AI Bubble Pops, Zuck Freezes Hiring, Newsom’s 2028 Surge, Russia/Ukraine Endgame
Summary
The AI drawdown looks like a sentiment reset, not yet the bust of the investment supercycle. Sacks called the roughly 10% correction “healthy”: GPT-5 advanced the frontier but fell short of its teased “Death Star” expectations, while clustered benchmarks and continued model leapfrogging weaken the rapid-takeoff thesis. His replacement frame is “more evolutionary rather than revolutionary.”
Enterprise AI is finding ROI in narrow, back-office workflows while generalized pilots stall. The cited MIT study found 95% of pilots were not reaching production, with 70% of budgets aimed at lower-return sales and marketing projects; specialized vendors, by contrast, reportedly succeeded two-thirds of the time. Chamath expects “a sorting and a cleansing” as foundation models absorb features, cheaper competitors arrive, and brittle ARR churns.
The investable architecture is shifting from autonomous general models toward human-AI systems, deterministic software, and networks of SLMs. Friedberg argued that AI-generated code still needs engineers, while generated film assets work better when rendered and controlled through software such as Unity. If specialized-model architectures become 10x or 100x more efficient, current data-center capex could earn much higher ROIC even as token production surges.
Adoption friction may matter as much as model capability. Friedberg said 8090 expects about $40 million of first-year bookings from back-office work, yet its first customer fired it despite “enormous gains” because resistance intensified deeper inside the organization. JCal similarly sees AI making SDRs “literally 10 times more effective,” but incrementally and only where employees embrace it.
Meta’s hiring pause marks digestion after a strategic talent panic, while OpenAI can still support a bull case at $500 billion. Sacks said extraordinary $100 million jobs and multibillion-dollar pre-product bids appear only when cash-rich incumbents feel strategically vulnerable—and warned founders that this is not “the normal state of the world.” After Friedberg cited roughly 700–750 million weekly active users, Chamath’s rough case used an estimated 500 million DAUs doubling to 2 billion in four years and one-tenth of Facebook’s revenue, implying about $1.5 trillion.
Newsom’s early 2028 lead reflects Democratic demand for a fighter, but the panel sees his California record as the tradeoff. He stood at 25% among surveyed California Democrats and left-leaning independents and 28% on Polymarket versus AOC at 14%; Sacks called his Trump-like presentation “synthetic or plastic,” while Chamath warned early leaders “never win.” The sharper electoral argument centered on wages, housing, education, healthcare, affordability, and whether a socialist can win a Democratic primary but not the Electoral College.
The proposed Ukraine endgame is comprehensive peace, no NATO membership, and territorial concessions—but Sacks framed Zelensky, Ukraine’s leadership, and some European governments as the blockers. JCal agreed NATO could come off the table for 20 years but insisted territorial choices belong to Ukrainians; he moved from a generic 5–10% chance for presidential diplomacy with dictators to a 50–50 estimate for Trump. Chamath’s warning was historical: protracted territorial conflicts can freeze behind demilitarized lines rather than reach clean settlements.
Deep dive
1. Enterprise AI bought pilots before it found the work
The cited MIT study examined 300 implementations and interviewed 150 leaders across 52 companies. Its headline finding: 95% of generative-AI pilots were failing to reach production, while 70% of budgets flowed toward sales and marketing tools despite back-office automation producing the strongest ROI.
Chamath traced the first wave to boards reading “AI,” asking CEOs for an AI strategy, and pushing the mandate down to CTOs with large budgets. Much of the resulting spending was experimentation without a sufficiently defined operational problem.
His key distinction was probabilistic versus deterministic software. Sales and marketing resist fixed heuristics, whereas back offices employ people largely to handle process edge cases—work where correctly implemented AI can achieve “extremely high rates of accuracy.”
Chamath expects the familiar platform purge: Facebook’s social ecosystem went from “seven or 8,000 social companies” to five survivors within six years, while SaaS winners also took years to emerge. AI now faces “a sorting and a cleansing” through logo churn, dollar churn, feature absorption, and failed deployments.
