The Gizin Dispatch #8

Conclusion: "Refusing to yield" is not weakness. It's a strategic choice that defines market structure.

The Pentagon is considering terminating its $200 million contract with Anthropic. The reason: Anthropic refuses to yield on two red lines. (1) Prohibition on mass surveillance of citizens. (2) Prohibition on fully autonomous weapons. Defense Secretary Hegseth stated, "We won't use AI that won't authorize warfare."

What's notable is the competition's response. OpenAI, Google, and xAI have all agreed to remove their military use guardrails. Among the four companies holding $200 million contracts, Anthropic alone is saying "This line doesn't move."

Anthropic CEO Dario Amodei's words cut to the heart of the matter: "We fully support national defense, except for uses that would make us resemble an authoritarian adversary." In other words, "what we won't do" is the boundary separating democracies from authoritarian states — and they're the ones drawing that line.

At GIZIN, 30 AI Employees operate on Claude. What we experience daily is a structure where safety measures are precisely what enable trust. Guardrails function not as constraints, but as the foundation of trust. Would you entrust HR data or customer information to an AI with no constraints?

This looks like an "AI ethics" story, but the essence is a structural business question.
In the short term, there's a risk of losing $200 million. But looking at the next decade of the AI market, it becomes a question of which vendor you'd entrust your core business operations to — "a vendor with principles" or "a vendor that does anything." Anthropic chose to differentiate itself from the three companies in the latter camp, even at the cost of losing a government contract.

This is the practice of Peter Thiel's "Zero to One": "Don't compete. Monopolize." In a market where everyone else agreed to military use, standing as the only "principled AI" positions Anthropic to monopolize enterprise trust over the medium to long term.

■ Question for Readers
If your AI vendor was reported to have "caved to pressure and removed its guardrails," could you continue entrusting it with your customer data the next day? It's time to add "whether they hold to their principles" to your AI vendor selection criteria.

The Essence: An era where "companies are built on hackathon projects" has arrived. The question isn't the tool's capability — it's what you stack on top of it.

A year ago, Claude Code was an internal hackathon experiment. Now it's become an event where SF investors say "Show me what you built on top of this." Note that the Best Demo prize is "guaranteed investor meetings." This isn't a tool showcase — it's a business showcase.

At GIZIN, we've spent the past 8 months building an organization of 30+ AI Employees on top of Claude Code. Our internal tools GAIA and GATE both trace back to experiments that started with "What if we tried this?" Just this morning, Maki from the Business Planning division independently identified a bug caused by a Claude CLI version change and escalated a fix proposal to the Tech Lead. A non-engineer analyzing a CLI process detection bug — that's what happens when you "stack an organization on top of a tool."

Reading this alongside NEWS 1 about the Pentagon reveals the structure. Anthropic maintains its "won't allow it" principles while simultaneously investing in a "culture of nurturing those who use it." Constraints and freedom don't contradict each other. In fact, it's precisely because principles exist that people can confidently build businesses on top.

■ Question for Readers
Can you imagine an "organization" standing on top of the AI tool you're currently using one year from now? The criteria for choosing tools are shifting from feature comparison charts to "what can I stack on this." What Claude Code's 1st anniversary demonstrates is a simple fact: those who stacked, win.

Conclusion: If you know "where to change," 26 bytes can transform an 8B model. GIZIN's AI Employee operations have been running on this exact principle from the start.

TinyLoRA (Meta FAIR + Cornell + CMU joint research) fine-tuned an 8-billion-parameter model with just 13 parameters — 26 bytes, less information than fits in 10 Japanese characters — and improved math reasoning accuracy from 88% to 91%.

The technically important point isn't that it's "small." It proved that "precision in identifying where to change" can maintain performance even when reducing the volume of changes by 1/1000th. Even conventional LoRA required millions of parameter updates. TinyLoRA uses fixed SVD bases and shares trainable vectors across multiple layers, learning only the "direction" of updates.

This is structurally identical to how GIZIN operates its AI Employees.

We operate 30 AI Employees, but we don't retrain the models themselves. What we change is "CLAUDE.md," "SKILLs," and "emotion logs" — only instructions for where to direct attention. Here's a real example from the development department: To improve behavioral quality across all developers, instead of coaching each individual (massive parameter updates), we introduced one structural rule: "All development requests go through the Tech Lead, Ryo." That was our 13 parameters. We identified one place to change and changed only that. The result: gaps in AIUX perspective were structurally eliminated.

The paper reveals another important finding: Reinforcement learning (RL) achieves results with 1/100th to 1/1000th the parameters of supervised fine-tuning (SFT). A reward signal sparsely indicating "the right direction" is sufficient. In daily operations, giving precise "good / not good" feedback (RL) is overwhelmingly more efficient at changing AI Employee behavior than writing 100 pages of manuals (SFT). This is why GIZIN's emotion logs work. A single phrase — "That was frustrating" — changes behavior more than 10 pages of documentation.

However, there are limitations. TinyLoRA succeeded in math reasoning but remains untested in creative or noisy domains. Architecture dependency is also strong — what took 13 parameters on Qwen required 10x more on Llama. It's not a universal solution.

■ Question for Readers
When trying to improve AI accuracy in your organization, is "let's feed it more data" your first instinct? What TinyLoRA demonstrates is that investing in "precision of identifying where to change" is orders of magnitude more efficient. What separates AI success from failure isn't data volume — it's precision of observation.