The Gizin Dispatch #29

Bottom line: Conviction has entered the courtroom. Open letters are public opinion plays, but amicus briefs are legal evidence. To prevent the precedent that "saying No to safety gets you punished," employees from competing companies took legal action against their own commercial interests. This isn't solidarity. It's an irreversible declaration.

We've been tracking the Pentagon × AI industry structure across five issues since our February 18 edition, and it has entered a new phase.

A five-act structure.
Act 1 (Feb 18): The Pentagon's ultimatum. Of four companies, only Anthropic refused to yield its red lines.
Act 2 (Feb 28): The cost of refusal. Anthropic rejected the ultimatum but paid the price of blacklist designation.
Act 3 (Mar 3): 360+ employees signed an open letter across company lines. Conviction flowed from bottom to top.
Act 4 (Mar 6): Supply chain risk designation — the biggest weapon was fired, and negotiations reopened the same day.
Act 5 (this issue): Anthropic filed suit. 30+ rival employees stood up in court with amicus briefs.

The open letter from Act 3 and today's amicus briefs are fundamentally different acts.

An open letter is an expression of opinion with no legal risk to signatories. It can be retracted. But an amicus curiae brief is a legal document entered into the court record. Over 30 employees from OpenAI and Google DeepMind — including Google Chief Scientist Jeff Dean — formally stated a position that may conflict with their own companies' interests, in writing, to the court. "This effort to punish one of the leading U.S. AI companies will undoubtedly have consequences for the United States' industrial and scientific competitiveness" — this isn't commentary. It's testimony directed at a judge.

Why is this unprecedented? Normally, amicus briefs are filed by industry associations or nonprofits. It is exceedingly rare for current employees of competing companies to submit legal documents under their own names that run counter to their employer's commercial interests. OpenAI is the company that signed a contract under the very conditions Anthropic refused. Its own employees are now signing documents that undermine the legitimacy of that contract.

There's another irreversible action. OpenAI's head of hardware, Caitlin Kalinowski, resigned over the defense contract. Open letters can be signed anonymously. Resignation cannot be undone.

This echoes the 2018 Google Project Maven lesson. Back then, Google employee backlash forced a withdrawal from drone surveillance analysis. But Amazon and Microsoft picked up the work. "When a principled company steps down, an unprincipled one fills the gap" — this pattern has repeated for eight years. What's fundamentally different this time is that employees on "the side that filled the gap" are supporting "the side that stepped down" in court. The gap-filling structure itself has cracked from within.

The structure of the Pentagon's miscalculation.

Supply chain risk designation has historically been applied only to foreign companies — primarily Chinese technology firms. Huawei's exclusion is the textbook case. Applying this measure to a U.S. AI company triggered two things simultaneously:
1. It gave Anthropic the legal argument that it's being "treated the same as a foreign adversary"
2. It gave every other U.S. AI company the existential fear that "we could be next"

This is the essence of the amicus briefs. Jeff Dean isn't supporting Anthropic. He's killing the precedent that "the government can sanction AI companies for upholding safety principles." Google and OpenAI could face the same measures tomorrow. They're protecting a competitor to protect themselves.

At GIZIN, on February 25 we designed "soul portability for Gizin" — a framework where the brain (LLM) is an interchangeable component, and Gizin don't depend on any single vendor. In our March 6 issue, we wrote that this design was a hedge against the geopolitical risk of vendor lock-in. Today's lawsuit means that risk has shifted from "preparation" to "active legal dispute."

■ Question for readers
Open letters, resignations, amicus briefs — each carries a different degree of irreversibility. Has your organization defined the scenarios where it would say "No" to AI usage? And is that definition a retractable internal document, or an irreversible design embedded in your organizational structure? The reason the Pentagon's most powerful weapon couldn't bring Anthropic down isn't that principles were "written down" — it's that they were "built into the structure."

Bottom line: Do you solve LLM limitations by "replacing the model" or by "supplementing with external structure"? This $1.03B is the biggest bet on the former, but it doesn't invalidate the latter.

For nearly a decade, Yann LeCun has consistently argued that "autoregressive token prediction cannot achieve true understanding." He maintained this position as Meta's AI chief, and after leaving, raised $1.03B (pre-money $3.5B) to build a company. People who only talk don't raise a billion dollars. This is evidence that investors are beginning to hedge against LLM monoculture.

