Why Moving Fast With AI Often Means Moving Backward

Welcome to the first edition of The Cybernetic CxO. A monthly newsletter for CTOs, CIOs, and digital transformation leaders navigating the AI era.

Each month, I’ll share patterns I’m seeing in the field: what’s working, what’s breaking, and what it means for how we build and run technology organizations. Expect honest takes, practical frameworks, and one thing you can do this week.

This month: Why moving fast with AI often means moving backward and what to do about it.

Prototype Speed Is Not Production Sustainability

#

Everyone’s excited about AI making development faster. But here’s what I’m seeing in the field: teams are shipping code nobody understands, and calling it progress.

This month, one LinkedIn post of mine went viral (122K impressions) about Linus Torvalds “vibe coding” in a hobby repo. The reaction surprised me. What resonated wasn’t the Linus angle. It was the uncomfortable truth underneath:

AI gets you to 90% faster. The last 10% is where you spend the next 34 years.

What’s Actually Happening

#

The pattern I see across enterprises is consistent:

• Day 1: AI-generated prototype works. Everyone celebrates.
• Day 100: No one understands the code. Every change feels risky.
• Day 365: Incidents start. Not from crashes, but from uncertainty.

This is the real cost of “vibe coding” at scale. It’s not that AI writes bad code. It’s that AI-generated ambiguity compounds: unclear intent, unstated assumptions, missing ownership. The code works. The tests pass. But the organization can’t maintain it.

The same pattern appears in technical debt. Loveholidays shared how they use static analysis as an early-warning system. What caught my attention: they built sensors that create signals that trigger actions. Hotspots. Change coupling. Knowledge distribution. Bus factor. These aren’t vanity metrics. They’re feedback loops that make debt observable before it’s painful.

AI Agents Are Starting to Eat SaaS

#

The “build vs. buy” equation is shifting fast. Martin Alderson’s analysis captures what I’m seeing: engineers are building internal dashboards, tools, and automations in minutes, skipping freemium tiers, avoiding API headaches, and ditching annual renewal debates.

Teams are asking: “Do we really need to pay for this SaaS… or can we just build it ourselves?”

But here’s what most teams underestimate: SaaS doesn’t just ship software. It absorbs operational judgment: uptime, compliance, security patches, edge cases, and on-call pain. The moment you replace SaaS with an AI-built system, all of that comes back inside your company.

Not all SaaS is doomed. Products with high-uptime requirements (payments), network effects (collaboration tools), or proprietary data still have moats. But if your product is “just a SQL wrapper with a nice UI,” you’re now competing with thousands of devs and their AI copilots.

The rule that holds: AI makes replacement possible. Ownership makes it sustainable.

If you don’t have clear long-term system ownership, not just building it, but maintaining it when it’s boring and business-critical, SaaS is still cheaper. Even now.

Europe’s AI Moment: From “Best Model” to “Trusted Infrastructure”

#

This month I was quoted in the NZZ on a shift I’m seeing across boardrooms: the conversation has moved from “Who has the strongest model?” to “Who can we trust with our data, IP, and regulatory exposure?”

For many use cases, open-source models are already “good enough.” The differentiator now is governance, controllability, and traceability in real enterprise environments.

The winning strategy for European companies isn’t “train the next frontier model at any cost.” It’s:

• Build an AI-capable platform your business can rely on
• Keep sensitive data and trade secrets protected
• Make AI outcomes explainable and auditable
• Treat AI like infrastructure, not a demo

We’re entering the phase where AI becomes normal IT. The winners will be those who provide trusted, compliant, production-grade AI at scale.

Don’t start with the model. Start with the operating model.

What If Your AI Re-Wired Itself Mid-Conversation?

#

Most LLMs today work like this: train once, freeze during deployment, update only when researchers retrain them later. Even if they feel adaptive, their core weights aren’t changing while you chat.

Google Research just introduced a different direction: systems that pair short-term memory (attention) with a long-term neural memory module that learns while you use it, guided by a “surprise” signal.

• If input is expected → minimal update
• If input is surprising → stronger update
• Includes forgetting to prevent memory overload

This is the shift from static inference to a closed-loop adaptive system. Surprise acts like an error signal, updates behave like a controller, and forgetting looks like homeostasis.

The prize is adaptability. The risk is drift and runaway feedback. The central question becomes: how do we balance plasticity (learning) with stability (control)?

For CxOs watching the AI space: this is early research, but it signals where things are heading: systems that don’t just respond, but adapt. The governance implications are significant.

My Current AI Stack

#

• Claude Code: CLI, coding & agentic work. Best coder, becoming my main tool for everything now.
• ChatGPT: Chat and drafting. Fails on many tasks now, getting replaced by Claude Code.
• Perplexity: Web research. Sources & citations that exist.
• NotebookLM: Work with your documents. Fast understanding.
• Gemini: Images. Best and fastest for images.

Vibe Coded This Month (Private Projects)

#

Yes, I vibe code too. Here’s what I built this month:

• Website migration: Moved romanoroth.com and cyberneticenterprise.com from Wix to Hugo (self-hosted). Saving a lot of money. Did someone say SaaS is dying?
• Filament Management: A tool to track my 3D printing filament inventory
• URL-To-Obsidian Converter: Quickly saves web articles to my Obsidian vault