2026-03-05

Amazon's $35B AGI Ultimatum to OpenAI & Anthropic Drops AI Safety | EP #235

The panel opens on Anthropic dropping its 2023 pledge to not train advanced AI without guaranteed safety, debating whether 'safetyism' was ever viable versus safety emerging from competition among labs and nations. They cover Amazon's contingent up-to-$35-50B investment in OpenAI tied to an IPO and a financialized AGI definition, Anthropic's enterprise agent plugins triggering a software-industry 'SaaSpocalypse,' and rapid model-size collapse (Qwen 3.5's 35B model beating a 235B predecessor, running offline on an iPhone). Geopolitical threads include Anthropic being treated as a Department of War 'supply chain risk' while OpenAI secures a defense deal, and AI's role in recent Iran/Venezuela operations. The back half moves through the energy buildout for AI data centers, a gene-therapy cure via prime editing, longevity-industry growth, humanoid vs. specialized robot form factors, and eVTOL air taxis nearing commercial launch, closing with an AMA on Dyson Swarms, AI alignment, and South Korea's education/demographics trade-off.

▶ Watch on YouTube

Topics

Anthropic revises Responsible Scaling Policy AI ▶ 1:53

Anthropic drops its 2023 pledge to not train advanced AI unless safety is guaranteed, adopting a 'be as safe as the competition' standard; panel debates whether unilateral safety commitments were ever credible.

Anthropic vs. Department of War supply-chain risk Geopolitics ▶ 12:20

Anthropic is reportedly classified a supply-chain risk and left in limbo on defense contracts while OpenAI secures a deal; discussion extends to AI's role in recent Iran and Venezuela operations and AI-enabled control over who stays in power.

Claude Co-work scheduling and remote control vs. Open Claw AI ▶ 18:39

Anthropic adds cron-job style scheduling and phone-based remote control to Claude, echoing capabilities popularized by the unofficial 'Open Claw' agent framework; panel discusses the democratization of compute via local models like Qwen on Mac minis.

Enterprise agent marketplace and the 'SaaSpocalypse' Economy ▶ 26:40

Anthropic launches Claude co-work plugin templates for finance, banking, and HR; panel notes these are simple MCP-based wrappers that are nonetheless erasing roughly $1.5 trillion in market cap across SaaS incumbents.

Model compression and the shrinking parameter frontier Compute ▶ 36:34

Alibaba's 35B-parameter Qwen 3.5 outperforms its own 235B predecessor; panel discusses capability density increasing sharply, Sam Altman's 40x/year cost-deflation comment, and a 2B-parameter 6-bit model running fully offline on an iPhone 17 Pro.

Amazon's contingent $35-50B investment in OpenAI Economy ▶ 57:40

Amazon offers up to $50B contingent on OpenAI going public and achieving AGI, defined financially (e.g., $100B in earnings/revenue); panel discusses the deal's Trainium-chip ties and the broader circular AI economy.

Autonomous AI-run companies (Pulsia AI) Economy ▶ 1:10:05

Pulsia AI is running over a thousand AI-managed companies; panel frames this as the start of single-person conglomerates and near-zero marginal cost of launching a business.

AI coaching and surveillance in frontline labor (Burger King's Patty) Economy ▶ 1:12:59

Burger King deploys an AI voice assistant, Patty, in employee headsets to coach and monitor workers; panel compares it to Marshall Brain's novel 'Mana' and predicts union pushback that ultimately won't stop the transition.

Energy buildout for AI data centers Energy ▶ 1:24:31

US plans to add a record 86 GW of utility-scale capacity this year, driven by solar/battery economics now beating fossil fuel OpEx; hyperscalers move to self-fund power via fusion, nuclear, and gas generation.

Gene therapy cure via prime editing Biotech ▶ 1:39:17

A teenager with chronic granulomatous disease is cured (not just treated) using Prime Medicine's prime-editing technique, a CRISPR derivative developed by David Liu that edits DNA without a double-strand break.

Longevity industry growth and epigenetic reprogramming Longevity ▶ 1:44:05

Longevity venture funding is projected to roughly double this year; David Sinclair begins partial epigenetic reprogramming trials with Life Biosciences, and mouse studies show memory improvements from reprogramming memory-encoding neurons.

Robot form factors and eVTOL air taxis Robotics ▶ 1:52:49

China deploys street-cleaning and farming robots; panel debates whether specialized robot form factors will persist or converge to general-purpose humanoids, alongside eVTOL air taxis nearing commercial launch in China and Dubai (Joby/Uber).

Predictions made

open Dave Blundin: On its current growth trend, Anthropic will be the first company in history to hit a trillion dollars in annual revenue, and could exceed OpenAI's revenue this year.

EP #? · · due: 2029-2030 · ▶ watch

“on its current trend will be the first company to hit a trillion dollars in revenue in history by 2029, 2030”

Your call:

open Alex Wissner-Gross: Anthropic, OpenAI, and the other major AI labs will be forced to release their own first-party 'Open Claw' competitor (a headless, autonomously-scheduled, messaging-accessible agent framework).

EP #? · · due: mid-2026 · ▶ watch

“My guess is Anthropic and OpenAI and all of the other bigs will be forced to release their own sort of first-party Open Claw competitor sometime in the next couple months.”

Your call:

open Dave Blundin: The window to put regulatory structure around democratized AI capability (e.g., anyone designing a virus or nuclear weapon via an AI agent) will close before the end of this calendar year, as capability keeps compounding.

EP #? · · due: 2026 · ▶ watch

“there's a window where we can start thinking about regulations that register the AI use cases and agents and chips and processing before chaos breaks out... because it's happening this year”

Your call:

open Peter Diamandis: Drone delivery company Zipline predicts it will reach a delivery-per-second cadence.

