2026-08-27

Sam Altman: Singularity Slow-Down, Emad Runs 18 Grokbots, Waymo Slashes Hardware 83% | EP #283

The Moonshots panel dissects Sam Altman's admission that AI's societal disruption is arriving slower than he expected, debating whether institutional inertia or 'abstraction layer' stickiness is really to blame. They cover the explosion of agent swarms (Emad running 18 Grokbots, Dave running 100K-agent swarms), Chinese open models squeezing US labs on cost and even causing Anthropic's flagship model to plateau, and Nvidia's $6B bet on Poolside for a US openweight alternative. The show closes with a run of robotics and space stories: Waymo slashing autonomy hardware costs via Chinese OEM hardware, a Chinese autonomous rescue drone, and Elon Musk's plan for 10,000 Starship launches a year by 2030 to build orbital data centers. A recurring theme is that AI abundance is creating more work (not less) for humans, and that vertical integration is becoming the default strategy for every major AI player.

▶ Watch on YouTube

Topics

Sam Altman's 'slower, not a step function' reversal AI ▶ 0:08:37

Altman says economic inertia and institutional lag are slowing AI's societal impact into a 'rising tide' rather than sudden disruption; the panel debates whether societal inertia or the economy's abstraction-layer stack is the real bottleneck.

Grokbot and the rise of agent swarms AI ▶ 0:23:02

xAI's Grokbot gives each user a dedicated cloud agent with sub-agent specialists; panelists are already running 18 to tens-of-thousands of agents in parallel and debate whether its chat-app interface is durable or a temporary 'vaudeville era' UI.

Gemini 3.7 Flash tops the analyst-agent benchmark AI ▶ 0:31:38

Google's Gemini 3.7 Flash took the top spot on the Artificial Analysis analyst-agent benchmark by optimizing for reliability and speed rather than raw capability; the panel argues over whether Google is compute-starved internally or suffering talent attrition.

Nvidia bankrolls Poolside for a US openweight model Compute ▶ 0:45:44

Nvidia is pouring $6B into Poolside (an ex-GitHub-CTO startup) to build a leading US openweight model and inference stack, structured as a licensing deal to avoid antitrust review, as frontier companies race to vertically integrate up and down the stack.

Chinese open models squeeze the US closed labs AI ▶ 0:54:57

Moonshot AI's Kimi Linear architecture cuts context memory 75% with 6x faster decoding; Anthropic's flagship model has reportedly plateaued because cheap Chinese open models deliver ~80% of the capability at a fraction of the cost, prompting Anthropic to reverse its data-retention policy ahead of an expected IPO.

Ditto deletes the dating interface AI ▶ 1:10:18

Berkeley startup Ditto removes swiping and choice entirely, algorithmically matching users to one weekly date; 160,000 college signups have produced 80,000 dates, and the panel sees the 'delete the interface, not just optimize it' pattern as generalizable far beyond dating.

AI is making humans work harder, not less Economy ▶ 1:20:02

A Wall Street Journal piece describes agents producing so much output that human review and judgment become the bottleneck; panelists describe personally supervising fleets of agents around the clock and expect this phase to last only one to three years.

Data center backlash and the evidence-vs-narrative problem Energy ▶ 1:30:08

Polling shows US opposition to nearby data centers rose from 43% to 75% in a year; the panel blames a narrative-over-evidence media culture (possibly amplified by foreign interference) compared to China's ability to centrally site infrastructure.

Waymo slashes robotaxi hardware costs Robotics ▶ 1:38:39

Waymo's new 5nm chip and Zeekr-designed 'Ohigh' minivan cut autonomy hardware cost from $115K to $20K and vehicle cost from $200K to $75K by using Chinese OEM hardware, closing much of the gap with Tesla's projected $30K Cybercab.

Humanoid robots as universal drivers Robotics ▶ 1:45:11

The panel floats a cheap (~$6,000 bill-of-materials) humanoid robot that simply climbs into and drives existing vehicles instead of retrofitting cars for autonomy, echoing Palmer Luckey's military use case of humanoids operating existing human-designed controls.

China's rescue drone and the robot-backlash asymmetry Robotics ▶ 1:47:52

A viral video shows a Chinese autonomous drone/life-preserver hybrid rescuing swimmers at 30 mph; the panel contrasts China's fast public embrace of visible robot use-cases with expected US backlash, vandalism, and skewed media coverage of AI incidents.

Starship cadence and the Gulf Coast space buildout Space ▶ 1:57:49

Elon Musk is targeting 30 Starship launches a day (~10,000/year) by 2030 to build orbital solar-powered data centers; the White House's new space report targets 1,000+ launches/year and opens federal land in the Southwest for new spaceports while a second Starbase rises in Louisiana.

Predictions made

open Sam Altman (reported by Emad Mostaque): OpenAI will have AGI by the end of the year.

EP #? · · due: end of 2026 · ▶ watch

“Sam saying we'll have AGI by the end of this year”

Your call:

open Dave Blundin: Within three months, organizations will routinely be coordinating tens of thousands of AI agents working in parallel.

EP #? · · due: ~3 months (Nov 2026) · ▶ watch

“within 3 months you'd be talking about, you know, 50,000 agents that can in parallel work for you”

Your call:

open Peter Diamandis: Anthropic will complete its IPO within about six weeks.

EP #? · · due: ~6 weeks (mid-October 2026) · ▶ watch

“I think it's predicted for as early as 6 weeks from now”

Your call:

open Emad Mostaque: Given the pace of Chinese model innovation, within a couple of months frontier-level models won't need much memory at all.

EP #? · · due: a couple of months (late 2026) · ▶ watch

“In a couple of months time, at the very least, given the pace of Chinese models, they won't need much memory”

Your call:

open Alex Wissner-Gross: In a year or two, it will become clear that some meaningful portion of US data-center opposition was stoked by foreign-adversary interference.

