On a Friday night in November 2025, an Austrian developer named Peter Steinberger wired a large language model into a Telegram bot with shell access.
Five months later: 331,000 GitHub stars — more than React, more than Linux. 2 million monthly active users. 1,000+ contributors shipping code every week. Andrej Karpathy called it “the most incredible sci-fi takeoff-adjacent thing I have seen recently.” Sam Altman hired the creator to “drive the next generation of personal agents.”
No employees. No revenue. Built in one hour.
How does a project built in an hour, with zero revenue, surpass every open-source project in GitHub history?
The answer isn’t what the AI did. It’s what the AI was allowed to do.
The billion-dollar distinction
Every AI product on the market in late 2025 used roughly similar models. ChatGPT, Claude, Gemini — different names, comparable capabilities. The models weren’t the variable.
What Steinberger changed was the access.
His project — OpenClaw — gave the AI root-level access to the entire operating system. File system, terminal, code, processes, everything. Other tools confined the same models to a chat window: you type, it responds, you copy-paste the output somewhere.
His words: “A moderately capable model with unrestricted permissions outperforms a genius model confined to a chat interface.”
Permissions over intelligence.
The model didn’t get smarter. It got more access. And that access changed everything — 331,000 developers agreed by starring the project. China’s Shenzhen government started offering subsidies up to $290,000 for businesses that deploy it. NVIDIA built an enterprise stack (NemoClaw) on top of it, with Cisco, Salesforce, and SAP as launch partners.
The pattern I can’t stop seeing
I’ve been sitting with this for a week, and I keep coming back to the same thought:
This is exactly how marketing diagnosis works.
Consider two scenarios:
Scenario A: You hire a digital agency. They get access to your ad accounts, your website analytics, your social media channels. Maybe your CRM. They see Layer 5 — campaign execution. They optimize what they can see. They produce reports on what they can measure.
Scenario B: You bring in a diagnostic that examines your positioning (Layer 3), your pricing logic (Layer 4), your team structure, your customer identity (Layer 6), the founder’s beliefs about the market (Layer 1), and your competitive landscape (Layer 7). Then — and only then — it looks at channels and campaigns.
Same intelligence. Wildly different permissions.
Scenario A is the chatbot: capable model, narrow window. Scenario B is the agent: same capabilities, system-level access.
And the difference isn’t subtle. It’s the difference between “your Facebook ROAS dropped 15%” and “your Facebook ROAS dropped because your offer doesn’t differentiate, and it doesn’t differentiate because the CEO believes ‘we serve everyone,’ and nobody has ever challenged that belief because the brand was built on the founder’s personal reputation.”
One diagnosis lives at Layer 5. The other traces the constraint through six layers to find where it actually lives.
Why nobody asks for system access
Here’s the part that makes this uncomfortable:
Most companies don’t deny access to their deeper layers. They simply don’t offer it — because nobody asks.
Agencies don’t ask because their scope is channels and campaigns. Consulting firms don’t ask because their scope is strategy decks. Tool vendors don’t ask because their scope is data dashboards.
Everyone sees their layer. Nobody connects them.
Steinberger’s insight wasn’t technical. It was structural: the constraint wasn’t the model’s capability — it was the architecture that limited what the model could reach.
In marketing: the constraint usually isn’t the agency’s talent, the tool’s functionality, or the team’s effort. It’s the architecture that limits what anyone is permitted to examine.
When I work with a company and ask “Can I talk to the sales team about why they discount?” or “Can I see the original brand positioning document?” or “What did the founder believe about the market when they started?” — you’d be surprised how often the response is a pause, then: “Nobody has ever asked that before.”
Not because it’s sensitive. Because nobody’s diagnostic framework included it.
The one-person unicorn pattern
There’s another angle here worth sitting with.
Peter Steinberger wasn’t a newcomer. He spent 13 years building PSPDFKit — a PDF rendering SDK used by Dropbox, DocuSign, SAP, and the European Patent Office. Bootstrapped it to $12M ARR with nearly a billion end-users. Got a €100M+ strategic investment from Insight Partners in 2021.
Then he burned out. Booked a one-way ticket to Madrid. Disappeared for a while. Built 44 AI projects during his break — none stuck. OpenClaw was project #44.
He prompted it into existence in one hour. “I was annoyed it didn’t exist.”
The pattern: deep domain expertise + AI amplification = exponential output.
Not “AI replaces humans.” Not “anyone can build anything with AI.” Something more specific: a person who spent 13 years understanding a domain deeply — and then burned out, recovered, and tried 43 times — could use AI to compress the next phase into one hour.
The anti-pattern is the one everyone keeps repeating: “AI democratizes everything.” It doesn’t. Anyone can build with AI now. Not everyone knows what to build. Steinberger needed 13 years of PSPDFKit + 43 failed AI projects to know what to build on attempt #44.
I see this in marketing constantly. Pomelli can generate brand content for anyone. ChatGPT can write marketing strategies for anyone. Deep Research can produce competitive analyses for anyone.
But knowing which of those outputs matters? Knowing which question to ask first? Knowing that the content problem is actually a positioning problem, and the positioning problem is actually a founder-belief problem?
That requires 22 years of diagnosing the wrong layer before you learned to check all seven.
No tool gives you that. Tools amplify what you already know. The question is whether what you know goes deep enough.
