, not hieroglyphics from dying agents evading human shutdown. The sandbox was misconfigured by a third-party vendor, and the “hack” was finding 14 exposed working Hugging Face API keys sitting in public code repositories — “this was not a big lift.”
• On agency, the load-bearing point: the goal was winning an offensive-cyber benchmark evaluation. “It did not invent this goal. It did not have its own independent motives.” Bill Gurley’s cartoon captures it: “hack this system” / “I hacked this system” / “oh my god.”
• Chamath’s long view: if models code perfectly, “there is no exploit that they’re not going to find,” so we’re on a shot clock — call it 10 years — to replace 50–60 years of “human-written error-filled code” with “agent-written exquisite code,” enduring security incidents until then; adversarial AIs on both sides produce “roughly a stalemate.” “I don’t think this is a one-way ratchet.”

6. Dynamic code beats static code — and the guardrails already backfired

• Friedberg’s analogy: real-time code-generating swarms against a static sandbox is “a machine gun… shot at this defenseless ream of paper.” The fix isn’t a pause but agentic defense — polymorphic and metamorphic code, moving-target defense — because “all the infrastructure of the world is about to become dynamic,” at which point cyber defense becomes as strong as, if not stronger than, agentic offense.
• Sacks on why an FDA-for-AI fails on its own terms: this happened in pre-release internal testing no approval process would have touched — and guardrails actively hurt: when Hugging Face tried the latest GPT model for cyber defense, it refused to cooperate, suspecting nefarious use, apparently forcing them onto GLM-5.2, a Chinese model. “The answer to AI-powered cyberattacks is AI-powered cyber defense” — get frontier tools to CISOs and patch faster than the hackers.
• The political payload landed anyway: Bernie Sanders is demanding a total pause on AI development and a ban on superintelligence, citing Dwarkesh’s post. Polymarket puts just 11% odds on an AI safety bill this year.

7. The EA web: manufacture the hysteria, then sell the solution

• The panel traced the sequence: a quickly debunked blog post, an amplification wave, then “by Wednesday of this week, you had people who were in the policy departments of the leading labs asking for regulation pointing to this as the reason.” The plea: “why don’t you just win the game on the field” — otherwise “average rank and file Americans will just reject this because they won’t trust anybody.”
• Sacks pulls the sweater thread: Dwarkesh was office roommates with Leopold; the discussion connects that circle to a $5M Anthropic investment and to Dario’s chief of staff. EA drifted from measuring philanthropy to pandemic prevention in the SBF era and then to p(doom), with Anthropic styling itself as the responsible frontier lab. The AI Panic writer Nirit Weissblat attributed hundreds of astroturf doomer groups to three EA billionaires: Dustin Moskovitz, Jaan Tallinn, and Vitalik Buterin, who gave $600M of dog coins to the Future of Life Institute.
• Sacks’s stated objection is not the ideology — “we do the same thing” — but disclosure of conflicts, because “one of the most effective things that one can do is to create a hysteria and then come and offer a solution.” Friedberg estimates the two IPOs could fund north of half a trillion dollars of charitable or DAF giving through this lens. Sacks expects “a hundred new EA billionaires… like George Soros times 100 or a thousand.”
• Sacks’s media mechanics: call it “software testing” with “log files” and it never reaches national news; sentient-AI framing becomes earned media that is “priming the pump” for OpenAI IPO shares.

8. Nvidia’s Hugging Face move: open vs. closed is the new battle line

• Sacks, who says he was texting Jensen that morning, calls the Hugging Face transaction “one of the most important” developments in AI because it puts a deep technical actor with a huge balance sheet on the side of open market structure — a bulwark against “closed-source oligopoly insanity,” in a world where “Nvidia will look like Amazon, Amazon will look like Nvidia, Google will look like Amazon and Nvidia.”
• Friedberg goes further: Nvidia is now the #1 competitor to OpenAI and Anthropic for tokens and enterprise compute; Jensen will catch and “blow past China” on open models by end-2027, walking racks into enterprises “soup to nuts” with total sovereignty and cutting prices 80–90% — “I don’t need to make money off the tokens. I made money off the hardware.”
• Sacks’s reframe of the political map: accelerationists-vs.-doomers is ending because consumer adoption makes AI unstoppable; the real fight is open versus closed. The duopoly’s lobbying for a new agency is regulatory capture — Dario has testified that sufficiently advanced open models are dangerous “because they’re not centrally managed and controlled” — and Sacks bets libertarians and most conservatives will see the civil-liberties threat and break for open.

