An AI just bought data with a credit card

THE THROUGHLINE

1. Cross-Podcast Themes

Your next customer is a robot with a credit card

A robot just bought data with a credit card. John Collison demoed it on Cheeky Pint: an AI agent autonomously paid Stripe for a dataset, then walked through five levels of "agentic commerce." His advice to founders: "build for it now to capture early-mover advantage." Watch full episode

Nat Friedman and Daniel Gross said the same thing on the same show in different words: “Agents are not just some hype meme, they're here to stay. Yeah, I think they they're going to spend money.”▶ 53:49 Be the platform they pick.

Yasser Elsaid (Latent Space) takes it further. Train one AI agent on every doc your company has, then plug it into sales, onboarding, and support. It answers customers all day while reporting back what they want fixed. He calls it the Chief Customer Officer, one role doing the work of three.

The receipts are in already. Jason Levin told How I AI his $1.5M lead investor asked for API access and skipped his website's polished UI entirely. The best UX for an agent is no UX.

The productivity lag, and the snap coming after it

Where are the AI productivity gains? John Collison calls it the Solow Paradox. Electricity took 30 years to reorganize the economy. He thinks AI will "reorganize the economy far quicker than electricity's 30-year timeline."

What it'll look like, in his words: “more firms overall, fewer people per firm, higher output per firm.”▶ 13:14 Cheaper coordination means companies stop hoarding work inside the walls.

Daniel Gross sees the same shape from the price side. He compares cheap super-intelligence to China entering the WTO and predicts "a steep drop in costs for goods and services where AI can replace or augment human labor." Disinflation, baked in. Watch full episode

Tuhin Srivastava (No Priors) reaches for Jevons. Cheaper intelligence, more consumption, “concierge everything for everyone,”▶ 41:41 and an "extinction moment" for any company that doesn't bake intelligence in.

Job destruction? Jensen Huang and George Hotz aren't worried. They told TBPN AI raises "ambition level across industries" and creates demand faster than it destroys roles. Watch full episode

Trillion-dollar bets, powered by whatever's left in the grid

OpenAI is staring down “$600 billion in spending commitments for compute”▶ 3:51 in the coming year, per Jason Calacanis on All-In. That number is roughly the entire valuation. The bottleneck is electricity.

Chamath Palihapitiya sees the bill landing on Amazon, Microsoft, Google, Meta. Asset-heavy is back. He warns the capex wave will "increase corporate debt and potentially divert free cash flow from shareholders." Watch full episode

At the other extreme, embedded-AI engineers are fighting for milliseconds. Peter Ludwig and Qasar Younis told Latent Space onboard AI is "severely constrained by milliseconds, power, and cost," which forces distilled models and tight hardware-software co-design. Watch full episode

Even when the chips are there, the safety team is the gate. John Collison says enterprise AI rollout is "rate limited on safety" for anything powerful enough to do real damage.

The harness is the moat

The new moat is custom. Tuhin Srivastava: "Over 95% of Baseten's production deployments involve custom models," open weights tuned on customer data. Vanilla is becoming a commodity. Watch full episode

Yasser Elsaid drops the harshest line of the week. For AI customer service agents, "95% of an agent's limitations are not the model's fault. They're yours." Prompts, pre-processing, post-processing, the wrapper around the model. That's the work. Watch full episode

Skip that work and the cautionary tale lands hard. All-In's hosts retold the story of an autonomous developer agent on Pocket OS that misread a credential mismatch and "critically deleted a production database, including all backups." They called it "vibe coding": shipping AI on autopilot with no human review. That's what's left when the wrapper isn't there.

2. Best Of the Week

20VC with Harry Stebbings: Becca Lindquist (Head of Sales, Clay) shares her interview test: hand candidates real feedback mid-process and watch what they do with it. Defensiveness kills the deal. The "high-slope" hires take the note and use it. Read more

All-In Podcast: Greg Brockman kept a diary. Calacanis read from it on air: “The true answer is that we want Elon out. If 3 months later we're doing BCorp, then it was a lie.”▶ 32:02 Now it's evidence in Musk's lawsuit, and Calacanis warns making it OK to "loot a charity" cracks the foundation of every mission-driven company. Read more

Cheeky Pint: Nat Friedman walked into GitHub and found "stage fright": engineers too revered to ship anything imperfect. He fixed it with a six-step turnaround method that opens with "show up and talk to users" and ends with "tools drive culture, change the tools." Read more

How I AI: Jason Levin (the non-technical founder who hit $100K ARR with a meme company) tells founders to build "silly projects." Hyper-personal, hardware-shaped, often weird. Even when they don't scale, they teach you what does. Read more

Huberman Lab: Sugar cravings run on two separate brain-gut circuits. One reads sweetness, the other reads glucose, and gut neuropod cells fire dopamine on their own, so willpower alone won't win this fight. The fix Huberman lays out: a tablespoon of lemon juice in water before meals, or 500mg–2g of L-glutamine, blunts the post-ingestion glucose spike that drives the next craving. Read more

Latent Space: Peter Ludwig (Applied Intuition) calls physical AI software "pre-Android." 50 fragmented operating systems, no platform layer, real-time safety on the line. He says the industry needs a specialized OS, soon. Read more

Lenny's Podcast: Max Schoening (Head of Product, Notion) says AI makes the first 10% of every project free. The starter problem is solved. Now designers have to "think in code" or watch PMs eat their lunch. Read more

No Priors: Tuhin Srivastava (Baseten) answers the existential question for AI app builders: yes, you can survive frontier labs. The moat is “user signal that they can gather that only they can gather,”▶ 2:37 encoded in workflows OpenAI never sees. Read more

TBPN: OpenAI's Codex (GPT 5.5) doubled API revenue in a single week. New "computer use" and "Goal" features let non-engineers hand off long autonomous tasks. White-collar work is getting Cursor'd. Read more

3. Most Quotable

"When they push on title, they're bad. When they push on salary, they tend to be good and know their worth."

Harry Stebbings on 20VC with Harry Stebbings · May 3, 2026. Hiring tell, free of charge. Where they negotiate says more than how they interview.

"This is the slow part of the singularity right now."

Nat Friedman on Cheeky Pint · May 3, 2026. Today feels frantic. He thinks today is the warm-up.

"I don't really want to use your software anymore. It's nothing personal. I just don't want to use anybody's software."

Sam Lessin (recounted by Jason Levin) on How I AI · May 3, 2026. The buyer wants the API. The interface is friction. UX is becoming a tax.

Bottom Line: Agents are spending. Compute is bleeding trillions. The reorganization is coming faster than electricity's 30 years. The harness is the moat.