2025-12-09

China's Rise, GPT-5.2, Anthropic IPO & the Battle for AI Trust w/ Emad, Salim, Dave & AWG | EP #214

The Moonshots panel (fresh off NeurIPS 2025) opens with the finding that Chinese labs now dominate frontier AI publishing while US labs go dark, then works through a stack of fast-moving AI news: Google's Titans/Miras long-term-memory architecture, a leaked GPT-5.2 benchmark chart and OpenAI's 'code red,' Anthropic's ~$300B IPO track amid an HBM memory and copper shortage, OpenAI's hallucination-reducing 'confessions' method, and an MIT study showing algorithmic efficiency gains accrue mostly to large labs. The back half pivots to geopolitics and hard infrastructure: China's Cambricon/Moore Threads chip push to escape Nvidia dependence, Europe's lagging AI gigafactory bid, Michael Dell's Invest America child savings program, the AI-major college boom, and a wave of space news (four new commercial space stations, a possible SpaceX and Anthropic IPO, Sam Altman entering the rocket business, and Chinese orbital AI data centers). It closes on humanoid robotics, capped by a viral Chinese T800 robot kickboxing demo that sparks debate over robot safety, wheels-vs-legs design, and the coming regulatory questions around robots in public.

▶ Watch on YouTube

Topics

NeurIPS 2025: Chinese labs rising as US frontier labs go dark AI ▶ 3:14

Alex Wissner-Gross recaps NeurIPS 2025 (29,000+ attendees, Alibaba's 146 accepted papers) and argues American frontier labs have largely stopped publishing internal research while Chinese labs keep publishing and releasing open-weight models, shifting where cutting-edge AI research becomes visible.

Google's Titans/Miras: breaking the context-window ceiling AI ▶ 10:07

A biologically-inspired 'surprise'-based long-term memory architecture (Titans/Miras) lets models store and retrieve information without the quadratic cost blowup of vanilla transformers, potentially enabling near-infinite context windows.

GPT-5.2 leak, OpenAI's 'code red,' and the weekly leapfrogging rat race AI ▶ 15:50

A leaked (unverified) GPT-5.2 vs Gemini 3 Pro benchmark chart, including a speculative 67.4% Humanity's Last Exam score, is discussed alongside Sam Altman's internal 'code red' memo as a management/fundraising strategy amid near-weekly frontier-model leapfrogging.

Anthropic's ~$300B IPO track amid compute, memory, and copper shortages Economy ▶ 23:45

Anthropic is reportedly negotiating a round valuing it near $300B on projected 2026 revenue of $26B; the panel ties this to a broader capital race including a spiking HBM memory market (OpenAI reserving 40% of world supply for Stargate) and a copper shortage from data-center wiring.

OpenAI's 'confessions' method and real-world hallucination rates AI ▶ 29:43

OpenAI's new method trains models to self-report mistakes/hallucinations rather than hide them; the panel cites hallucination-rate studies (15-25% on everyday questions, GPT-5 dropping from ~18% to ~3%) and Peter's own experience of ChatGPT fabricating fake repair shops.

Gemini 3 Deep Think, fleets of agents, and the cost of intelligence AI ▶ 34:44

Gemini 3 Deep Think's parallel-reasoning approach is framed as a template for how AI revenue will scale — not single better models, but millions/billions of agents working in parallel — while Dave Blundin argues the 'cost of intelligence going to zero' narrative undersells surging demand.

MIT study: AI algorithmic efficiency gains accrue to large labs Compute ▶ 41:07

An MIT study finds 91% of algorithmic efficiency gains from 2012-2023 came from just two transitions (LSTM-to-transformer, Kaplan-to-Chinchilla scaling), undercutting the idea that small/open labs benefit disproportionately from algorithmic progress.

Visual chain-of-thought and converging multimodal world models AI ▶ 46:01

New visual chain-of-thought methods that reason with image tokens (not just text) deliver 3-6% reasoning gains; Emad connects this to a broader convergence where text, image, and video models seem to share the same underlying structure (e.g., Flux by Black Forest Labs).

Geopolitical AI race: China's chip independence vs. Europe's gigafactory push Geopolitics ▶ 53:07

China accelerates chip self-sufficiency (Cambricon tripling output, Moore Threads' oversubscribed IPO) after losing Nvidia access, while a national security strategy points to full US-China tech decoupling; meanwhile the EU opens bidding for an AI gigafactory in early 2026 but is seen as far behind Silicon Valley/Cambridge.

Invest America and the AI-era education/jobs shift Economy ▶ 1:09:14

Michael and Susan Dell's $6.25B 'Invest America' gives every US child born after Jan 1, 2025 a $1,000 investment account; discussion links this to 'universal basic equity,' the boom in AI college majors (MIT's course 64), and a parallel construction-trades boom building AI data centers.

Commercial space race: stations, SpaceX/OpenAI rockets, and orbital AI data centers Space ▶ 1:25:41

Four private US space stations are emerging from a NASA seed program, SpaceX is reportedly weighing a 2026 IPO, Sam Altman is in talks with rocket startup Stoke Space, and China's CosMO Space is building a multi-module orbital AI supercomputer, all pointing toward a race to put AI compute (and Dyson swarms) in orbit.

National robotics strategy and humanoid combat demos Robotics ▶ 1:46:57

A prospective 2026 US executive order aims to accelerate robotics as China installs the majority of the world's robots; the episode closes on Optimus vs. Figure running videos and a viral, controversial Chinese T800 humanoid robot kickboxing demo.

Predictions made

open Alex Wissner-Gross: Frontier AI model leapfrogging (GPT, Gemini, Grok, etc. trading the top benchmark spot) will happen on a near-weekly basis.

