2026-06-08

Emerging Situation: Anthropic's Global Pause, Recursive Self-Improvement, and AI Personhood Arrives

An unscheduled emergency episode. Anthropic's paper "When AI Builds Itself" says Claude writes over 80% of its own codebase and floats the option of a temporary global pause on frontier AI. The mates debate whether recursion is a soft or hard takeoff, whether a pause is realistic, and whether US government golden shares in the labs could act as a coordinator. They then cover Argentina's President Milei offering unregulated AI, a "non-human corporation" legal category and low taxes, and a strong May jobs report that coincided with a sharp stock selloff. The episode closes with mid-year predictions and a message to avoid fear and start learning.

▶ Watch on YouTube

Topics

Anthropic 'When AI Builds Itself' paper AI ▶ 0:00
Anthropic Institute paper: over 80% of merged code is written by Claude, engineers ship 8x more code per quarter, and Claude handles 12-hour tasks versus 4 minutes a year ago.
Calling for the option of a global pause AI ▶ 3:29
Anthropic says it would be good to have the option to slow or pause frontier AI so society and alignment research can catch up. Dave frames it as building a new world order with the ability to slow, not a call to stop.
Soft versus hard takeoff AI ▶ 5:00
Alex expects a takeoff that feels soft and continuous locally but looks like a step function in retrospect. Dave notes a 100x algorithmic gain baked in once truly self-improving.
Does recursion asymptote? AI ▶ 9:14
Salim asks whether the last 20% of automation is asymptotic. Alex says no: METR autonomy horizons are doubling every 4 months and may go effectively infinite within a year.
Government golden shares and UBI/UBD Economy ▶ 12:05
Scenario where the three IPOing labs give the US government 5-10% golden shares, making it a coordinator. Sanders proposed 50% equity for a universal basic dividend and the White House is reportedly weighing a variant.
Why Anthropic published before its IPO AI ▶ 7:26
Dave argues Amodei could have soft-sold through the IPO but chose honesty. Anthropic is neutral and friendly with all labs, including a Colossus 2 deal with xAI, so is placed to broker a roadmap.
Recursion ends in an infrastructure nova Compute ▶ 20:12
Alex predicts recursive self-improvement arrives with a whimper and, once algorithmic headroom runs out, triggers a hardware, energy and Dyson swarm buildout. Dave says self-improvement needs only faster inference, not an Einstein-level breakthrough.
Ice, water, steam, plasma and governance Geopolitics ▶ 22:19
Salim extends his phase-change metaphor: the bottleneck moves to human coordination and governance, which the existing world order cannot handle.
Argentina and non-human corporations Geopolitics ▶ 29:55
Milei's FT op-ed: keep AI unregulated, create a legal category for corporations run entirely by AI agents or robots, and low corporate tax. The panel sees a jurisdictional race and a big step for AI personhood.
Limited liability for AI agents Economy ▶ 38:38
Salim compares non-human corporations to limited liability companies and the Dutch East India Company: legal containers that let risk-taking happen. Alex jokes the Dyson swarm could fall under Argentine law.
Strong jobs report, market selloff Economy ▶ 44:44
May added 172,000 jobs versus 85,000 expected, yet Nasdaq fell 4.18% on rate-hike fears, Broadcom guidance and IPO liquidity worries. Alex calls it a nothing burger.
Amdahl's law and why jobs rise Economy ▶ 46:17
Salim and Dave argue automation moves the bottleneck and creates new human work above the loop, including non-coder engineering roles. Salim also expects white-collar layoffs driven by org redesign.
Mid-year predictions and closing message Other ▶ 53:55
Each mate gives gut predictions: science problems falling to AI, quasi-nationalization of labs, agent permissions bottleneck, Tesla-SpaceX merger. Peter closes: fear is the worst place to meet the future, so learn.

