2025-09-03

AI Insiders Reveal Elon Musk's Master Plan to Win AI w/ AWG & Dave Blundin | EP #192

Peter Diamandis, Alex Wissner-Gross, and Dave Blundin (Salim Ismail was traveling in India) run through a dense week of AI news: Musk's Colossus 1/2 data-center buildout and xAI's ultra-cheap Grok Code Fast 1 versus Meta's researcher exodus back to OpenAI, framed through Richard Sutton's 'bitter lesson' now applied to brute-force hardware scaling. They dig into Google's Nano Banana image model and live-translate feature disrupting Adobe, Duolingo, and Chegg, debate whether the frontier-model value chain accrues at the chip/data-center layer or the application layer, and weigh the NASDAQ's dot-com-beating valuation as a signature of imminent superintelligence. The back half covers health and longevity (an AI stethoscope, a psilocybin longevity study, Peter's own stem-cell re-education treatment), humanoid robotics (Nvidia Jetson Thor, Tesla Optimus going vision-only, Waymo dwarfing Uber's per-driver efficiency), and closes on rising US electricity prices as a leading indicator of AI's energy hunger and the prospect of recursive self-improvement.

▶ Watch on YouTube

Topics

Colossus 2 and the AI data-center arms race Compute ▶ [10:22]
Elon's Colossus 2 (1 gigawatt, 500,000 Blackwell GPUs, doubling again in 2026 for Grok 5) is neck-and-neck with OpenAI's Stargate; the panel argues this is winner-take-all, not a five-to-seven-lab equilibrium, and notes Colossus runs on self-generated natural-gas power rather than the grid.
The bitter lesson applied to hardware AI ▶ [11:05]
Alex explains Richard Sutton's 'bitter lesson' (scale beats hand-crafted algorithms) and argues it now applies to brute-force hardware scaling too, with the software/algorithm layer being the next place to expect similar brute-force gains.
Grok Code Fast 1 price war AI ▶ [20:31]
xAI's coding model undercuts GPT-5 and Claude Sonnet 4 by 10-15x on token price; Dave argues this isn't a race to the bottom but a 'free first hit' market-share grab because coding-AI demand is effectively infinite.
Musk poaches Meta talent; purpose vs. cash AI ▶ [25:50]
Elon lures 14 Meta engineers while Meta's own superintelligence lab loses researchers back to OpenAI despite nine-figure pay packages; the panel argues mission and a track record of delivering (not money) is what retains top AI talent.
Google Nano Banana image model AI ▶ [35:44]
Gemini 2.5 Flash Image ('Nano Banana') edits photos via natural language with strong character/scene consistency at ~3.9 cents/image; Alex argues it's a distilled sliver of a larger world model rather than just a Photoshop killer.
Google Live Translate disrupts language-learning stocks Economy ▶ [48:13]
Gemini-powered real-time conversational translation lands the same week Chegg (down from $90 to $1.40) and Duolingo (down 10% on the announcement) get hit, illustrating how frontier models cannibalize narrow SaaS products from below.
OpenAI Realtime API and 'streaming interactive models' AI ▶ [55:29]
GPT Realtime powers low-latency voice agents (demoed on a Zillow-style home search); Alex proposes 'streaming interactive models (SIMs)' as the missing term for a future where every pixel and audio segment is generated live on demand.
Nvidia earnings and the US-China AI chip race Compute ▶ [1:01:16]
Nvidia posts 56% YoY revenue growth and a $4T market cap even as it cedes 15% of China sales to the US government; simultaneously investors back China's Cambricon, which Alex reads as the start of a post-Nvidia, post-CUDA multi-architecture world.
Export restrictions accelerate Chinese chip innovation Geopolitics ▶ [1:04:44]
Dave and Peter argue that restricting chip exports to China (as with satellites before) only forces indigenous competition, and that software/algorithmic gains (100-1000x) matter more than the shrinking 7nm-vs-2nm hardware gap.
India's AI infrastructure push Geopolitics ▶ [1:07:00]
Mukesh Ambani launches Reliance Intelligence Ventures with Google and Meta to build India's AI backbone, echoing his disruptive Jio telecom playbook; Dave notes OpenAI's parallel push into India for its uniquely young, untapped talent pool.
NASDAQ valuations vs. the dot-com bubble Economy ▶ [1:19:20]
NASDAQ market cap hits 176% of US money supply and 129% of GDP, above dot-com peaks; Alex frames the ratio spike as exactly what you'd expect on the eve of superintelligence rather than a pure bubble, since underlying AI products are real and profitable.
AI-driven health and longevity breakthroughs Longevity ▶ [1:23:02]
Covers a UK AI stethoscope detecting heart disease in 15 seconds, a 2025 mouse study where psilocybin/psilocin extended survival to 80% vs 50% (57% lifespan gain) via telomere effects, and Peter's own stem-cell re-education treatment showing dramatic recovery in a type-1-diabetes teen and an ALS patient.
Humanoid robotics and vision-only training Robotics ▶ [1:34:00]
Nvidia's Jetson AGX Thor (2 petaflops FP4, 10x Orin, 2M developers) pushes AI compute into robots; Tesla drops motion-capture suits for vision-only Optimus training; China's humanoid robot sales exceed 10,000 units in 2025 (+125% YoY) en route to a projected 10 billion units by 2040.
Waymo vs. Uber and the cost of mobility Transportation ▶ [1:41:20]
One Waymo robotaxi outperforms 99% of Uber drivers on daily trip count despite far smaller fleet scale (700K trips/month vs Uber's 30M/day); the panel speculates driving costs could fall 4x, reshaping suburbs, parking lots, and telecommuting.

