2025-09-26

The Latest in AI: Job Loss, Elon & Sam Altman Chip Race & the "AI Bubble" w/ Brian (Blitzy) & Emad

Peter Diamandis, Dave Blundin, Emad Mostaque and guest Brian Elliot (CEO of Blitzy) open with a wakeup call on college's collapsing perceived value, then dig into the AI model wars (Gemini overtaking ChatGPT, Grok 5's AGI chase, Qwen's rapid catch-up) and a looming compute scarcity driven by gigawatt-scale data center buildouts from OpenAI, xAI, and Meta. They argue the current AI spending wave is not a dot-com-style bubble because revenue tracks capex in real time, debate whether a 3-4 day work week is coming or just CEO spin to avoid panic, and cover humanoid robotics, the US-China solar gap, OpenAI's nonprofit-to-for-profit restructuring, and AI's accelerating impact on drug discovery and healthcare.

▶ Watch on YouTube

Topics

College's collapsing value proposition Economy ▶ 1:40

Perceived value of a college degree has dropped from 75% to 35% since 2010 as tuition inflates (180% since 2005, 900% since 1983) while curricula lag what AI has made learnable; the panel argues what matters most is admission brand, not graduating, and AI-native programs like Alpha School are already outperforming.

The AI compute scarcity era Compute ▶ 16:00

xAI's Colossus 2, the $100B Nvidia-OpenAI deal (10GW / 4-5M GPUs), and Zuckerberg's $600B pledge are framed as the start of a structural compute shortage, with demand growing 10x a year against much slower supply growth.

Is AI a bubble? Economy ▶ 1:14:15

Panel compares Nvidia's stock (earnings tracking price) to Cisco's dot-com-era pattern (price detached from earnings) to argue AI spending is not a bubble, while warning that tokenized digital assets are the real bubble to watch.

Model wars: Gemini, Grok, Qwen AI ▶ 12:04

Gemini overtakes ChatGPT in US iOS app-store rank on Google's distribution power; Grok 5 chases AGI benchmarks; Alibaba's Qwen closes the gap fast on reach and near-daily releases despite smaller compute budgets.

Work, jobs, and the 3-day work week debate Economy ▶ 1:20:44

Zoom, Gates, Nvidia and JPMorgan CEOs predict a shorter work week; the panel is skeptical about who actually benefits, cites a study where human-in-the-loop doctors performed worse than GPT-4 alone, and argues tech leaders have shifted from candor about job losses to reassurance to avoid public panic.

AI agents can finally plan over long horizons AI ▶ 46:23

A Mustafa Suleyman clip describes today's models as one-shot predictors without long-horizon planning; the panel counters that planning capability has effectively already arrived at the application layer (self-driving road trips, Blitzy's engineering agents).

Humanoid robotics and embodiment race Robotics ▶ 1:09:02

Figure raises $1B at a $39B valuation with a Brookfield partnership for data-gathering access to homes and offices; OpenAI re-enters robotics hiring; panel estimates 1-10 billion robots by 2040, far above Morgan Stanley's $5T-by-2050 forecast.

OpenAI's nonprofit-to-for-profit restructuring Economy ▶ 32:12

OpenAI targets a $500B valuation as it converts to for-profit, leaving its nonprofit arm with roughly $100B (larger than Harvard's endowment); Microsoft's ~30% stake and shifting leverage with OpenAI are discussed alongside Elon Musk's rivalry with Sam Altman.

Energy: solar's quiet takeover vs the US-China gap Energy ▶ 1:04:16

US Secretary of Energy dismisses solar's future even as DOE's own NREL projects 40%+ of US electricity from solar by 2035; China's 880GW installed base dwarfs the US's 177GW, exposing a US solar supply chain dependency on China.

AI drug discovery and personalized medicine Biotech ▶ 1:35:00

DeepMind's Demis Hassabis says AI could shrink drug discovery to months; 150 AI-discovered molecules emerged in 2025 alone with 80-90% phase-1 success, and one drug went from design to human trials in 12 months versus the normal 4-5 years.

AI glasses and the next data flywheel Robotics ▶ 57:07

Amazon's 'J-Hawk' AI glasses for delivery drivers (100,000-unit pilot across a 390,000-driver workforce) are built explicitly to harvest training data for future robots, following Meta's Ray-Ban smart glasses.

Tokenization of markets Economy ▶ 1:28:21

NASDAQ pushes to trade tokenized securities; Robinhood's EU arm already offers tokenized stock access to private companies like OpenAI and SpaceX; panel sees AI-driven compliance as the key unlock for a blockchain-based alternative to the IPO process.

Predictions made

open Alex Wissner-Gross: AI will have 'solved' all of mathematics via parallelized formal proof systems within about 18 months.

· due: 2027 · ▶ watch

“Alex was predicting I think it was 18 months we'll be solving all math.”