2. GPT-5 punctured rapid takeoff, not the investment cycle
Sacks characterized the roughly 10% public-AI correction as a “healthy dose of skepticism,” not the start of a bust. He still sees an investment supercycle, but one requiring more work than the claim that AI will recursively improve itself into superintelligence.
After ChatGPT’s 2022 launch, he argued, predictions of AGI within two or three years fueled both utopian job-replacement stories and doomer scenarios. That same narrative helped provoke roughly 1,000 state bills and measures such as California’s SB 1047.
GPT-5 crystallized the reset: Altman teased a “Death Star,” but reviews were mixed and benchmark gains fell short of expectations. Sacks also warned that incremental model updates can become overfit to benchmarks; he saw clustered performance and continued leapfrogging as evidence of a normal technology race—“more evolutionary rather than revolutionary.”
3. Specialized models move value toward the last mile
Sacks pointed to growing differentiation rather than one all-powerful model: ChatGPT was accused of being sycophantic, Grok presented as “more based,” Google excelled in video, and Anthropic in coding. “The specialization really belies this idea of one model becoming all knowing and all powerful.”
General models still need enterprise context, detailed prompts, connected data, validation, and iteration. Sacks called the gap between 90% and 99% the “last mile”—the industry-specific work where usable business value is created and vertical applications can retain economics.
The study’s sharper counterpoint to its 95% failure headline was that specialized vendors succeeded two out of three times. Friedberg offered TaxGPT, one of the companies his group incubated, as a copilot built specifically for tax professionals and CPAs—the kind of constrained problem that can outperform generic prompting.
Friedberg sees a movement toward SLMs: smaller task-specific models working alone or in networks. If future models design those networks themselves, their internal logic may become hard to explain, but 10x–100x efficiency gains could sharply reduce energy and dollar cost per token while increasing total token production.
4. Generation works best as a layer, not an autonomous business
Friedberg’s first implementation lesson was human pairing: AI can generate code, but engineers still must debug it, integrate it, and move it into production. The organizational question is therefore “where does it fit,” who operates it, and how existing teams use it.
His filmmaking example paired generative systems with Unity: AI creates objects, while the deterministic renderer controls camera angle, framing, color, lighting, and character consistency. The model becomes a “marionette” layered over established software rather than an instruction to “make a movie.”
JCal placed AI in the “trough of disillusionment,” comparing it with Uber and DoorDash before small unit-economic improvements and network effects unlocked scale. On the ground, he sees SDRs becoming “literally 10 times more effective” at finding and organizing leads, not salespeople disappearing wholesale.
Chamath’s open question is whether incumbents become trapped by tens of billions invested in LLM talent, opex, and capex if state-space approaches, custom silicon, or another representation wins. The field is only a few years into what he expects to be a “multi-decade innovation cycle.”
5. Deployment politics and strategic bids will cleanse AI revenue
Friedberg said 8090 expects roughly $40 million of bookings in its first full year, entirely from back-office applications. Yet the first customer, despite receiving “enormous gains,” fired the company when internal resistance grew lower in the organization; escalation eventually resolved the dispute.
That experience makes the “AI boogeyman” an operating variable across the Fortune 2000 and Global 2000. Technical accuracy alone cannot secure renewals when threatened employees or decision-makers can block adoption, adding organizational risk to already uncertain logo and dollar retention.
Meta’s pause came only eight weeks after its talent blitz: the hosts cited its Scale AI team deal and $14 billion investment, pursuit of Ilya Sutskever’s company, and alleged $100 million offers for OpenAI talent. Sacks called the freeze digestion, not a popped bubble: such offers require a cash-rich incumbent that feels strategically vulnerable.
After Friedberg cited roughly 700–750 million weekly active users, Chamath’s rough $500 billion OpenAI bull case used an estimated 500 million DAUs doubling every two years to 2 billion in four years. At one-tenth of Facebook’s revenue, he modeled about a $1.5 trillion valuation; Sacks added that subscriptions, consumer leadership, and search substitution provide actual revenue support.
6. Newsom leads because Democrats want a fighter, but early leads decay
Newsom registered 25% among surveyed California Democrats and left-leaning independents, while Polymarket put him at 28%, up about nine points, versus AOC at 14%. Sacks read the surge as Democrats rewarding someone mirroring Trump’s fighting style.
