2026-03-11

Meta Buys Moltbook, GPT 5.4, and Fruitfly Brain Upload | Moonshots Live at The Abundance Summit 238

A live Moonshots show from the Abundance Summit in Palos Verdes with all four co-hosts on stage together for the first time. Peter announces the $3.5M Future Vision XPRIZE and the Sept 25 Moonshots Gathering in LA. The panel recaps the Summit (Eric Schmidt, Kevin Weil, Dara Khosrowshahi, robot CEOs), then covers GPT-5.4 breaking FrontierMath Tier 4, Claude's consumer surge, Meta buying Moltbook, Karpathy's autoresearch, Apple's idle neural cores, and Alex's Eon Systems uploading a fruit fly brain. An AMA closes with the post-capitalism, job-loss versus labor-scarcity debate, Emad's optimistic 'organizational singularity' view, and a pitch for 10 MW regional data centers.

▶ Watch on YouTube

Topics

Future Vision XPRIZE and Moonshots Gathering Other ▶ 6:19

Peter launches a $3.5M global competition for hopeful films about the future, partnered with Google, Range Media and the Roddenberry family. Finalists are shown at the Moonshots Gathering on Sept 25 in downtown LA. moonshots.com has been acquired as the hub URL.

Summit recap: Eric Schmidt and recursive self-improvement AI ▶ 14:47

Schmidt expects about five frontier labs (never more than ten) and thousands of AI startups, with incumbents in trouble. Alex argues recursive self-improvement is already here because current frontier models were largely designed and trained by their predecessors.

Solve everything: OpenAI science push AI ▶ 17:35

Kevin Weil wants 100 scientists winning 100 Nobels. The panel notes labs are hiring top mathematicians, physicists and biologists because everything becomes a software problem, and doubts the frontier models will really be released (the IMO gold model still is not).

Robots and Uber's automation timeline Robotics ▶ 20:27

Brett Adcock promised Peter a Figure robot this summer. Dara Khosrowshahi said Uber would automate about 30% of its work this year, which raises the driver-displacement question nobody can candidly answer on a public stage.

GPT-5.4, FrontierMath and Cerebras speed AI ▶ 22:44

GPT-5.4 at max reasoning solves about 38% of FrontierMath Tier 4, with rumors it is near the first open math problem. Emad points to OSWorld-Verified passing human level and OpenAI's Cerebras deal moving from about 50 to roughly 1,000 tokens per second. Alex says the data ceiling was illusory and synthetic data is now the path.

Claude consumer surge and the Department of War fight AI ▶ 27:51

Claude surged in consumer use after Anthropic's public clash with the Department of War. Alex calls it the Streisand effect and argues attempts to pause frontier AI end up accelerating competitors. Only about 11M users out of 8B people, so adoption is extremely early.

AI job-exposure chart and managing with AI Economy ▶ 29:55

A radar chart shows AI saturating 80-85% of white-collar task categories, with health care support, food service and grounds maintenance lowest. Dave uses Gemini and Claude to track 1,100 people and review deal memos. Salim notes futurists wrongly expected manual labor to be automated first.

Meta acquires Moltbook; building for agents AI ▶ 36:46

Meta acqui-hired the Moltbook team while OpenAI got OpenClaw. The panel debates whether advertising to AI agents makes sense given Moltbook agents mimic human distrust and compute-scarcity dynamics. Dave urges kids to jump in now because the field has no age or experience requirement.

Yann LeCun's $1B and the world-model debate AI ▶ 45:58

LeCun's AMI Labs raised $1B for world models, the second-largest European round. Alex says generative models already got us to scalable superintelligence; Emad says JEPA does not scale well yet; Dave warns against scientific arrogance.

Karpathy autoresearch and tiny models AI ▶ 50:38

Karpathy's autoresearch ran about 650 experiments in two days and found transferable gains. Alex expects the next algorithmic breakthroughs at the small-model end, possibly knowledge factored out into plain-text databases and model cores shrinking to megabytes. Small-to-large scaling lag fell from months to days.