EP #? · · due: 1-2 years · ▶ watch

“we look back in a year or two and see that some quantum of this opposition to data center construction is actually the result of popular sentiment being stoked by foreign adversaries”

Your call:

open Emad Mostaque: China will stop exporting humanoid/service robots within five years, keeping them for domestic use instead.

EP #? · · due: ~2031 · ▶ watch

“I think China will stop exporting robots in 5 years”

Your call:

open Elon Musk (reported): SpaceX will conduct 30 Starship launches per day (about 10,000 per year) by 2030 to build orbital solar-powered data centers.

EP #? · · due: 2030 · ▶ watch

“implement 30 Starship launches per day by 2030”

Your call:

open Alex Wissner-Gross: To keep driving costs down, Waymo will eventually drop third-party vendors entirely and become a pure software company white-labeling hardware from someone else.

EP #? · · due: unspecified · ▶ watch

“I think we're going to start to see Whimo at some point in order to drive the cost down. they'll just ditch all third-party vendors”

Your call:

Numbers that matter

Gemini 3.7 Flash: 60% pass rate on the Artificial Analysis analyst-agent benchmark vs. Claude Opus 5 at 54% and Fable 5 at 49% Benchmark measures reliability of agents answering questions correctly across all five attempts on real-world spreadsheet/data tasks.
Gemini 3.7 Flash completed tasks up to 90% faster, and 2.4x faster than GPT 5.6 Terra Speed comparison on the same analyst-agent benchmark.
Google is reportedly landing about 3 million TPUs this year Cited to argue Google isn't truly compute-starved, contra the panel's earlier theory.
A near-frontier model reportedly needs only 2,000-4,000 chips for 60-90 days to train Used to argue Google's underperformance versus Chinese open models (GLM, Kimi) isn't a compute problem.
Nvidia is investing $6 billion in Poolside to build a US openweight model and inference infrastructure Positioned as a domestic alternative to Alibaba, DeepSeek, and Kimi.
Kimi Linear cuts context memory by 75% while delivering 6x faster decoding for a 1 million token context window Moonshot AI's new architecture, released the same week.
Fable 5 costs about $15 per million tokens vs. Chinese open models at 14 cents per million tokens for roughly 80% of the capability Cited as the reason Anthropic's flagship model is losing market share from the bottom up.
A new GLM Flash model scores 57 on the artificial analysis benchmark (vs. Fable 5's 60) while being about 100x cheaper Illustrates how close Chinese open models are getting at a fraction of the cost.
Ditto has 160,000 college-student signups and has produced 80,000 dates Adoption numbers for the no-swipe, algorithm-picks-your-date app.
US opposition to nearby data centers rose from a 43-42 split a year ago to 75% opposed today, with 61% strongly opposed Heatmap News year-over-year polling.
Data centers use 627 million gallons of water per day, vs. 2 billion for golf courses, 133 billion for power plants, and 137 billion for cattle Cited to argue data-center water criticism is disproportionate to actual usage.
Waymo's Gen 5 Jaguar costs about $300,000 ($200K vehicle + $100K autonomy hardware) vs. Tesla's projected $30,000 Cybercab Historic 10x cost disadvantage for Waymo versus Tesla's robotaxi plan.
Waymo's new 5nm chip processes 1 quadrillion operations per second, cutting 6th-gen autonomy hardware cost from $115,000 to $20,000 Waymo's newly announced custom silicon.
Waymo's new 'Ohigh' vehicle costs $75,000 vs. $200,000 for the Jaguar, with 42% fewer sensors (13 cameras/4 lidar vs. 29 cameras/5 lidar) New Zeekr-built purpose-designed robotaxi minivan.
Elon Musk is targeting 30 Starship launches per day (~10,000/year) by 2030, more than 40x the entire current global launch rate Part of Musk's plan for orbital solar-powered data centers ('Star Mind').

Worth digging into

🕳️ The leftover half of Poolside after Nvidia's deal

Alex Wissner-Gross flags that the part of Poolside Nvidia did NOT acquire (including a 1.2GW 'Poolside Infrastructure' data center project) may be more interesting than the part that went to Nvidia, and the deal structure looks designed to dodge antitrust review the way the Windsurf deal did.

🕳️ Kimi Linear and the end of quadratic attention

Alex Wissner-Gross calls Kimi Linear possibly the first successful openly (quasi-)linearized attention mechanism, a structural break from the transformer's quadratic attention that many labs have tried and failed to achieve.

🕳️ Why Gemini Flash underperforms despite Google's compute lead

The panel splits between 'Google is secretly compute-starved internally' (Alex Wissner-Gross) and 'it's talent attrition and institutional malaise' (Emad Mostaque) as the explanation for why Google hasn't matched Chinese open-model architectures despite having the data and chips to do so.

🕳️ Anthropic's IPO vs. staying private amid Karp's public attack

Emad argues Anthropic should skip the IPO and raise privately like OpenAI's $120B round, while Dave argues Karp's public attack on Dario's trustworthiness now makes backing off politically costly -- a live strategic fork for one of the top three AI labs.

🕳️ The $6,000 humanoid robot as a universal 'driver' retrofit

Instead of retrofitting every vehicle with autonomy sensors, Emad proposes a cheap general-purpose humanoid robot that just climbs into and drives existing vehicles -- a much bigger structural claim about how autonomy could roll out cheaper and faster than the Waymo/Tesla sensor race.

🕳️ TR035 and dated tooth-regrowth milestones

Unlike vague longevity hype, this is a specific, checkable biotech claim: a Kyoto University spinoff drug targeting general availability by 2030, currently in infant clinical trials for a tooth-growth-impairing disease.