2026-08-13

Bernie Demands the Labs Stop, Wall Street Turns GPUs Into Bonds, Grok 4.7 Takes #1 ft. Emad Mostaque

The Moonshots panel, joined again by Emad Mostaque, covers nine stories spanning the $101M Health Span X Prize finals, the collapse of Hollywood film economics under AI video generation (Higgsfield's $2M feature film, LTX 2.5, Chinese dominance of text-to-video), and Grok 4.6's frontier-tying performance built on Cursor's reasoning-trace data. Nvidia's $500B Wall Street financing partnership for GPU compute sparks a debate over whether compute-backed securities repeat the mortgage-backed-securities playbook, followed by a skeptical look at the "Dragon Hatchling" post-transformer architecture. Bernie Sanders' AI-pause letter to Anthropic, OpenAI, and Meta and Stanford's AI-designed bacteriophage draw a sharp defense of accelerationism from Alex Wissner-Gross, while Anthropic's text watermarking, the EU's AI-labeling rules, and Zuckerberg's "personal superintelligence" essay get mixed, often cynical reviews. The episode closes on Archer Aviation's acquisition of three Boeing units consolidating the eVTOL race, and a fast-fire AMA on alignment, escape velocity, simulation ethics, and abundance.

▶ Watch on YouTube

Topics

Health Span X Prize finals Longevity ▶ [3:08]

Peter reports from the $101M Health Span X Prize finals in Utah: 800 teams entered, $1M awarded to 10 teams so far ($20M given away toward the $80M grand prize), all requiring human trials to reverse functional aging in cognition, muscle, and immune system by 2030.

AI-generated feature films (Higgsfield, LTX 2.5) AI ▶ [19:46]

Higgsfield's "Cully Hill Boys" is the first full-length AI-generated movie with licensed celebrity likenesses, made for $2M/4 weeks/$1M compute versus Hollywood's $20-100M/12-18 months, with the entire 10-step workflow open-sourced. LTX 2.5 runs high-definition real-time video generation on a MacBook Pro.

Chinese dominance of text-to-video AI AI ▶ [21:51]

Bloomberg reports 9 of the top 10 text-to-video models are Chinese, raising questions about US cultural export dominance and noting these models are simultaneously learning physics/motion/causality needed for robotics and autonomous driving.

Grok 4.6 launch and xAI's Cursor strategy AI ▶ [40:26]

xAI's Grok 4.6 ties GPT-5.6 on the Artificial Analysis Intelligence Index at $2-6/M tokens, built on post-training with Cursor's acquired reasoning-trace data. The panel debates whether this lets xAI catch up to but not leapfrog the frontier, with Grok 4.7 (2T params) rumored in two weeks and eventually a 10T-parameter model.

Nvidia's $500B GPU financing partnership Economy ▶ [51:58]

Nvidia partners with Apollo, BlackRock, Blackstone, Brookfield, and KKR to mobilize over $500B in third-party capital for AI infrastructure, turning GPU compute into a financial asset class. Debate follows on whether this echoes mortgage-backed securities risk.

Dragon Hatchling post-transformer architecture AI ▶ [1:04:50]

A new non-transformer architecture (Dragon Hatchling / BDH) climbs the Arc AGI benchmark at much lower compute cost. Alex Wissner-Gross reads the paper and calls it a "hot mess" of biomimetic motifs rather than a simple, bitter-lesson-compliant successor to transformers.

Bernie Sanders' AI pause letter Geopolitics ▶ [1:12:01]

Senator Sanders sends a letter to the CEOs of Anthropic, OpenAI, and Meta demanding a pause on AI development, citing AI-designed viruses and warning "my colleagues and I in the US Senate will" act if labs don't. The panel largely dismisses it as political positioning.

Stanford's AI-designed bacteriophage (Evo 2) Biotech ▶ [1:20:51]

Stanford researchers used the open-source Evo 2 model to design ~300 novel bacteriophage DNA sequences, synthesizing 16 viable phages effective against E. coli with no prior natural resistance -- called biology's "Wright brothers moment" and a dual-use biosecurity concern.

AI content watermarking and EU labeling rules AI ▶ [1:33:41]

Anthropic begins embedding invisible statistical watermarks in Claude-generated text; the EU launches an AI-content icon/labeling system. Watermark-removal tools appeared within 24 hours, and the panel (especially Alex) calls both moves regressive, akin to cookie banners.

Zuckerberg's "personal superintelligence" essay AI ▶ [1:41:50]

Mark Zuckerberg publishes a 6,500-word vision statement proposing personal AI agents for every person (Muse Glimmer/Muse Spark models) plus community benefit-sharing for data center host towns. The panel is split between praising open-weight commitment and cynicism about Meta's incentives.

eVTOL consolidation: Archer acquires Boeing units Transportation ▶ [1:56:37]

Archer Aviation acquires Wisk Aero, Insitu, and SkyGrid AI from Boeing (Boeing takes an equity stake) as the panel surveys the flying-car race: Joby (FAA stage 4, $3/seat-mile), Archer, EHang (fully autonomous, operating in China/Dubai at $330K/unit), Beta Technologies, and Eve.