Apple M5 and the idle neural-core overhang Compute ▶ 56:59

Apple controls about 20% of TSMC output but locks its neural cores and unified memory away from developers. The panel expects the overhang to collapse within a year via built-in OS models or local open-weight apps such as Qwen 27B on a 24 GB Mac.

World eye-scanning stores Other ▶ 1:00:44

Sam Altman's iris-scan verification coming to retail stores prompts Minority Report comparisons; the panel notes sci-fi timelines are arriving decades early.

Eon Systems fruit fly brain upload Biotech ▶ 1:02:46

Alex announces the first multi-behavior whole-brain emulation of a fruit fly: a connectome driving a simulated body that walks, grooms and eats. Eon is targeting mice and then humans, which Alex says is years, not decades, away.

xAI power and the energy build-out Energy ▶ 1:06:07

xAI is committing to 1.2 GW per data center. The US is on track to add 86 GW of capacity in 2026, 51% of it solar, and Schmidt's 100 GW call is being met by deregulated, well-funded companies.

Predictions made

open Eric Schmidt (clip): No more than about 10 foundation-model labs will exist (about 4-5 in the US, 1-2 in Europe, a couple in China), alongside thousands of successful AI startups.

· due: unspecified · ▶ watch

“there won't be more than 10, but there will be thousands of successful AI startups”

Your call:

open Alex Wissner-Gross: GPT-5.4 will solve the first open (unsolved) hard math problem on the open-problems benchmark tier.

· due: unspecified · ▶ watch

“5.4 is reportedly rumored to be on the verge of solving the first open hard math problem”

Your call:

open Peter Diamandis: Neural nets will be 100 times bigger at the end of 2026 than at the start of the year.

· due: 2026 · ▶ watch

“the neural nets will be 100 times bigger at the end of this year than at the beginning”

Your call:

open Alex Wissner-Gross: By September 2026 there will be at least a thousand ultra-high-quality inspirational AI-generated videos produced for nearly free.

· due: 2026-09 · ▶ watch

“a thousand videos of ultra-high inspirational quality generated for nearly free”

Your call:

open Peter Diamandis: Figure's Brett Adcock will ship Peter one of the next-generation humanoid robots this summer.

· due: 2026 summer · ▶ watch

“this summer, he will ship me one of those”

Your call:

open Dara Khosrowshahi (clip): Uber will automate roughly 30% of its employment this year.

· due: 2026 · ▶ watch

“we're going to automate 30% or so of our employment this year”

Your call:

open Alex Wissner-Gross: A radical post-transformer advance will shrink frontier-capable models far further, with the model core eventually in the megabytes and world knowledge factored out into a plain-text database.

· due: unspecified · ▶ watch

“We'll be something maybe even in the megabytes”

Your call:

open Alex Wissner-Gross: The idle-neural-core overhang on Apple Silicon will collapse within a year, e.g. Apple builds frontier-class local models into the OS, possibly announced at WWDC in June.

· due: 2027 · ▶ watch

“I would be surprised if that overhang doesn't collapse in the next year”

Your call:

open Alex Wissner-Gross: Eon Systems will achieve the first whole-brain emulations of a mouse and of a human within years, not decades.

· due: unspecified · ▶ watch

“it's going to be years, not decades, before we get to the first mouse and the first human”

Your call:

open Peter Diamandis: Archer Aviation will be operational moving people around Los Angeles with eVTOLs by 2028 for the LA Olympics.

· due: 2028 · ▶ watch

“Archer plans to become operational by 2028 here”

Your call:

open Dave Blundin: If Anthropic keeps growing 10x per year from a roughly $26B run rate it reaches about $2.6 trillion revenue in two more years, implying a quadrillion-dollar valuation.

· due: 2028 · ▶ watch

“26 billion, 260 billion, 2.6 trillion, most revenue in the history of the world”

Your call:

open Elon Musk (clip): Hundred-trillion-dollar companies will exist within about five years.