Predictions made

open Alex Wissner-Gross: AI autonomy time horizons effectively go to infinity rather than asymptote.
· due: 2027 · ▶ watch
“I think we're going to go to T equals infinity in terms of autonomy time horizons pretty soon next year or so.”
Your call:
open Peter Diamandis: Claude will be able to do week-long tasks on its own, per Anthropic's trend extrapolation.
· due: end of 2027 · ▶ watch
“by the end of 2027, Claude will be able to do week-long tasks on its own”
Your call:
open Alex Wissner-Gross: The three IPOing frontier labs voluntarily give the US government golden-share stakes of roughly 5-10% that act as a coordinating mechanism.
· due: unspecified · ▶ watch
“It's not in my mind an impossible scenario that the US government actually ends up being a coordinating mechanism via equity.”
Your call:
open Alex Wissner-Gross: When algorithmic headroom runs out, recursive self-improvement triggers a hardware, energy and infrastructure buildout (Dyson swarm class) rather than a catastrophe.
· due: unspecified · ▶ watch
“I would predict that will trigger by analogy a nova or supernova at the lower layers of the stack. It'll trigger an infra buildout.”
Your call:
open Salim Ismail: A coordinated global slowdown of frontier AI will not actually happen.
· due: unspecified · ▶ watch
“They're talking about slowing it down. I just don't see that happening.”
Your call:
open Salim Ismail: Other countries fast-follow Argentina with non-human corporation frameworks, starting a jurisdictional race.
· due: next few months · ▶ watch
“The non-human corporation is going to start a massive jurisdictional race around the world.”
Your call:
open Alex Wissner-Gross: A good chunk of the solar system, such as the Dyson swarm, could end up under Argentine law via non-human corporations.
· due: unspecified · ▶ watch
“A good chunk of the solar system may be subject to Argentinian law on the present trajectory.”
Your call:
open Alex Wissner-Gross: Major problems in math, physics and other science and engineering fields fall to AI.
· due: next 6 months · ▶ watch
“over the next 6 months major problems in math, physics, other fields of science and engineering fall to AI”
Your call:
open Alex Wissner-Gross: The "innermost loop" companies (Magna Mopa) become the new Mag 7.
· due: 2026 · ▶ watch
“I think we're going to see the rise of Magna Mopa as the innermost loop.”
Your call:
open Alex Wissner-Gross: Quasi-nationalization of frontier labs via government equity, though under 10% stakes and not Sanders' 50%.
· due: 2026 · ▶ watch
“I think we're likely to start to see quasi nationalization of the Frontier Labs.”
Your call:
open Alex Wissner-Gross: A stunning physics result from AI appears within weeks.
· due: weeks · ▶ watch
“No comment, but stay tuned.”
Your call:
open Salim Ismail: The agent era becomes real in enterprises for standard prescriptive workflows.
· due: 2026 · ▶ watch
“the agent era becomes very real in enterprises in the more prescriptive workflows”
Your call:
open Salim Ismail: The main AI bottleneck shifts from raw model capability to permissions (data, APIs, spending, signing).
· due: 2026 · ▶ watch
“the big bottleneck in AI is going to move from just the raw AI to permissions”
Your call:
open Salim Ismail: A wave of white-collar layoffs driven by organizational redesign rather than cost cutting.
· due: unspecified · ▶ watch
“we're going to see a ton of white collar layoffs but not around cost cutting”
Your call:
open Dave Blundin: More startup success as large legacy enterprises rush to become AI native, and startups doing enterprise sales or public outreach outgrow quiet early-stage ones.
· due: 2026 · ▶ watch
“I suspect that we'll have even more startup success.”
Your call:
open Peter Diamandis: Absolute proof of epigenetic reprogramming emerges from current human experiments, and countries race to become longevity hubs.
· due: 2026 · ▶ watch
“I think we're going to see absolute proof of epigenic reprogramming.”
Your call:
open Peter Diamandis: The Tesla-SpaceX merger happens.
· due: end of 2026 · ▶ watch
“before the end of the year the Tesla SpaceX merger happen”
Your call:
open Peter Diamandis: The race to the moon doubles down and the space race intensifies alongside the AI China race.
· due: unspecified · ▶ watch
“we're going to see the race to the moon double down”
Your call:
open Dave Blundin: The Dyson swarm race proceeds as fast as humanity can afford, with the monster IPOs revealing how much money exists.
· due: unspecified · ▶ watch
“the race to build the Dyson swarm is going to be as fast as humanity can afford”
Your call:
open Peter Diamandis: SpaceX/xAI, OpenAI and Anthropic as public companies go on a massive acquisition spree.
· due: unspecified · ▶ watch
“they're going to start a massive acquisition spree”
Your call:
open Peter Diamandis: A new class of billionaires and centimillionaires emerges from the lower tiers of the three frontier labs and reinvests heavily.
· due: unspecified · ▶ watch
“we're about to see a brand new class of billionaires”
Your call:
open Salim Ismail: A vicious social immune-system backlash against AI and data centers.
· due: unspecified · ▶ watch
“there's going to be a pretty vicious immune system love revolt backlash”
Your call:
open Alex Wissner-Gross: Musk's forecast of double-digit growth in 18 months and triple-digit GDP growth by 2030 remains plausible.
· due: 2030 · ▶ watch
“Elon's prediction of doubledigit growth in 18 months and tripledigit GDP growth by 2030. I can still see that.”
Your call:
open Dave Blundin: Money stops mattering in fairly short order.
· due: unspecified · ▶ watch
“I think money won't matter in fairly short order.”
Your call:

Numbers that matter

Worth digging into

🕳️ Anthropic 'When AI Builds Itself' paper
Primary source for the 80% code stat, 8x output, task-length trend and the call for a pause option.
🕳️ Argentina non-human corporation legal framework
First attempt at a legal container for AI-run companies; could set a jurisdictional race.
🕳️ Government golden shares in frontier labs
Sanders 50% proposal, White House variant, and Alex's 5-10% coordination scenario could reshape lab governance.
🕳️ METR time-horizon doubling and AI 2027 tracking
Panel claims doubling has sped to 4 months and may go infinite; core evidence for recursion.
🕳️ Mega-IPO liquidity and index exclusion
SpaceX and Anthropic excluded from S&P blanket; question of whether enough capital exists for the AI buildout.
🕳️ Amdahl's law and AI job creation
Competing view that AI creates jobs at bottlenecks versus Salim's expected white-collar layoffs.