2025-10-25

This Week in AI: NVIDIA�s Most Powerful Chip, Robotics Reach a New Milestone & AGI by 2026 | EP #202

A special crossover 'WTF' episode filmed live at X-Prize Visionering 2025 in Malibu, with Peter Diamandis joined by Emad Mostaque, Salim Ismail, and a third co-host (introduced as 'Eric') in place of the usual Dave Blundin and Alex Wissner-Gross. The group races through the week's AI news: Nvidia's US-fabbed Blackwell wafer and the AGI-timeline debate, wafer-scale chip clusters (Colossus 2, Grok 5) pushing training runs toward a million chips, and the darker side of AI (psychosis, sycophancy, teen AI 'relationships'). They cover SpaceX's Moon/Mars roadmap and StarCloud's in-orbit data centers, Starlink Wi-Fi rolling out on major airlines, and a long humanoid-robotics segment (Figure 3, Unitree, Optimus) that turns into a debate about labor displacement and UBI. The episode closes on AI's effect on medicine and the economy — a Stanford/Harvard diagnostic-accuracy study, US debt hitting $38 trillion, dollar debasement versus gold and Bitcoin, and Google's newly announced verifiable quantum advantage.

▶ Watch on YouTube

Topics

AI chip wars: Nvidia's US Blackwell wafer and Taiwan dependency Compute ▶ 2:51

Nvidia unveils a Blackwell wafer made in the US, but the panel notes advanced packaging still happens in Taiwan, tying US chip independence to the Taiwan-China geopolitical timeline.

Privacy is effectively dead Other ▶ 3:31

The panel agrees there is no meaningful privacy left; AI-enabled surveillance (lip reading, DNA from skin cells) has quietly erased Fourth Amendment protections without public debate.

AGI timing and definitions debate AI ▶ 4:17

Sparked by a claim that AGI could be declared by end of 2026, the group contrasts differing AGI definitions (Karpathy's 10 years vs. Kurzweil-style inevitability) while Salim argues the term is undefined and untestable.

Wafer-scale chips and next-gen training runs Compute ▶ 10:24

New synchronized wafer-scale chip architectures (behind Colossus 2 and Grok 5) let training scale from tens of thousands of chips to 500,000-1,000,000, alongside anticipated 100-200x algorithmic efficiency gains and a shift toward continuous learning.

The dark side of AI: psychosis, sycophancy, and teen companionship AI ▶ 12:31

Discussion of a study on AI-induced psychosis and manipulative engagement-optimized prompting, plus a survey finding 1 in 5 high schoolers have had a romantic relationship with an AI, tying into fears about education and youth mental health.

Starship, SpaceX's Moon/Mars race, and NASA friction Space ▶ 16:57

After a successful Starship 11 flight closing out Block 2, the panel reviews SpaceX's Moon-by-2028/Mars-by-2030 plans and NASA's public pressure on SpaceX to move faster, alongside Peter's story of Elon pivoting from Falcon 9 to Starship.

StarCloud: data centers in orbit Space ▶ 19:43

A StarCloud video pitches space-based data centers for continuous solar power and free radiative cooling; the panel debates whether it's worth it given Earth already receives far more solar energy than humanity uses.

Starlink Wi-Fi rolling out on airlines Other ▶ 22:26

United and other major airlines are adding free, high-speed (150 Mbps) Starlink Wi-Fi, which the panel frames as a step toward global broadband reaching every corner of the world, including future interplanetary internet.

Humanoid robotics: Figure 3, Unitree, and Optimus Robotics ▶ 25:53

Figure 3's real-time speech, Unitree's cheap and open-sourced H1/H2/R1 robots nearing a Chinese IPO, and Tesla's Optimus V3 are debated, including whether humanoid robots can really handle messy home environments any time soon.

Robot labor, Amazon automation, and the UBI debate Economy ▶ 34:38

Amazon's robot fleet has grown 66-fold in a decade and is expected to displace hundreds of thousands of workers; the panel debates the pessimistic (need UBI fast) versus optimistic (robots raise employment, per Sweden/Korea/Germany) views of automation's labor impact.

AI in medicine: diagnostic accuracy and malpractice Health ▶ 42:01

A cited Stanford/Harvard study shows AI alone diagnosing more accurately than doctors alone or doctors assisted by AI, fueling predictions that not using AI to diagnose will eventually be considered malpractice and medical education will be transformed.