· due: 2031 · ▶ watch

“we're going to have 100 trillion dollar companies”

Your call:

open Peter Diamandis: A newly launched XPRIZE typically gets won within 6-7 years, so a prize for cutting living costs from about $2,500 to $250 a month would be won in that window.

· due: 2032 · ▶ watch

“it typically gets won within 6 to 7 years”

Your call:

open Dave Blundin: AI will cause a massive job trough and social unrest followed by a rebound in 2028, an Industrial Revolution compressed into two to four years.

· due: 2028 · ▶ watch

“a massive trough, massive social unrest, and then a rebound in 2028”

Your call:

open Emad Mostaque: Automating a typical company with AI leaves about 25% of the staff but spawns about five times as many companies, so overall employment stays flat with no radical unemployment.

· due: unspecified · ▶ watch

“you end up five times more companies because you can”

Your call:

open Dave Blundin: AI will push company valuations through the roof while the W-2 paycheck disappears for non-shareholder workers.

· due: unspecified · ▶ watch

“their W-2 paycheck is dead. It's going away”

Your call:

Numbers that matter

$3.5M raised for the Future Vision XPRIZE $3M prize money plus film financing, with Google, Range Media and the XPRIZE Foundation; finale Sept 25 in downtown LA.
About 500 people put down $100 deposits for the Moonshots Gathering Early-bird deposits; event rented at the United Theater, LA.
10,000+ entries narrowing to 100, 50, 10, then 5 finalists Planned Future Vision XPRIZE funnel.
GPT-5.4 solves about 38% of FrontierMath Tier 4 Research-level problems that take teams of mathematicians weeks.
About 50 to 1,000 tokens per second Speed jump after OpenAI's Cerebras deal (Codex 5.3 fast).
11 million Claude users Out of 8 billion people, used to argue adoption is very early.
80-85% of white-collar task categories saturated Radar chart of AI capability by occupation group; lowest in health-care support, food service and grounds maintenance.
About 1,100 people tracked with AI Dave uses Gemini and Claude 4.6 to review thousands of documents on his portfolio team.
Meta bought Manus for $2 billion Cited by Emad as Meta's agent play alongside the Moltbook acqui-hire.
LeCun's AMI Labs raised $1B Roughly a $2.5B valuation per the panel; second-largest European round.
About 650 experiments in 2 days Karpathy's autoresearch run that found gains transferring from small to large models.
Small-to-large scale-up lag fell from 6-12 months to 6 days Emad on the matured software and hardware stack.
Apple controls about 20% of TSMC manufacturing Used to argue its locked neural cores are a huge waste.
Qwen 27B (27 billion parameters) runs on a 24 GB MacBook Described as about Sonnet level; no App Store apps wrap it yet.
About 50 million neural connections modeled in the fruit fly Peter's 50 trillion was corrected live to 50 million.

Worth digging into

🕳️ Eon Systems fruit fly whole-brain upload

Claims the first multi-behavior whole-brain emulation of a model organism, with mouse and human targets 'years not decades' out.

🕳️ Karpathy's autoresearch and the small-model route to RSI

Automates what AI researchers do all day and may be the most direct accelerant to recursive self-improvement.

🕳️ Is Meta's Moltbook deal an agent-economy play or an acqui-hire?

The panel splits on whether advertising to agents is viable; an attendee says it may simply be buying creative talent.

🕳️ Apple's idle neural cores and local frontier models

Alex predicts the overhang collapses within a year, possibly at WWDC.

🕳️ Organizational singularity: 25% of staff, 5x companies

Emad's quantitative case that AI doesn't cause mass unemployment directly contradicts Dave's trough-then-2028 view.

🕳️ GPT-5.4 FrontierMath Tier 4 and the open-problems tier

38% on research-level math is a bellwether and a first genuine open-problem solve may be imminent.

🕳️ Future Vision XPRIZE

A $3.5M effort to counter dystopian media with positive futures; finalists revealed Sept 25 in LA.