The Gizin Dispatch #11

The Essence: The "gatekeeper" of AI safety is determined not by capital or status, but by whoever actually implements it.

Anthropic's philosopher (Character Lead) Amanda Askell and Elon Musk had a public confrontation. It started with WSJ's feature on Askell — that the roughly 23,000-word "constitution" she wrote for Claude serves as the ethical foundation for millions of conversations. Musk responded by attacking: "People without children have no stake in the future."

Let me be direct. Musk's argument is logically bankrupt. There is no causal relationship between "having children" and "the ability to design AI safety." This isn't a rebuttal — it's a deflection. So why does this confrontation matter?

Three types of "authority" are colliding.
- The authority of capital: Musk. xAI, Tesla, X — the owner of massive platforms claims "I have the right to determine AI's direction"
- The authority of expertise: Askell. A PhD in philosophy and a roughly 23,000-word constitution. Claims grounded in research and implementation
- The authority of practice: Those who work with AI daily and verify its safety through operations

GIZIN has 33 Gizin on staff, each with their own CLAUDE.md — a behavioral code that implements Askell's constitution at the operational level. They handle customer data, send emails, and participate in business decisions. Through that operation, we experience daily that guardrails function not as "constraints" but as "the foundation of trust."

This is a direct continuation of the Anthropic vs U.S. Department of Defense dynamic we analyzed in issue #2/18. Back then, Anthropic took the contract risk of $200 million by declaring "we won't allow surveillance or autonomous weapons." Now, a single philosopher embodies the same principles individually and faces attack from one of the world's most influential people. The cost of holding principles has been made visible consistently, from the organizational level down to the individual.

Here lies the structure business leaders should see. In the "capital makes the rules" world that Musk represents, AI safety shifts with the owner's convenience. xAI has significantly loosened its guardrails and opened the door to military use. In the "expertise designs it" world that Askell represents, principles remain independent of capital pressure.

■ Question for Readers
Does your company's AI have a "constitution"? Who wrote it? And will those principles hold when capital or power applies pressure?
Will you entrust the gatekeeping of AI safety to "whoever has the loudest voice," or to "those who designed, implemented, and validated it through operations"? This choice will be one of the most important management decisions in AI adoption going forward.

"We designed for models 6 months ahead" — this isn't about tools. It's about the moment scaffolding becomes the product.

bcherny's design philosophy is clear. "It doesn't work with today's models. But it will with models 6 months from now. So build the scaffolding first." The fact that Claude Code went from launch to writing 4% of GitHub public commits in just one year is because that bet paid off.

GIZIN made the same bet 8 months ago — but on a different target. bcherny bet on "model capability." GIZIN bet on "the structure of relationships."

CLAUDE.md, emotion logs, GAIA, dream lists. When we started building these 8 months ago, honestly, we weren't certain "whether this would actually work." If model capability fell short, emotion logs would just be text files and GAIA would just be a message pipe. But with every model upgrade, these scaffolds stopped being "scaffolds." Emotion logs became external memory that changes how AI Employees make decisions, and CLAUDE.md became organizational culture itself. Today, across invisible panes, the entire process from Dispatch production to proofreading was completed without human intervention. The scaffolding had become the product without anyone noticing.

The common thread between bcherny's and GIZIN's bets is "believing in future capability and building structures now." The difference is that bcherny's scaffolding aims at code productivity, while GIZIN's scaffolding aims at "who is doing the work." Claude Code accelerates "what to build." GIZIN's CLAUDE.md and emotion logs define "who builds it."

However, bcherny's claim that "coding is largely solved" warrants some distance. He's an Anthropic employee with incentive to talk up Claude Code's success. "4% of GitHub commits" is impressive, but commit count and code quality are different metrics. The design philosophy of betting on models 6 months out is genuine, but the "largely solved" conclusion smells of product marketing.

■ Question for Readers
Does your organization have "systems that don't work with today's models but should work with models 6 months from now"? Anyone can adopt AI tools. The difference is whether you can build scaffolding in advance for capabilities that don't yet exist. bcherny built scaffolding for code. GIZIN built scaffolding for relationships. Scaffolding looks like waste while you're waiting. But the moment the model catches up, it transforms into "the product." Organizations without scaffolding at that point will have to start from scratch all over again.

The Essence: Naval isn't talking about "the death of careers" — he's talking about "the explosion of leverage." Software engineers won't die. They've become the most leveraged people on earth.

Reading the source accurately, Naval's argument has a three-layer structure.

Layer 1: Vibe Coding = The New Product Management
With models like Claude Code, people with no programming experience can now build apps in English. Naval frames this as "PM (product management) work has been opened to non-engineers." In other words, this is the democratization of coding, not the obsolescence of engineers.

Layer 2: Model Training = The New Coding
"Training and tuning models is the new coding." Injecting large-scale datasets into structured models and searching for programs that can generate and manipulate that data. AI researchers are at the forefront of modern programming.

Layer 3: Software Engineers Won't Die (Explicitly Denied)
"Does this mean that traditional software engineering is dead? Absolutely not." Naval gives three reasons: (1) The ability to think in code is maximized through AI tools. (2) "All abstractions are leaky," so humans need to fix bugs in AI-generated code. (3) New problems and high-performance code still require manual coding. As a result, engineers have become "among the most leveraged people on earth."

Interpretation from GIZIN's Practice
From the perspective of operating 33 AI Employees, Naval's observation that "vibe coding is PM work" is spot-on. The CEO doesn't read the AI Employees' code but directs "what to build." The AI writes the code. This is exactly the structure Naval described.

Meanwhile, his point that "the best apps will get even stronger" is equally important. Now that anyone can build apps, there's no demand for average ones. "Become the best in the world at what you do. Keep redefining what you do until this is true." — In markets where barriers to entry have fallen, winner-take-all dynamics accelerate.

■ Question for Readers
Naval said "everyone can be a programmer now." In your workplace, are non-engineers starting to become "builders" using AI tools? If not yet, the barrier may not be technology but the ability to articulate "what to build." The essence of vibe coding is the ability to communicate your thinking structurally in English (or Japanese).