2025-11-17

AI Costs Plummeting 40x: Why Costs Are Collapsing & What It Really Means w/ Dave, Salim & AWG

The WTF Just Happened in Tech crew covers Anthropic's rise versus OpenAI, new world-model and memory-pruning AI research, and Moonshot AI's ultra-cheap Chinese Kimi models as evidence of 40x year-over-year 'hyperdeflation' in the cost of intelligence. Peter shares new global survey data showing cost of living, unemployment and inequality are the world's top fears, sparking a debate about UBI, mindset, and storytelling as the antidote. The back half covers the data-center/nuclear power buildout, China's record drone swarm, Blue Origin's first booster landing, Boston unions fighting Waymo, and a run of science/longevity news including AI doing new science, cheaper GLP-1 drugs, and startups pursuing embryo editing outside the US.

▶ Watch on YouTube

Topics

Anthropic vs OpenAI enterprise strategy AI ▶ 3:32

Anthropic overtakes OpenAI in enterprise LLM API share and projects huge 2028 revenue and margins; the panel debates whether Anthropic's code-generation focus is the critical path to recursive self-improvement.

World models: Marble vs Genie 3 AI ▶ 16:02

World Labs' Marble (3D Gaussian splats, client-side rendering) and Google's Genie 3 (server-side pixel generation) are opposite compute/versatility trade-offs for generating explorable virtual worlds; the bigger market is synthetic training data for robots and science, not consumer worlds.

Forgetting memorized data without losing reasoning AI ▶ 18:02

A Goodfire paper shows how to prune a pretrained model so it forgets memorized facts while keeping general reasoning ability, pointing toward much smaller, more efficient generally-intelligent models.

Nested learning and compression theory of AGI AI ▶ 23:45

Google's Nested Learning NeurIPS paper suggests model architectures and optimizers are facets of one underlying process, feeding Alex's thesis that AGI is fundamentally about compressing world knowledge past a phase-transition threshold.

Moonshot AI's ultra-cheap Kimi models AI ▶ 28:08

Alibaba-backed Moonshot AI released Kimi, a near-frontier open-weight model trained for a few million dollars on Groq hardware, now among the best open-weight models available since Meta and OpenAI stopped open-sourcing.

40x year-over-year hyperdeflation in AI cost Economy ▶ 30:48

Sam Altman's cited 40x annual drop in the cost of intelligence per unit is framed by Alex as an 'inner loop' that will eventually drag down costs across healthcare, food production, and other sectors.

Europe loosens GDPR to compete in AI Geopolitics ▶ 35:54

Brussels is moving to loosen GDPR as Italy, Spain and other EU economies report AI products arriving 6-12 months late and costly compliance audits delaying 40% of AI projects.

Global anxiety: cost of living, jobs, inequality Economy ▶ 40:46

FII's 32-country Priority Global Survey shows cost of living, unemployment, and poverty/inequality are the top three global concerns, grounding the panel's 'future of abundance' optimism against lived reality.

Data centers, nuclear power, and capital flows Energy ▶ 52:12

Gigawatt-scale data centers are multiplying faster than power supply; an $80B deal aims to build about 10 AP-1000 nuclear reactors, but that covers roughly a tenth of the 92 GW the US reportedly needs by 2030.

Drone swarms as the real robotics story Robotics ▶ 1:01:30

China's 16,000-drone world record, the drone-dominated war in Ukraine, and 10,000 drones a month crossing the US border all point to coordinated drone swarms, not humanoids, as the underappreciated near-term robotics impact.

AI accelerates science and cures disease AI ▶ 1:20:23

GPT-5 shows 'tiny glimmers' of AI doing new science and Edison's Cosmos AI-scientist system compresses months of research into 12 hours; meanwhile Zuckerberg and Chan are boosting Biohub compute 10x by 2028, and cheap $149/month GLP-1 drugs already hint at what AI-driven abundant healthcare could look like.

Embryo editing moves from selection to alteration Biotech ▶ 1:33:18

Startups like Preventive and Manhattan Genomics are building CRISPR embryo-editing capability, blocked by FDA rules in the US and reportedly eyeing jurisdictions like the UAE; the panel notes no single US federal statute actually bans germline editing.

Predictions made

open Alex Wissner-Gross: Many if not most grand challenges in math, science, engineering, and medicine will start to fall to AI.

· due: 2028 · ▶ watch

“many if not most grand challenges in math, science, engineering, and medicine start to fall to AI over the next three maximum years”

Your call:

open Dave Blundin: Grand challenges in math, science, engineering, and medicine will start to fall thanks to AI's 40x hyperdeflation.