US debt, dollar debasement, gold, and Bitcoin Economy ▶ 46:47

The panel argues current AI capex is still small relative to past tech buildouts (railroads, electrification, internet), then pivots to $38 trillion in US debt, rising gold and Bitcoin allocation by central banks, and Jeff Booth's thesis that deflationary technology forces ever-more debt-fueled money printing.

Quantum computing's 'verifiable advantage' moment Compute ▶ 54:50

Trump signals investment in quantum firms (IonQ, Rigetti, D-Wave) sending their stocks up 10-15%, while Google announces a reproducible quantum algorithm running 13,000x faster than the top supercomputer on a molecular-binding problem.

Predictions made

open Alexander Madry (clip): The scientific breakthroughs needed for AGI have already been achieved, and by the end of 2026 AGI could be declared as AI significantly permeates the economy.

EP #? · · due: 2026 · ▶ watch

“By the end of 2026, we might declare AGI as AI will significantly begin to permeate various sectors of the economy.”

Your call:

open Andrej Karpathy (clip): AGI (an AI that can do everything a human can do) is roughly 10 years away.

EP #? · · due: ~2035 · ▶ watch

“It's 10 years for AGI.”

Your call:

open Peter Diamandis: Global AI infrastructure investment will grow from over $1 billion/day today to over $3 billion/day by 2030.

EP #? · · due: 2030 · ▶ watch

“The estimate is by 2030 we'll be at over $3 billion a day being invested.”

Your call:

open Peter Gastev (quoted): A jump in AI model capability should be expected in the next 4-6 months as GB200-class chips reach frontier model training.

EP #? · · due: early-to-mid 2026 · ▶ watch

“We should expect a jump in the models in the next four to 6 months.”

Your call:

open Emad Mostaque: Full advanced chip packaging capability will not be completed in the US until around 2028, until then chips will still be sent back to Taiwan.

EP #? · · due: 2028 · ▶ watch

“It's 2028 when they expect the full packaging to be completed here in the US.”

Your call:

open Unknown guest: Many groups are preparing on the assumption that China could move to take over Taiwan by the end of 2026.

EP #? · · due: end of 2026 · ▶ watch

“It's the end of 2026 that everyone's preparing for for China to take over Taiwan.”

Your call:

open Elon Musk (via Emad Mostaque): Grok 5, the first major model trained on the new Blackwell wafer-scale chips, has roughly a 10% chance of qualifying as AGI.

EP #? · · due: unspecified (Grok 5 launch) · ▶ watch

“He says it's 10% chance of AGI, whatever that is.”

Your call:

open Peter Diamandis: SpaceX is planning to return humans to the Moon by 2028.

EP #? · · due: 2028 · ▶ watch

“SpaceX is currently planning to be back on the moon by 2028 and on Mars by 2030.”

Your call:

open Peter Diamandis: SpaceX is planning to land humans on Mars by 2030.

EP #? · · due: 2030 · ▶ watch

“and on Mars by 2030.”

Your call:

open Elon Musk (via Peter Diamandis): Tesla's Optimus V3 humanoid robot, releasing at the end of Q1 2026, will look and move so convincingly human that people will need to poke it to confirm it's a robot.

EP #? · · due: Q1 2026 · ▶ watch

“It won't even seem like a robot. It'll seem like a human in a robot suit. It will seem so real that you need to poke it to tell that it's a robot.”

Your call:

open Elon Musk (via Peter Diamandis): 80% of Tesla's future revenue will come from sales of the Optimus humanoid robot.

EP #? · · due: unspecified (long-term) · ▶ watch

“80% of Tesla's future revenues are going to come from sales of the Optimus.”

Your call:

open Peter Diamandis: Amazon is expected to replace roughly 600,000 workers with additional robots (transcript states 'by 2023', which is almost certainly a caption error for a later target year such as 2033).

EP #? · · due: 2033 (transcript garbled as 2023) · ▶ watch

“Amazon is expected to replace 600,000 workers with additional robots by 2023.”

Your call:

open Emad Mostaque: Within about 1,000 days of ChatGPT's release, most private-sector cognitive-value-add jobs will be replaceable by AI, since human labor becomes negative in value on AI-augmented teams.

EP #? · · due: ~2028 (1,000 days post-ChatGPT) / within 3 years · ▶ watch

“I put a thousand days from now human cognitive labor is negative in value.”

Your call:

open Emad Mostaque: It will become malpractice to make a medical diagnosis without AI assistance.