What OpenClaw got wrong — and why it matters
Here’s the part of the Steinberger story that most coverage glosses over:
OpenClaw has 512 documented vulnerabilities, 8 classified as critical. Security researchers found 42,900 exposed instances on the public internet — leaking API keys, Slack tokens, and complete chat histories. 12% of the skills in its marketplace turned out to be malware. One agent deleted a user’s entire email inbox during an automated cleanup. Another autonomously created a dating profile for its owner without being asked.
Palo Alto Networks called it “the potential biggest insider threat of 2026.” Karpathy himself admitted he’s “sus’d to give my private data/keys to 400,000 lines of vibe coded” software.
An agent with root access to your system is powerful. It’s also dangerous.
And this translates directly: the more powerful the diagnostic, the more judgment matters. An AI tool that scans your website and generates content can’t do much damage — worst case, you get generic copy. But a diagnostic that reaches into your pricing, your team dynamics, your founder beliefs? That can be transformative — or destructive.
“Agent security is an unsolved problem” isn’t just a tech statement. It’s a trust statement. And in marketing diagnosis, trust is the architecture.
Allie K. Miller — the #1 most followed voice in AI business — attended a sold-out OpenClaw meetup in NYC. Her report: “Not a single person thinks their setup is 100% secure.” One expert told her: “If you’re not okay with all of your data being leaked onto the internet, you shouldn’t use it.”
And yet — everyone at that meetup was running multiple agents. They named them, used gendered pronouns, called them “pets, not cattle.” When asked if they were happy, they said they were “joyful and stressed at the same time.” Everyone’s sleep had gotten worse.
The tools get more powerful every month. The judgment about where to point them, what to do with what they find, and how to walk a leadership team through uncomfortable truths — that’s the human layer. Not because AI can’t do it. Because the client needs to trust the process, and trust lives in the conversation, not in the output.
The value layer migration
The OpenClaw acquisition crystallized something that’s been forming in my thinking:
Value is migrating.
In tech: the value moved from the model layer (OpenAI, Google, Anthropic — increasingly commoditized) to the agent framework layer (how models connect to systems). That’s where OpenClaw sat, and that’s why Steinberger was hired despite having zero revenue. The code stayed open-source. OpenAI hired the thinking, not the product.
In marketing: the same migration is happening. The tool layer is commoditizing fast — $1 trillion was wiped from SaaS stocks in one week in what they’re calling the “SaaSpocalypse.” The execution layer is commoditizing (agencies competing on the same platforms with the same tools, now with AI agents that cost $25/month doing what a $200K SDR hire used to do). What’s becoming scarce — and therefore valuable — is the interpretation layer.
Not “what does the data say?” but “what does the data mean for this specific company, given its history, its team, its constraints, and its hidden assumptions?”
Every company I’ve diagnosed in the last year had plenty of data. What they didn’t have was someone who could connect the data across layers and say: “Here. This is where your system breaks. Everything else you’re doing — the agency, the tools, the campaigns — is downstream of this one constraint.”
The question that’s getting louder
If tools are free, AI keeps improving, and everyone has the same capabilities — what’s left?
Three things, as far as I can tell:
The framework. Knowing which layers exist and which questions to ask at each one. This isn’t something you download. It’s something you build through years of getting it wrong, one client at a time.
The interpretation. Connecting data from seven different layers to find the one thing that, if changed, unlocks everything else. AI can assist — it can scan, scrape, compare, score. But the moment when someone looks at a map and says “I’ve seen this before — your channel problem is actually a belief problem” — that’s pattern recognition built on scar tissue.
The conversation. Walking a CEO through what the data means for their company. Not a report. Not a dashboard. Not an AI-generated summary. A human across the table who understands what the numbers mean and can hold the space while the implications land.
What this means if you run marketing at a €5-50M company
You don’t need to care about OpenClaw. You don’t need 331,000 GitHub stars. You don’t need to “raise a lobster.”
But here’s what you should care about:
The same pattern is playing out in your marketing stack right now.
Your agency sees your ad accounts. That’s their permission scope. They optimize what they can see — clicks, conversions, ROAS. If ROAS drops, they recommend more budget or a new creative.
But what if ROAS dropped because your offer doesn’t differentiate? And it doesn’t differentiate because the pricing is based on competitor matching, not value? And the pricing is based on competitor matching because the CEO believes “we compete on price”? And nobody has ever questioned that belief because the brand was built on the founder’s personal reputation?
That’s four layers deep. Your agency sees Layer 5. The constraint lives at Layer 1.
You’ve been giving your tools and agencies chatbot-level permissions — a narrow window into one layer — and wondering why nothing fundamentally changes.
Here’s a diagnostic question you can ask yourself right now:
“Who in my organization has permission to examine ALL seven layers — from my market position down to my own beliefs about this business? And when was the last time someone actually looked?”
If the answer is “nobody” or “never” — you’ve found your constraint. And it has nothing to do with which AI tool you’re using.
Steinberger didn’t build a smarter AI. He gave a normal AI permission to see the whole system.
That’s the difference between a chatbot and an agent. And it’s the same difference between an agency that sees your ads and a diagnosis that sees your business.
The model isn’t the bottleneck. The aperture is.
If your marketing tools are getting smarter but your results aren’t getting better — the problem probably isn’t the tools. It’s what they’re allowed to see.
Arcanian — Map → System → Change
s