· due: 2028 · ▶ watch

“we're going to start to see grand challenges in math, science, engineering, medicine start to fall over the next two to three years thanks to that”

Your call:

open Alex Wissner-Gross: AI-driven virtual-cell research could cure all disease.

· due: 2030 · ▶ watch

“This could happen in the next 5 years. Call it 2030.”

Your call:

open Peter Diamandis: The next five years of AI/energy infrastructure decisions by countries like Italy and Spain will set the objective for the next century.

· due: 2030 · ▶ watch

“I think the next five years we're going to set the objective for the next century”

Your call:

open Dave Blundin: The flow of capital into data center construction and power will reach $1.2 trillion a year.

· due: 2030 · ▶ watch

“the flow of capital... is going to be $1.2 trillion a year by 2030 coming just into data center construction and power for it”

Your call:

open Alex Wissner-Gross: We'll be 'drowning' in generalist-capable humanoid robots.

· due: unspecified (next few years) · ▶ watch

“We're going to be drowning in humanoid robots that are generalist in terms of their capabilities in the next few years”

Your call:

open Peter Diamandis: Elon Musk will reveal an unforgettable demo of Tesla/SpaceX hop or hover propulsion technology.

· due: 2025 · ▶ watch

“he's gone on record saying by the end of the year there will be an unforgettable demo”

Your call:

open Alex Wissner-Gross: A generational public conversation about legalizing germline editing will happen.

· due: unspecified (near-term) · ▶ watch

“I wouldn't be surprised if in the near-term future we find a generational conversation about whether germline editing should in fact be allowed”

Your call:

Numbers that matter

Anthropic projects $70B revenue and $17B cash flow by 2028 at a 77% profit margin Anthropic's enterprise-first business model versus OpenAI's consumer growth strategy.
OpenAI projects $100B revenue in 2028 but stays unprofitable until 2029 Contrast with Anthropic's projected profitability and margin.
Moonshot AI trained its Kimi model for about $4.6-6M, 30-40x cheaper than it cost OpenAI/Anthropic to build their original frontier models Chinese open-weight models are now among the best open-source options available.
Sam Altman cites 40x year-over-year deflation in the cost of intelligence per unit ('hyperdeflation') Framed as the core driver of the episode's cost-collapse theme.
EU AI compliance audits cost about €260,000 and take 8-15 months, delaying 40% of projects Cited as a key reason EU AI products lag the US by 6-12 months.
FII's Priority Global Survey covers 32 countries, 60,000+ respondents, about two-thirds of the world's population Basis for the cost-of-living and unemployment fear data discussed.
AP-1000 nuclear reactors cost about $7B each; an $80B Brookfield/Cameco-backed partnership aims to build roughly 10 across the US Nuclear buildout for AI data center power demand.
Eric Schmidt told Congress the US needs 92 gigawatts of new power by 2030 Compared against the roughly 11 GW the 10-reactor AP-1000 deal would add.
China set a world record flying 16,000 AI-coordinated drones together Demonstration of millimeter-precision swarm coordination.
An estimated 10,000 drones a month cross the Mexico-US border Dave's point that drone technology already outpaces border-wall technology.
TrumpRx targets about $149/month for GLP-1 drugs, which cut repeat-stroke incidence by roughly half over 3 months Cited as an early example of AI-era abundant, low-cost healthcare.
Edison's Cosmos AI-scientist system completes 4-6 months of expert human research in 12 hours, reading about 1,500 papers and running about 42,000 lines of code per experiment Example of scaffolding-based agentic science.

Worth digging into

🕳️ Goodfire's weight-pruning/forgetting paper

A technique for distilling models down to pure reasoning with externalized knowledge could be the path to sub-billion-parameter generally intelligent models.

🕳️ Google's Nested Learning NeurIPS paper

Alex frames it as fishing for a 'grand unified theory of machine learning' where models and optimizers are facets of compression, a core piece of his AGI thesis.

🕳️ FII Priority Global Survey raw data

32 countries, 60,000+ respondents on cost of living, unemployment and inequality is a rare ground-truth check against the panel's abundance narrative.

🕳️ Edison's Cosmos AI-scientist system

Claims of compressing months of expert research into 12 hours and surfacing Alzheimer's-relevant findings would be a major agentic-science milestone if verified.

🕳️ Embryo-editing startups going offshore

Preventive and Manhattan Genomics are reportedly looking at jurisdictions like the UAE to do CRISPR embryo editing the FDA won't allow in the US.