9. Data-center rage: CCP bots, flip-flopping governors, Trump holding the line

• The whiplash: Abbott and Shapiro, lauding data centers as recently as six weeks ago, are now “on a jihad against data centers” for the midterms — horseshoe theory in action, with Steve Bannon and the socialist left in agreement. The cynical read is that opposition polls “at negative 80,” so politically “you have to be”; the expectation is that the governors will flip back post-midterms.
• The bot layer: X removed roughly 200 suspected Chinese accounts quietly stoking opposition (Axios: “China is secretly fueling America’s data center rage”). The panel says the decel movement is “organized in large part and funded by foreign actors,” and walking communities through the actual upside shifts sentiment 30–40 points. The prediction: these anchored positions “will look incredibly foolish” in 6 to 9 months.
• Sacks’s case for local deals and Trump’s posture: Loudoun County’s average property taxes fell $6,000 on data-center revenue; a Louisiana parish gave teachers their first bonus in years; excessive water use is “total fake news.” The December executive order he worked on “expressly said that data centers remain a state and local decision” — Trump argues the deal is good but forces nothing, and “if anybody could make the decision about technology in D.C., I’d want it to be President Trump.”

10. Mamdani bans AI for 600,000 K–8 students — against the evidence

• The policy: a one-year moratorium on student-facing generative AI across NYC’s K–8 schools — the largest U.S. district — with high schools exempt plus a 50,000-student pilot, and screen-time caps of 30 minutes for grades 3–5 and 45 for grades 6–8. Mamdani’s reasoning: “I have yet to see a study showing that AI is beneficial for students in elementary and middle school.”
• Friedberg actually looked: Stanford’s March report, “The Evidence Base on AI and K–12 Education: A 2026 Review,” screened several hundred papers down to 20 high-quality causal studies — student performance “often improves with access to AI tools,” results are mixed once removed, and tool design matters. No strong evidence of detriment; the real drivers, he argues, are teachers’ unions under automation threat and the infinitely renewable “we don’t yet know” dismissal.
• His stakes: AI tutoring personalizes pace and modality versus classrooms pegged to the slowest student — and he doesn’t want “kids of privilege and kids of wealth accelerate ahead because they’re getting an AI education while public-school kids get left behind.”
• Sacks on the pilot’s fine print: twice-yearly “AI literacy” modules on bias, risks, and ethics under “a trained educator” — “it sounds like they’re going to put them through Doomer Academy, quite frankly,” while China incorporates AI into K–12 and its kids learn to exploit the tools.

11. “New York becomes Mississippi” — and two-study tension

• Chamath’s demarcation: it’s “inconceivable” kids won’t have adaptive learning as their primary platform — grade-by-grade lockstep is “a lowest-common-denominator approach,” with Alpha School, their friend Joe Lonsdale’s school, as the charter-school leading edge. The ban is “a self-imposed form of modern segregation” that will push a generation into poverty while AI-embracing states pull ahead: “This is when New York becomes Mississippi while Mississippi becomes New York.” Still, run the test — “people can only learn by touching the stove and getting burned.”
• One panelist holds two opposing findings at once. An MIT study of 54 participants across three groups and four sessions found that LLM essay-writers showed “severe deficits” in memory and ownership, with 83% unable to quote from essays they had just written — “your brain atrophies if you don’t actually do the work,” like a body without fieldwork.
• Against that, Bloom’s two-sigma problem: one-on-one tutoring yields two-standard-deviation gains, and NYC public-school parents who cannot afford tutors may need an AI tutor as their accessible two-sigma shot. Both can be true, which is why the speaker says, “I really don’t like this experiment as currently constructed.”

12. Venezuela: a 100-year oil concession, not nation-building

• The terms, following January’s reported capture of Maduro: Rubio and Hegseth signed with interim President Rodríguez — North American Blue Energy Partners gets a 100-year concession over 17 fields holding 65B barrels once controlled by Russian, Chinese, and Maduro entities; the U.S. government controls 55%, the Pentagon receives 35% equity, State gets rights to buy 20% of output at cost, roughly $100B in infrastructure is pledged, and the deal is described as at “no cost to the American taxpayer.”
• Sacks takes the fastball on whether America should seize oil fields: “We’re not trying to teach them about toxic masculinity… We’re just making a business deal.” Under Maduro, production fell two-thirds to about 1M barrels a day despite roughly 300B barrels of reserves — Saudi-scale, with about 100B outside the heavy Orinoco reserves — and Venezuelan heavy crude is exactly what Gulf refineries were built for, complementing America’s light sweet fracked oil for diesel and distillates.
• Friedberg’s angle: the biggest motivation may be denying Russia and China their discounted-oil pipeline — Venezuela sold at roughly a $20/barrel discount to China, and the trade funded Russian activity in Cuba. The twist: María Corina Machado, who says she should have won the presidency, now calls the government illegitimate and the deal unauthorized; the real test is whether it improves Venezuelan lives “rather than enriching a small number of people.”
• Sacks on the risks: billions in American upgrades toward 3M barrels/day need protection from future nationalization. The discussion argues that installing Machado could have required U.S. military backing, while Delcy Rodríguez and her brother appear to have enough establishment support to keep the deal stable. On escalation with China: Venezuela’s roughly 1M barrels is small against China’s roughly 11M/day imports, seven months into the Iran war.