Your call:

open Dave Blundin: The ARC-AGI-v2 benchmark will fully saturate within three years, implying beyond-superhuman intelligence.

· due: 2028 · ▶ watch

“If this one saturates in three years, we're in another universe.”

Your call:

open Mark Zuckerberg: Meta will invest $600 billion in US data centers by 2028 to avoid falling behind on superintelligence.

· due: 2028 · ▶ watch

“Better to lose billions than be late to super intelligence.”

Your call:

open Mustafa Suleyman: AI will gain the ability to take actions and plan over long time horizons, not just answer one-shot questions, by the end of 2026.

· due: 2026 · ▶ watch

“I think that we basically have that by the end of next year.”

Your call:

open Emad Mostaque: Ordinary stocks will be tradable on the X platform within about a year.

· due: 2026 · ▶ watch

“I think you'll be able to trade stocks on X by next year.”

Your call:

open Emad Mostaque: Humans will carry negative value on cognitive-labor teams that include AI within a few years, since the weakest team member drags down performance.

· due: ~2028 · ▶ watch

“Humans will have negative value in cognitive labor in a few years.”

Your call:

open Peter Diamandis: Between 1 billion and 10 billion humanoid robots (per Elon Musk and Brett Adcock's high-end case) will exist by 2040, implying a $25-250 trillion market far above bank forecasts.

· due: 2040 · ▶ watch

“The low end of this is a billion robots by 2040.”

Your call:

open Emad Mostaque: Part of the FDA drug-approval process will incorporate in-silico (AI-simulated) predictions of trial outcomes within five years.

· due: 2030 · ▶ watch

“I reckon in the next five years you might have part of the FDA process be in-silico predictions of these drug trials.”

Your call:

Numbers that matter

College value perception: 'very important' fell 75% (2010) to 35% (2025) Chart Peter presented on Americans' declining view of a college degree's importance.
Tuition up 180% since 2005 (5.6%/yr average) Driver of rising student debt without matching job outcomes.
Tuition up ~900% since 1983 Cumulative college tuition price change vs other expenses since 1983.
Oxford tuition: $13,000/yr domestic vs $60,000/yr international Emad citing UK tuition as unsustainably expensive relative to value.
xAI Colossus 2: gigawatt-scale data center, 110,000 GB200 GPUs, 119 air-cooled chillers Elon Musk's latest data center buildout, part of his stated goal to be first to 1, 10, then 100 gigawatts.
Nvidia-OpenAI $100B deal ≈ 10GW / 4-5 million GPUs Equal to everything Nvidia sold in all of 2024 and double what it sold the year before, per Jensen Huang.
Lab compute tripled in one year; data center capacity to go from 44GW to 156GW by 2030 Panel calls the 2030 projection a likely lowball estimate given current demand trends.
US solar: 18GW installed H1 2025, 50% of new US generating capacity, now 4th-largest US electricity source Cited to rebut the Secretary of Energy's dismissal of solar's growth.
China: 880GW solar capacity in 2024, growing 45.6%/yr, vs US 177GW growing 27%/yr Illustrates the widening US-China solar manufacturing and deployment gap.
Robotics market: Morgan Stanley forecasts $5T by 2050 vs panel's 1-10 billion robots ($25-250T) by 2040 Panel argues bank estimates ignore self-improvement and self-manufacturing feedback loops.
Full Waymo-level self-driving could save 40,000 lives/year and $1 trillion in societal costs Cited as an example of how a human in the loop can be a net negative.
Hypertension affects 1.5B adults (30-45% of population, up to 60% over 60); 46% undiagnosed, only 21% controlled Context for Apple Watch's new FDA-cleared hypertension alert feature.

Worth digging into

🕳️ Alpha School and AI-native education models

Panel claims two hours a day of AI tutoring puts kids in the top 0.5% globally, a striking claim worth checking against the school's own published outcomes and the Erik Brynjolfsson early-career-job-loss paper they also reference.

🕳️ OpenAI's nonprofit-to-for-profit restructuring mechanics

Panel claims the nonprofit retains ~$100B, making it the world's second-largest endowment, a consequential structural move with limited public detail discussed.

🕳️ Physician + GPT-4 diagnostic accuracy study (74%/76%/92%)

A striking claim that human-in-the-loop clinicians perform worse than AI alone has major implications for how AI gets deployed in medicine.

🕳️ US vs China solar manufacturing gap

Panel frames solar as the US's best shot at keeping pace with China but says the US is already badly behind (880GW vs 177GW) due to supply-chain dependence on China.

🕳️ AI-designed drugs in human trials (Insilico Medicine, DSP-1181)

150 AI-discovered molecules in 2025 alone and a 12-month design-to-trial timeline would be a major acceleration in drug development if verified.