Predictions made

open Alex Wissner-Gross: Superintelligence will exist within 2-3 years, so any current college freshman's career plan should assume this timeline.
EP #? · · due: 2027-2028 · ▶ watch
“Assume that we're going to have super intelligence in the next 2 to 3 years and guide your career plans accordingly.”
Your call:
open Peter Diamandis: Google will release Gemini 3 imminently.
EP #? · · due: late 2025 · ▶ watch
“We're on the verge of the release of Gemini 3, which will be coming out any day now.”
Your call:
open Dave Blundin: Duolingo is doomed unless it transforms itself into an AI-native company.
EP #? · · due: unspecified · ▶ watch
“Duolingo is absolutely doomed unless it becomes an AI company.”
Your call:
open Alex Wissner-Gross: Voice-AI customer service company Vocara will 10x its annual recurring revenue by year-end after already doubling in the past week.
EP #? · · due: end of 2025 · ▶ watch
“Vocara doubled in ARR during the past week and is planning to 10x between here and the end of the year.”
Your call:
open Peter Diamandis: Humanoid robots will reach 10 billion units in use.
EP #? · · due: 2040 · ▶ watch
“The prediction of 10 billion humanoid robots by 2040. It's coming.”
Your call:
open Alex Wissner-Gross: Inference-time robotics will let sovereign countries redomesticate their entire manufacturing supply chains, onshoring production once dependent on global trade.
EP #? · · due: unspecified · ▶ watch
“One can imagine a not too distant future where robotics enables essentially every sovereign country to redomesticate its entire supply chain.”
Your call:
open Dave Blundin: Autonomous ride-hailing will make driving roughly four times cheaper than owning a car, letting the poorest people be chauffeured first.
EP #? · · due: unspecified · ▶ watch
“Driving will be four times cheaper than owning a car.”
Your call:
open Dave Blundin: Chip and algorithmic efficiency gains will eventually close most of the 100,000-1,000,000x energy gap between frontier AI models and the human brain, multiplying effective AI capability at fixed power.
EP #? · · due: unspecified · ▶ watch
“If we get a 10 to the 5th, 10 to the 6th improvement in energy efficiency, that means we get that level of improvement on our total AI capabilities.”
Your call:

Numbers that matter

Worth digging into

🕳️ Leopold Aschenbrenner's 13F filing
The panel flagged his fund's largest position (Intel) plus a broader list of energy/compute-adjacent holdings as an investable map of the AI-energy thesis, but didn't dig into the actual filing.
🕳️ Bitter lesson hitting the software/algorithm layer
Alex argues brute-force scaling has hit hardware but not yet software; if it does, it could restructure how AI labs allocate R&D spend.
🕳️ Cryptographic provenance for cameras vs. deepfakes
Alex floats camera-to-browser cryptographic chains of trust as a fix for eroding visual trust, but is skeptical of adoption -- a real open infrastructure question given industrial-scale deepfake capability.
🕳️ Non-CNS psilocin analog for longevity
The mouse study's telomere/anti-aging effects seem separable from psilocybin's hallucinogenic CNS effects, per Alex -- a potentially major longevity drug lead if isolated.
🕳️ AI value-capture: infrastructure vs. application layer
Alex's 'pyramid vs. inverted pyramid' framing is a live, high-stakes debate for where investment dollars should flow in the AI stack.