Atoms

Travis Kalanick's industrial holding company. Vision: digitize the physical world. Not atoms.dev (MetaGPT "AI team for solo founders").

Source: atoms.co/vision — captured 2026-09-02.

Karthik wrote almost the same thesis in a physical notebook years ago. Alignment is the framework, not the holding company. Skcript copying kitchens, mines, or "wheelbase for robots" is cosplay.

Their vision, in their words

1. Understand the current state of the physical world.
2. Predict the future state.
3. Control the future state.

Treat atoms like bits. CPU = manufacturing. Storage = real estate. Network = transport. First computer was food (CloudKitchens, Otter, Lab37, Picnic, ProFood). Then mining (Pronto AI). Transport still "announcement soon."

Gainfully employed robots — specialized machines with jobs, not dancing humanoids. Pancake machine at 1000/hour, not a humanoid that also folds laundry. Humanoids are for messy, low-volume, human-shaped rooms. Industrial work wants throughput.

Eight years of kitchens as City Storage Systems. Rebrand to Atoms ~Mar 2026. Manifesto is downstream of a business, not the other way around. ~$1.7B equity (Jul 2026, a16z).

What that means for Skcript

Same energy. Not their category.

Atoms	Skcript
Physical-world computer	Customer-conversation computer for small SaaS teams
Food first, then expand	FeatureOS first ($15K MRR), then SupportWire
Specialized robot with a job	One job, one machine, closed loop
Anti-humanoid	No general "AI employees" / agent team

The OS in this vault is for Karthik. Otter's OS is for the restaurant. The product has to be the customer's computer, not ours.

Coach already killed the Grok-bot-as-headcount fantasy (Coach/2026-08-22-grok-bot.md). Rashim already showed isolated FeatureOS "agent" tools that fail the real customer workflow (Team/CEO-Feedback/Rashim — March 2026 Feedback.md).

The wedge (small)

Rule: one job, one machine, closed loop. No humanoid PM. No third product.

The computer we already own: feedback → roadmap → changelog → support. That's the kitchen. FeatureOS is the kitchen. SupportWire is the specialized arm.

Step 01 only: understand current state. A PM opens FeatureOS Monday morning and in 30 seconds knows what paying customers are actually demanding this week. Job that disappears: archaeology across Intercom + Slack + Canny + Linear. If we cannot sense, predict and control are fanfic.

The gainfully employed robot: SupportWire resolves the repetitive ticket. Not an AI teammate that "runs support." Price the work, not the chair. Wedge = FeatureOS customers who already hate Intercom — not Fin's general war.

Compound loop (only after sense works): shipped in changelog, same answer dies in support. One record.

What not to do

• Don't start a robotics / physical-AI company.
• Don't publish an 11-section vision letter. Travis earned his with kitchens.
• Don't confuse atoms.co with atoms.dev.
• Don't build general agents.
• Don't treat this vault as the product.
• Don't hide in vision while FeatureOS still needs names on a calendar. Distribution is the bottleneck.
• Don't ship a daily digest. Daily is the humanoid. Weekly is the pancake machine.

One thing (locked)

Weekly digest. Not daily.

Daily is a notification tax. People mute it. Canny already ships "here's what happened" and PMs ignore it. Weekly is the job that matches the product story: a PM opens FeatureOS Monday morning and in 30 seconds knows what paying customers demanded this week.

The machine

One email, Monday, plus the same view in-app. Not a new product. Not Slack until email is a habit.

Fixed body. No AI essay.

1. This week's demand — top 5 posts from paying / identified customers, ranked by new votes and comments this week, not all-time votes.
2. What moved — 3 posts that jumped (new, or +N from paying accounts).
3. If you only look at one thing: [that post]. One line.

Click goes to the post. Empty state is honest: "No paying-customer signal this week." That empty state is the reason to identify customers, not a prompt to hallucinate insights.

Not in v1

AI summary paragraphs. Sentiment charts. Predicted ship list. Team activity. Changelog draft. Daily cadence.

Kill test

If a paying PM would not miss it on a Monday, it's an activity dump. Kill it. Success is they forward one item to the team or reply "we're doing #2" without opening six tabs.

The 10-star version at our scale is not hardware. Skcript digitizes the customer conversation the way Atoms digitizes a kitchen — sense, then predict, then control — and refuses the humanoid demo.