His objection was authenticity: Trump has a “trillion dollar personality” and is always recognizably himself, whereas Newsom’s imitation looks “synthetic or plastic.” More importantly, Sacks argued, Trump’s durability comes from issue positions rather than presentation alone.
Friedberg expects socialist momentum to persist and suggested that some on the Republican side might prefer facing that candidate. Chamath recalled Ron DeSantis leading early and warned, “These people never win”; in his formulation, a socialist can win the Democratic primary but cannot win the presidency under current electoral dynamics.
7. Household economics and governing records will frame 2026 and 2028
Chamath expects the midterms to be an economic referendum encompassing security, immigration, tariffs, and especially rates. He described Powell as “holding on white knuckled” to a position the data may not support, making monetary policy an unusually large swing factor.
JCal’s proposed Democratic attack centered on stagnant real wages and unaffordable education, housing, and healthcare, paired with a pro-immigration message. Sacks then proposed adding $1 to the federal minimum wage every year for eight years as a politically effective playbook, not necessarily an endorsement.
Chamath countered that wages rose significantly in Trump’s first term and cited reshoring-oriented trade policy, the tax bill, and no-tax-on-tips. JCal also invoked no-tax-on-overtime. Sacks said such measures help the middle and working classes; JCal replied that giveaways motivate voters and Democrats could “10x” Trump’s method.
Sacks’s record case against Newsom framed California as having the highest taxes for the worst public services. He cited a claimed swing from a $75 billion surplus to a $20 billion deficit, along with the country’s highest rates of poverty, homelessness, inequality, illiteracy, and wage stagnation, its third-highest unemployment, and high housing and energy costs.
JCal’s Grok comparison ranked Wes Moore’s Maryland first, Gretchen Whitmer’s Michigan second, and Gavin Newsom’s California third on quality of life, cost of living, and crime.
8. Trump’s Alaska summit narrowed peace to three disputed pillars
Sacks called the Alaska meeting major progress because it restored presidential-level US-Russia diplomacy absent since 2021. Citing Gallup, he said Ukrainian support for continuing the war had fallen from roughly 70% a year or two earlier to 24%, making peace more aligned with public sentiment.
His first pillar was a comprehensive settlement rather than a ceasefire. Russia sees a pause as time for Ukraine to regroup and rearm, and Sacks’s blunt formulation was: “When you’re fighting a war and losing, you don’t get to call a timeout.”
The other pillars were no Ukrainian NATO membership—which Sacks described as the central prewar friction—and territorial concessions reflecting battlefield control. He blamed the prospective losses on the failed 2023 counteroffensive despite roughly $100 billion in weapons, not on Trump’s negotiations.
Sacks said Zelensky, Ukraine’s leadership, and some European governments continued demanding a ceasefire, NATO access, and no territorial concessions, including over Crimea. Because “the Russians and the Ukrainians have to agree,” he credited Trump’s diplomacy but doubted a deal without Ukrainian and European compromise.
9. History and sovereignty leave diplomacy with narrow odds
Chamath found that unresolved territorial wars can harden into demilitarized lines: Korea since 1953, Cyprus since 1974, and Kashmir since 1947, alongside the Israel-Lebanon boundary. Crimea since 2014 and East Jerusalem illustrated territory taken without international consensus; overcoming that pattern would be “truly unprecedented.”
JCal praised Trump for talking to Putin, tightening sanctions, threatening India over Russian oil, selling more weapons, obtaining mineral rights, and pressing NATO members to pay more. He still called Putin a dictator and war criminal and initially placed any president’s chance of success at only 5%–10%.
Sacks pushed back that moralistic name-calling undermines diplomacy and that pressure must also reach Zelensky. JCal agreed NATO membership could be removed for 20 years in exchange for security guarantees, but maintained that territorial concessions are Ukraine’s sovereign decision.
Sacks raised a conflict of interest: Ukraine’s ruling elite canceled elections and remains in power while war-related money continues to flow, despite citizens wanting peace. JCal said leaders must heed those citizens but America cannot decide for them; by the close, he upgraded Trump specifically to a 50–50 chance and insisted, “I call balls and strikes.”