Finance sizes up the AI machine: 88.6% judgment match, Griffin's GPU warning, Schmidt's cycle

The 88.6% claim

Terminal X, a company building an agentic operating system for financial firms, says its platform matches senior investor judgment 88.6% of the time. The system institutionalizes firm memory — the memos, mandates, and processes that usually live in people's heads — using OpenAI's GPT-6.1 Sol to power an Eval Agent that automates due diligence and investment memo workflows, reducing what the company calls memory debt. m:…8210

The measurement apparatus is the point. Terminal X says its Eval Agent "grades outputs against a 54-criterion rubric", catching 25% more deviations from a fund's mandate than the next-best model, with agent agreement to human reviewers 220 basis points higher on GPT-6.1 Sol. The company also claims 30% lower cost per query than the next-best model — and cites one customer with $2.4 million in annualized savings on due diligence and investment workflows. m:…8210

Deployment detail matters to this audience: Terminal X says its agents run on-premise or in the customer's own cloud, and that GPT-6.1 Sol handles the judgment-heavy stages — synthesizing the investment thesis, identifying key considerations, assessing fit with the firm's mandate. Engineers spend weeks onsite reverse-engineering trusted memos and capturing processes that were never written down, the company says. m:…8210

The company also says its engineers are former finance and investment professionals from JPMorgan and Blackstone — a deliberate hiring signal for an audience that buys judgment, not just software. m:…8210

Griffin's pricing warning

Ken Griffin, CEO of Citadel, told Bloomberg that AI models now complete in hours the high-level white-collar work that used to take his PhDs and masters-level employees four to six weeks — a line that doubles as a hiring thesis and a reskilling alarm. He called for a profound national reskilling effort for workers in what he described as AI's capability sweet spot, while arguing startups are positioned to give people without perfect pedigrees new career lanes. m:…7258

On markets, Griffin argued the reset is driven by investors moving from broad thematic AI buying to discerning analysis of supply-chain choke points and economic rent capture. DRAM and MLCC makers, he noted, earn extremely high margins on real supply constraints — while GPU pricing is vulnerable to a free-fall if provisioning overshoots demand. His analogy: GPUs have very low marginal cost to run, like airline seats, so overcapacity crushes price. m:…7258

Corporate buyers, in Griffin's telling, are already price-discriminating — choosing between high-priced state-of-the-art models from OpenAI and Anthropic and lower-cost open-source alternatives. And he offered a hedge against the automation story: the vast majority of alpha generated at Citadel still comes from human creativity, ingenuity, and fundamental research, not AI. m:…7258

The frontier cycle

Eric Schmidt, the former Google CEO, told the In Good Company podcast that frontier models now iterate fast: roughly six months to build and about two months to test, in his telling. The cycle time is the story — a machine that rebuilds its frontier roughly twice a year. m:…3602

For investors, the three threads read as one machine: models iterating on a roughly six-month cadence, vendors selling judgment-matching agents into finance at claimed 30% lower query cost, and market veterans pricing the hardware underneath — with Griffin warning that the seat being sold has near-zero marginal cost. m:…8210 m:…7258 m:…3602