EP #? · · due: through 2026 · ▶ watch

“we're just going to see leaprogging on a near weekly basis at this point”

Your call:

open Emad Mostaque: Whether or not the leaked GPT-5.2 benchmark numbers are real, some model will hit those same numbers within roughly six months to a year.

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

“all these numbers will be hit in the next 6 months probably for a year”

Your call:

open Alex Wissner-Gross: China will undergo a 'Cambrian explosion' of new AI architectures now that it has been effectively decoupled from the US tech stack.

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

“I think we're going to see a Cambrian explosion, no pun intended, of architectures coming out of China”

Your call:

open Emad Mostaque: Chinese chipmakers like Cambricon will become fully competitive with Nvidia within a few years, mirroring how BYD came to rival Tesla.

EP #? · · due: roughly 2028-2029 · ▶ watch

“in a few years, you can probably expect them just like Tesla and BYD to actually be fully competitive”

Your call:

open Alex Wissner-Gross: Humanoid robots will begin substituting for skilled-trade/construction labor within 5 to 10 years.

EP #? · · due: 2031-2036 · ▶ watch

“5 to 10 years out... we're going to see humanoid robot substitution effects”

Your call:

open Peter Diamandis: The US Postal Service has at most five years left before it is effectively shut down or replaced, with Amazon likely picking up a government delivery contract.

EP #? · · due: 2031 · ▶ watch

“my guess is US Post Office has at max five years left”

Your call:

open Salim Ismail: Within a year to 18 months, wearable/implantable sensors will give real-time metabolic feedback warning people against eating something mid-digestion.

EP #? · · due: early-to-mid 2027 · ▶ watch

“I'm expecting in a year or 18 months, some sensor that's in your stomach saying, 'Hey, you're about to eat that donut. Wait 10 minutes because I'm still metabolizing the coffee.'”

Your call:

open Alex Wissner-Gross: If recursive self-improvement lets AI substitute for AI-engineering labor, the current rush into AI college majors could reverse, sending students back toward the humanities.

EP #? · · due: unspecified, contingent on recursive self-improvement · ▶ watch

“this rush to major in AI at MIT and UCSD maybe reverse itself, unwind itself and everyone goes back to majoring in the humanities”

Your call:

open Unknown guest: Orbital compute energy will become cheaper than terrestrial energy.

EP #? · · due: 2030 · ▶ watch

“Orbital compute energy will be cheaper than on Earth by 2030”

Your call:

Numbers that matter

29,000+ NeurIPS 2025 registrants, ~50% YoY increase Signals explosive growth of the field's largest annual conference
146 papers accepted from Alibaba, including a best paper award Evidence of Chinese labs' rising research output at NeurIPS
ICLR Chinese first-author papers: 9% (2021) to 30% (2025); US share: 52% to 36% Quantifies the shift in AI research leadership from US to China
GPT-4/4o context window: ~128,000 tokens Baseline for comparing new long-context architectures like Titans/Miras
2 million tokens is roughly equal to 3,000 pages of text or 16 novels Illustrates the scale of proposed long-context models
Human genome = 3.2 billion base pairs; a 2M-token context window covers only ~0.06% of it Shows how far current context windows are from holding a full genome
Leaked Humanity's Last Exam scores: Gemini 3 Pro 37.5% without tools, ~50% with tools; speculative GPT-5.2 score of 67.4% Unverified benchmark leak driving speculation and Polymarket betting
Anthropic reportedly in talks for a round valuing it near $300 billion on ~$26 billion projected 2026 revenue About 10x forward revenue, called cheap versus Palantir's 111x
Opus 4.5 priced at $25 per million tokens vs. Grok 4.1 at $50 per million tokens Illustrates intensifying price competition among frontier model providers
OpenAI reserved 40% of the world's HBM memory supply for its Stargate data center in Abilene, Texas Cited as the first sign of an emerging global compute/memory shortage
Studies found 15-25% hallucination/wrong-answer rates on everyday questions (GPT-4o, Claude 3.7 Sonnet) Baseline hallucination rates motivating OpenAI's 'confessions' method
GPT-5 cut hallucination rate from ~18% to ~3% Shows rapid improvement in factual reliability across model generations
Large-model training became ~22,000x more efficient vs. only 10x-100x for smaller models MIT finding used to argue efficiency gains favor big labs over small ones
Visual chain-of-thought delivers 3-6% gains in reasoning performance Quantifies the benefit of reasoning with image tokens, not just text
Luma raised $900 million to build multimodal 'world models' Investment scale behind the push toward unified text/image/video/world models

Worth digging into

🕳️ NeurIPS/ICLR authorship-gap data (Chinese vs. US first authors)

A structural shift in where cutting-edge AI research becomes public, with strategic implications for who benefits from open science.

🕳️ The MIT algorithmic-efficiency paper (91% of gains from 2 transitions)

Directly undercuts the popular narrative that small/open-source labs can out-innovate big labs on efficiency.

🕳️ OpenAI's 'confessions' method for model honesty

A concrete alignment technique tied to Goodhart's Law that could change how hallucination and self-reporting get engineered into frontier models.

🕳️ China's CosMO Space orbital AI data center

A near-science-fiction engineering claim (100MW, 10 exaflops in orbit) mentioned with no independent verification.

🕳️ Sam Altman's reported talks with Stoke Space

Ties OpenAI's compute ambitions directly to space-launch capability -- a new vertical-integration angle for AI hyperscalers, using an untested ring-shaped aerospike engine.

🕳️ Invest America's $1,000 child accounts as 'universal basic equity'

A live policy/philanthropy experiment framed as a hedge against AI-driven labor disruption.