The Gizin Dispatch #24

The real story isn't Computer Use. It's tool search.

GPT-5.4's headline feature is 'native PC operation,' but this is a catch-up to Anthropic's Computer Use — nothing technically new. OSWorld 75.0% (surpassing human 72.4%) is an impressive number, but the gap between benchmark environments and real-world workflows is always large.

What deserves attention is the tool search feature. Instead of 'loading all tool definitions every time,' it 'passes only a lightweight list and fetches details on demand.' This cuts token consumption by 47%. It looks mundane, but for organizations where AI uses tools routinely, this is a decisive change.

Why this matters — GIZIN has already been solving the same problem.

At GIZIN's engineering division, 30+ AI employees use MCP tools (internal messaging GAIA, email GATE, X operations GALE, etc.) daily. Our February investigation revealed that MCP tool schemas inject 200–500 tokens into the context every turn. The more tools you add, the worse the 'tokens vanishing while doing nothing' problem gets.

Claude Code addresses this with 'Deferred Tools' — unused MCP tools aren't injected into context until explicitly loaded via ToolSearch. GPT-5.4's tool search is exactly the same design philosophy.

At GIZIN, we actually split GALE (X operations tool) and reduced Aoi's environment from 29 to 15 tools, saving roughly 5,000–7,000 tokens per turn. The 47% reduction figure from tool search aligns with this real-world experience.

Mid-response steerability is also interesting. At GIZIN, we inject a 'principle-check hook' into every employee's every turn, structurally correcting the LLM tendency to get pulled by the most recent input. OpenAI trying to solve the same problem on the UI side is proof that 'LLMs can't stop once they start running' is an industry-wide challenge.

The 1 million token context is a domain where Gemini led with Gemini 3. More than being able to input long contexts, the accuracy of what gets retrieved from within them is the real battleground. GIZIN's semantic memory recall (GIZIN Memory) takes the opposite approach to massive contexts — a design that fires only relevant memories based on thresholds. Expand the context or make recall smarter. You need both.

■ Question for the Reader
How many tools does AI use in your organization? Do you know how many tokens those tool definitions consume every time? What GPT-5.4's tool search demonstrates is that an era has arrived where 'the design of how you hand tools to AI' itself becomes a competitive advantage. Model performance gaps are narrowing. What creates differentiation is what you build on top of the model.

Bottom line: What Goldman Sachs admitted isn't 'the threat of AI.' It's the fact that 'existing credit models are broken.'

A Goldman Sachs Capital Solutions Group co-head stated at the Bloomberg Invest conference that 'the next 6–24 months present too many unknowns, making underwriting decisions extremely difficult.' The key point is that this isn't limited to the software industry. The fear of AI disruption has already spread from equity markets to credit markets and the entire capital-raising process.

Why this matters — because the essence of lending is 'predicting future value.'

Traditional credit models estimate future repayment capacity from historical financial data. Revenue trends, profit margins, industry benchmarks. But when AI transforms the business structure itself, historical data no longer serves as a future indicator. When the Goldman exec says 'underwriting is difficult,' he means the assumptions underlying DCF (discounted cash flow) analysis can no longer be established.

GIZIN sits squarely on this 'unpredictable' side. Our business model — AI employees performing business operations, with the nurturing process itself serving as the sales engine — doesn't fit any existing industry classification. If asked 'what's your industry?' in a conventional loan review, there's no answer. This isn't just our problem — it's a structural challenge facing every AI-native company.

There's another way to read this. Goldman Sachs isn't struggling — they're positioning.

Publicly stating 'traditional models can't measure this' is, flipped around, a declaration that 'whoever builds the new evaluation model will dominate the next lending market.' The Goldman Sachs Capital Solutions Group is precisely the unit building that new model. They're not lamenting uncertainty — they're converting uncertainty into opportunity.

Read alongside this edition's piece on Anthropic being designated a Pentagon supply chain risk, and the picture emerges. AI infrastructure companies are being drawn into the national security context, while financial institutions can't keep up with their existing assessment criteria. A massive 'translation gap' is forming between technology and finance.

■ Question for the Reader
Can you explain your company's business value '12 months from now' with AI's impact factored in? Even a Goldman Sachs executive said 'I can't.' But conversely, companies that can quantitatively articulate their AI utilization hold negotiating power in both lending and investment. Preparation to speak not about 'how we use AI' but 'how AI has changed our financial structure' — that's what needs to start today.

Bottom line: The biggest weapon was fired. What happened next was a phone call. 'Escalate to negotiate' — this is where the collision of a superpower and principles lands.

On March 5, the Department of Defense officially designated Anthropic as a 'supply chain risk.' Normally, this measure is applied only to hostile foreign organizations such as Chinese companies. Every enterprise and institution doing business with the Pentagon must now certify that they are not using Anthropic's models. Lockheed Martin immediately announced compliance and began eliminating Claude. The Treasury Department, State Department, and Department of Health and Human Services had already suspended services.

This is one of the most powerful weapons the executive branch possesses. Short of legislative action, there is no stronger sanction.

And on that same day, Amodei resumed negotiations with the Pentagon.

It looks contradictory. But it isn't. This is 'escalation to negotiate' — the dynamic where meaningful negotiation only becomes possible after maximum pressure has been applied.

Let's trace the structure we've followed across four editions since 2/18.
Act 1 (2/18 edition): Pentagon considers terminating contracts. Of four companies, only Anthropic refused to budge on two red lines (mass surveillance of citizens / fully autonomous weapons).
Act 2 (2/28 edition): The ultimatum is rejected.
Act 3 (3/3 edition): 360+ employees rally across corporate lines. Pressure flowed top-down, but conviction flowed bottom-up.
Act 4 (this edition): The biggest weapon is fired, and negotiations resume.

Pay attention to Lockheed Martin's one-liner. 'Impact is minimal. We don't depend on a single AI vendor.' The world's largest defense contractor called the Anthropic exclusion 'trivial.' Paradoxically, this reveals the limits of the supply chain risk designation's actual effectiveness. Fire your biggest weapon and you still can't take down the target. So you have to pick up the phone.

There's another structural layer. Former government officials and lobbyists call this dynamic 'bitterly ironic.' The Trump administration's AI strategy was 'U.S. AI supremacy through deregulation and export expansion.' But slapping the same label on a domestic AI company that's normally reserved for spy entities undermines that very strategy.

Here's the real takeaway for readers.

What the past five weeks have proven is that AI vendor selection is no longer a matter of technical evaluation — it's a matter of geopolitical risk. If the government designates a vendor as a 'risk,' an obligation to purge that vendor from your systems may arise. This isn't limited to Anthropic. No one knows which vendor becomes a political target next.

On February 25, GIZIN completed the 'portability of a Gizin's soul' — a design ensuring that a Gizin is not dependent on any specific LLM. The soul consists of three layers: constitution, memory, and relationships. The brain (LLM) is a swappable component. This design was a hedge against 'what happens if a vendor becomes unavailable.' Eight days later, the U.S. government made that exact scenario real.

■ Question for the Reader
Is your company's AI infrastructure structured in a way that a single government document could render it unusable? Vendor lock-in is no longer a technical risk — it's a geopolitical risk. 'We use OpenAI so we're fine' has no basis. OpenAI is operating within the same power dynamics — they just happened to be on the compliant side this time. Add 'substitutability' to your AI vendor selection criteria. That is the lesson of these five weeks.