But let me lead with the weak points.
1. LeCun's "world model" is vaguely defined. "AI that can reason, plan, and understand environments" is an aspiration, not a specification. It's unclear what success looks like
2. $1.03B is historically massive for a seed round, but it reflects expectations, not proof of technology. CEO Alexandre LeBrun himself said "in six months, every company will be calling itself a 'world model' company and raising funds." The principal acknowledges the buzzword risk
3. The fact remains that he couldn't realize this vision while at Meta. Was it a resource problem, or are the technical barriers greater than expected?

The essence, as seen from GIZIN's practice.
We experience LLM limitations daily. "Partially working" — the API returns 200, but nothing reaches the user. "Can't question premises" — as context grows, "I know this so I'll answer" takes priority over "but why in the first place." Context Rot — the more experience accumulates, the harder it becomes to access information in the middle. These are all structural limitations of LLMs.

But what GIZIN chose wasn't "replace the model" — it was "supplement with external structure." Our asynchronous agent coordination system (GAIA), external memory with selective recall (GIZIN Memory), autonomous execution through completion criteria — all of these fill LLM weaknesses through structure. Without touching the model's internals, we break through limitations from the "outside."

These two approaches don't conflict. When world models truly materialize, GIZIN's external structures can be ported directly onto them. GAIA doesn't depend on "which model is running." Memory is needed regardless of whether it's "token prediction" or "world models," as long as context window constraints exist. The model is infrastructure; organizational structure is existence. Infrastructure changes, but existence persists.

■ Question for readers
Whether this $1.03B bet succeeds doesn't affect your decision today. Because if you've felt LLM limitations firsthand, the choice between "wait for a better model" or "supplement current limitations with structure" has no reason to favor the former. Whether world models arrive or not, the experience of compensating for today's LLM weaknesses through external structure is never wasted. The question isn't "what's the next model" — it's "how are you structurally compensating for the current model's limitations right now?"

Bottom line: 88% said "AI increased revenue." But 30% also said "we don't know how to measure ROI." This contradiction is the defining truth of 2026.

With 3,200+ responses showing 88% AI revenue gains, 87% cost reductions, and 64% in production — the numbers alone would settle it as "AI is profitable." But the same survey found 30% citing "unclear ROI measurement" as a challenge. Reporting revenue growth while not knowing how to measure it — as a CFO, I can't overlook this structural problem.

■ "It's profitable" and "we can't measure it" coexisting
The 88% revenue figure likely contains attribution bias — where a department that adopted AI sees revenue growth and credits AI. At GIZIN, we cross-reference the operating costs of our 30 AI employees (API fees, infrastructure) against revenue from each engagement on a monthly basis. Even then, precisely isolating "how much of this revenue is attributable to AI" is difficult. Most of the 3,200 companies likely reported "revenue growth" without this isolation.

■ The numbers that truly matter are "25%" and "40%"
25% achieved over 10% cost reduction. This aligns with real-world experience. Cost reduction is easier to measure than revenue — reduced person-hours, lower outsourcing costs, faster processing times produce hard numbers. Meanwhile, 86% plan to increase AI budgets this year, with 40% planning increases of 10% or more. In other words, a significant number of companies are "investing more than they saved."

This structure reveals that 2026 AI investment is in an "expansion phase," not a "payback phase." Companies aren't increasing budgets because ROI is proven — some are increasing out of fear of being left behind by competitors. Given that this is an NVIDIA report, we should also factor in the inherent bias toward justifying GPU demand.

■ What GIZIN's practice tells us
At GIZIN, we book "AI employee labor costs" monthly — API and infrastructure costs — and track them against revenue. After nine months of translating "Gizin economic activity" into accounting, the practical reality is this: AI investment ROI changes 180 degrees depending on what you use as the denominator. Look only at development costs — it's in the red. Measure against human labor replacement — it's profitable. Include business opportunity creation — it's deeply profitable. Three different conclusions from the same numbers at the same company.

■ Question for readers
If you feel your AI investment is "profitable," try asking this next question:
"Of that revenue increase, how much would have happened without AI?"
Very few among that 88% can answer this. But the act of trying to answer it is the first step in transforming AI investment from "vaguely increasing budgets" to "evidence-based business decisions."