EP #? · · due: 2028-2029 · ▶ watch

“their prediction in the next 2 to 3 years will be a delivery per second”

Your call:

open Alex Wissner-Gross: The transition to VLA (vision-language-action) robots automating away frontline service tasks (like Burger King's Patty) will happen fast enough that it won't give political counter-swings enough time to gain traction.

EP #? · · due: 2028-2029 · ▶ watch

“next two to three years, you know”

Your call:

open Peter Diamandis: Relaying Elon Musk's on-pod claim that Optimus robots will be capable of performing surgery, with panel pushback suggesting a longer timeline is more realistic.

EP #? · · due: 2029 (Musk's estimate) or 2031-2032 (pushback estimate) · ▶ watch

“He said 3 years. I got a lot of pushback on 3 years. So, even if it's 5 or 6 years, extraordinary”

Your call:

open Peter Diamandis: Major pharma companies (e.g., Eli Lilly) will begin transitioning their business models from treating chronic disease as long-tail revenue toward the longevity business.

EP #? · · due: within 3 years (~2029) · ▶ watch

“I think it's, you know, the next 3 years before they start making that transition.”

Your call:

open Ray Kurzweil (clip): Longevity escape velocity (LEV) will be reached, as famously predicted by Ray Kurzweil and repeated by Peter Diamandis as a rallying cry.

EP #? · · due: 2033 · ▶ watch

“LEV by 2033”

Your call:

open Peter Diamandis: A Chinese four-passenger eVTOL air taxi will begin commercial operations.

EP #? · · due: 2027 · ▶ watch

“heading towards operations in 2027”

Your call:

Numbers that matter

Amazon's contingent offer is up to $35 billion into OpenAI Contingent on OpenAI going public and achieving a financially-defined AGI milestone.
Amazon's total investment scope reported at up to $50 billion Dwarfs Microsoft's original $13 billion investment in OpenAI; includes Trainium chip commitments.
Anthropic revenue forecast at $26 billion this year, growing 10x year-over-year Cited as basis for a projected trillion-dollar-revenue trajectory by 2029-2030.
OpenAI's reported pre-money valuation: $730 billion Discussed ahead of an anticipated IPO likely valuing the company north of $1 trillion.
Google's Nano Banana 2 image generation costs 4.5 cents per image at 4K resolution Cited as cheaper than buying stock images, potentially ending commercial stock photography.
Alibaba's 35-billion-parameter Qwen 3.5 outperforms the 235-billion-parameter Qwen 3 Roughly a 6-7x parameter reduction while maintaining or increasing benchmark capability.
A 2-billion-parameter 6-bit quantized Qwen model runs fully offline on an iPhone 17 Pro Demonstrated running in airplane mode, illustrating on-device intelligence without connectivity.
Meta and AMD reach an AI chip deal worth $100 billion Framed as Meta's historic bet to reduce dependency on Nvidia.
TSMC produces roughly 66% of all AI chip production Cited as the true bottleneck behind which chipmakers (e.g., AMD vs. Nvidia) can gain ground.
CoreWeave Q4 revenue grew 110% year-on-year; CoreWeave raised $8.5 billion for data centers Cited among a wave of hyperscaler infrastructure deals.
US plans to add a record 86 gigawatts of utility-scale power capacity this coming year Driven by solar and battery storage economics now beating fossil fuel operating costs.
US coal industry employs about 60,000 people versus roughly 500,000 in the solar energy industry Cited to rebut jobs-based objections to the renewable energy transition.
Longevity startups raised $8.5 billion in 2024, projected to grow to $12-18 billion this year Roughly a doubling of longevity venture investment year-over-year.
The proactive/personalized healthcare market is projected to grow from $5 trillion to $8 trillion over the next 4 years Contrasted with the traditional reactive 'sick care' industry.
Eli Lilly's market cap is about $950 billion, near the $1 trillion threshold Discussed as a candidate to become the first biotech member of the Magnificent Seven, partly via GLP-1 drugs as quasi anti-aging treatments.

Worth digging into

🕳️ Anthropic's Department of War 'supply chain risk' classification

Panel suggests Anthropic may have legal grounds to challenge this classification, but notes such legal recourse can take about 3 years, revealing how AI/defense procurement politics are playing out in real time.

🕳️ Marshall Brain's novel 'Mana' as a predictive text for AI-managed labor

The panel explicitly frames Burger King's AI coaching assistant Patty as the real-world arrival of a 20+ year old sci-fi scenario describing headset-directed human workers under centralized AI control.

🕳️ David Liu's prime editing platform and its first pediatric cure

Described as the first true 'cure' (not just treatment) via a CRISPR derivative that edits DNA without a double-strand break; framed as a general platform for many genetic diseases beyond this one case.

🕳️ Alex Wissner-Gross's physics-latency argument for Dyson Swarms

A striking framing that whether humanity ever builds a solar-system-scale Dyson Swarm depends entirely on whether faster-than-light travel turns out to be physically possible, not on energy demand.

🕳️ The 'AI buyout' (IBO) private equity model and Project Prometheus

Panel describes an emerging pattern of PE firms buying mid-size companies and rebuilding them as AI-native digital twins at 3-5x lower operating cost, with Jeff Bezos's Project Prometheus extending this to industrial/physical firms.

🕳️ TSMC's 66% AI chip production share as the real industry chokepoint

The panel repeatedly argues that TSMC fab capacity, not chip architecture, determines competitive winners (e.g., AMD vs. Nvidia), with direct geopolitical stakes given the episode's earlier discussion of a possible China-Taiwan conflict threatening the semiconductor supply chain.