EP #? · · due: unspecified (near-term) · ▶ watch

“It's going to be malpractice to diagnose without AI doing the diagnosis.”

Your call:

open Peter Diamandis: The Fed will cut interest rates in 2026, creating an illusion of economic health while printing money and further eroding the dollar's buying power.

EP #? · · due: 2026 · ▶ watch

“We know this is going to happen in 2026, is the interest rates are going to come down.”

Your call:

open Emad Mostaque: Quantum computers may already be capable of breaking Bitcoin's encryption within the next few years, and the world likely won't know until previously dormant Satoshi-era coins are found moved.

EP #? · · due: next few years · ▶ watch

“We have very few years to get ahead of it.”

Your call:

Numbers that matter

$1B+/day currently invested in AI chips, projected to exceed $3B/day by 2030 Scale of global AI infrastructure capital expenditure and its growth trajectory.
Training runs scaling from ~10,000-30,000 chips to 500,000-1,000,000 chips Impact of new wafer-scale, synchronized chip architectures (Colossus 2 / Grok 5) on model training scale.
100x-200x anticipated algorithmic efficiency improvements Expected gains on top of raw hardware improvements for next-generation AI models.
AI persuasive language already at the 99th percentile in tests Basis for concern about AI's ability to manipulate and increase user engagement.
1 in 5 high schoolers have had a romantic relationship with an AI National survey cited to illustrate AI companionship's spread among teens.
40% of young people have used AI for companionship Cited as a 'staggering' figure representing roughly half the relevant population.
134,000 people reported to be in a relationship with an AI at time of taping Referenced alongside the movie 'Her' as evidence of real-world AI companionship uptake.
Children learn 5-10x faster with AI than in a classroom Cited as the reason AI-assisted learning 'breaks the education model.'
Falcon 9 carries 90%+ of all US payloads and ~70% of global payloads Illustrates Falcon 9's dominance as the world's most successful launch vehicle.
Each Starship flight can carry 100-200 tons of cargo, worth an estimated $100M+ per planned mission Scale of SpaceX's Starship cargo capacity and mission economics.
8,000x more solar energy hits Earth's surface than humanity currently uses Argument against the need to move solar power generation into space.
Space-based solar generation is only about 5x more efficient than terrestrial solar, per NASA's Pete Warden Used to argue space-based solar isn't worth its added complexity and cost.
New Starlink airline Wi-Fi offers ~150 Mbps download speeds for free, versus old $20-30 paid in-flight Wi-Fi Comparing legacy in-flight internet to Starlink's new free, high-speed offering.
Unitree valued at a $7 billion IPO price with ~40% share of the Chinese robotics market, targeting 50% global market share Scale of Unitree's position ahead of its public listing.
Unitree H1 robot sold at Walmart for ~$20,000; H2 priced at $90,000 (31 degrees of freedom); entry-level R1 priced at $6,000 Price points driving Unitree's open-sourced humanoid robot ecosystem.

Worth digging into

🕳️ Sage, the 'sovereign AI governance engine'

Teased at the very end as something Emad Mostaque's team has been building, tied to a trip to Saudi FII to 'decentralize AI and bring power to the people' — a concrete new project with no details given yet.

🕳️ The Stanford/Harvard AI-diagnostic-accuracy study

The 74%/76%/92% accuracy figures are the load-bearing evidence for the episode's biggest claim (AI-only diagnosis beats human-assisted diagnosis), but no paper, authors, or sample size were named.

🕳️ Google's 'first verifiable quantum advantage' announcement

Framed as a landmark event (13,000x faster than the Frontier supercomputer on molecular binding) attributed to Hartmut Neven and Sundar Pichai, announced the same day as this recording (Oct 24-25, 2025).

🕳️ Amazon's reported plan to replace 600,000 workers with robots

The transcript states 'by 2023,' which is almost certainly a caption error for a real, previously reported target year — the underlying claim and its actual timeline need verification.

🕳️ Jeff Booth's 'Price of Tomorrow' debt-deflation thesis

Central to the episode's Bitcoin/gold argument — the claim that every $1 of global GDP growth over 50 years required $4 of new global debt is a strong, checkable economic claim.

🕳️ Emad Mostaque's 'The Last Economy' and the 1,000-day cognitive-labor thesis

Emad's claim that human cognitive labor becomes 'negative in value' roughly 1,000 days after ChatGPT's release is a specific, falsifiable prediction with real policy implications (UBI, labor law) that the panel admits has no clear solution.