AMA: alignment, escape velocity, simulation ethics, abundance Other ▶ [2:07:00]

Rapid-fire audience Q&A covering whether smarter models become more or less aligned (Emad: models are still "bacteria level" internally), whether any lab can reach permanent model escape velocity (Salim: no, models commoditize), Terafab obsolescence risk, and whether the singularity has a resource cost (Alex: scarcity premise is outdated).

Predictions made

open Peter Diamandis: The Health Span X Prize competition (reversing functional aging by 20 years) will be won.

EP #? · · due: 2030 · ▶ watch

“So, this is a competition that will be won by 2030.”

Your call:

open Dave Blundin: Dave will personally do more productive work in the second half of 2026 than in his entire life combined up to 2026.

EP #? · · due: end of 2026 · ▶ watch

“There's no doubt in my mind that I'm going to do more productive work in the second half of 2026 than in my entire life combined up till 2026.”

Your call:

open Emad Mostaque: The compute cost to shoot a feature-length AI-generated movie will drop from roughly $1M to around $10,000.

EP #? · · due: January 2027 ("the new year") · ▶ watch

“You could shoot a movie just like that one probably for 100,000 of compute and then 10,000 of compute... Probably by the new year.”

Your call:

open Alex Wissner-Gross: Grok 4.7 will surpass Anthropic's Opus and take the #1 spot on frontier model rankings.

EP #? · · due: late August 2026 ("a couple of weeks" from Aug 13, 2026) · ▶ watch

“He thinks 4.7 will go above Opus. So, it'll take number one in a couple of weeks.”

Your call:

open Alex Wissner-Gross: Bad actors will weaponize AI content watermarks (e.g., Anthropic's) for malicious purposes.

EP #? · · due: within weeks · ▶ watch

“I I think we're only 5 minutes away from bad actors weaponizing these watermarks to do bad things.”

Your call:

open Alex Wissner-Gross: Elon Musk will announce a flying car.

EP #? · · due: 6-12 months from August 2026 (roughly Feb-Aug 2027) · ▶ watch

“Elon's going to announce his flying car within 6 to 12 months.”

Your call:

Numbers that matter

$101 million Total purse of the Health Span (longevity) X Prize, with an $80 million grand prize.
800 teams Number of teams that entered the Health Span X Prize.
$20 trillion/year Global cost of age-related chronic disease, against a $120-130 trillion global economy.
47 -> 79 US average life expectancy in 1900 versus today; healthy life expectancy stops around age 63.
~70% LEV Alex Wissner-Gross's back-of-envelope estimate of longevity escape velocity achievable via third-gen GLP-1 therapy in a subpopulation.
$2M / 4 weeks / $1M compute Total budget, production time, and compute cost for Higgsfield's AI-generated feature film "Cully Hill Boys," vs $20-100M and 12-18 months traditionally.
$700M revenue run rate Higgsfield's revenue run rate reached in about 1.5 years since founding.
61 / $2-6 per million tokens Grok 4.6's score on the Artificial Analysis Intelligence Index (tying GPT-5.6) and its input/output token pricing.
1.5T -> 2T -> 10T parameters Reported parameter-count progression from Grok 4.5 to 4.7 to future Grok models.
$500 billion+ Third-party capital Nvidia is mobilizing with Apollo, BlackRock, Blackstone, Brookfield, and KKR for AI infrastructure financing.
contracts through 2029 CoreWeave clients signing multi-year compute contracts for 6-year-old A100 GPUs (introduced 2020).
83.3 vs 83.2 on CyberGym DeepSeek V4 Pro's cyberattack-capability benchmark score narrowly exceeding Anthropic's Methuselah.
300 designs / 16 viable phages Number of bacteriophage DNA sequences Evo 2 generated and how many proved viable against E. coli in Stanford's experiment.
$3/seat-mile vs $330,000/unit Joby's target eVTOL price point per seat-mile versus EHang's autonomous two-seat aircraft purchase price.
71% of Americans Share of Americans who don't want a data center in their backyard, more than oppose a nuclear plant nearby.

Worth digging into

🕳️ Dragon Hatchling post-transformer architecture

Alex calls it a hot mess of biomimetic motifs, but the underlying question of what a true post-transformer architecture looks like remains open and consequential for compute economics.

🕳️ Evo 2 open-source bio-design model

A 40B-parameter open-source model that can already design viable novel bacteriophages runs on a consumer MacBook, well ahead of any biosecurity regulation of DNA/RNA synthesis equipment.

🕳️ Nvidia's $500B compute-backed securities structure

Could become the largest new asset class of the decade or repeat the mortgage-backed-securities collapse if a new architecture strands GPU-based compute assets.

🕳️ AI content watermarking arms race

A watermark-removal tool for Claude, Gemini, and OpenAI outputs appeared within 24 hours of Anthropic's launch, suggesting the whole approach may be short-lived.

🕳️ eVTOL industry consolidation and capital flight to AI/robotics

Brett Adcock's move from Kitty Hawk/Archer-adjacent ventures to Figure AI is cited as a proxy for venture capital abandoning flying cars for AI and robotics.

🕳️ Higgsfield's fully open-sourced 10-step AI filmmaking workflow

If genuinely replicable, it could collapse Hollywood's cost structure to roughly 2% of a traditional feature film budget and 6% of the time.