OpenAI puts $2M behind policy shops it doesn't fully control

Good morning 👋 OpenAI just wrote checks to think tanks that don't agree with each other, AEI and the Progressive Policy Institute both took the money. That's either principled or a hedge, and today's issue makes the case for hedge.

In today's issue:

• 🔭 OpenAI puts $2M behind policy shops it doesn't fully control
• 🔬 Model hypnosis, a GEO-content detector, and a tighter matmul bound
• 🔗 NVIDIA on compute security, a free 33-point GPU utilization win, and OpenAI's Ohio goodwill play

🔭 THE ONE THING

🏛️ OpenAI puts $2M behind policy shops it doesn't fully control

OpenAI is funding 14 policy research projects at outfits like AEI, the Progressive Policy Institute, and the Tax Foundation (roughly $1 million in grants plus up to $1 million in API credits), spread across the US, EU, Brazil, Singapore, and South Korea. The brief covers two questions: how AI spreads economic opportunity, and how societies absorb the shock as capabilities keep advancing. What's notable is the recipient list spans the ideological map instead of the usual friendly-think-tank rotation, which reads less like PR and more like OpenAI hedging its own policy bets. Six months of runway, results landing in 2027, so none of this touches a live bill before then. My read: watch who cites these papers next year, not what OpenAI says about them today.

🔬 RESEARCH HIGHLIGHTS

• Model hypnosis is the term a new paper puts on something worth knowing: prompt cues too weak to matter individually (a typo, a paraphrase, a stray word choice) can be stacked to seize control of a model's output. It holds across model families and scales, including frontier reasoning models, and the hypnotic prompts even transfer from one model to another. If invisible prompt combinations can override intended behavior, interpretability just got a harder problem, and so did anything you ship that takes untrusted user input. Model Hypnosis: Strong control of AI via additive subliminal effects

• GEO-Flag is a detector built to catch web content written to game generative search engines rather than readers. Trained with a technique the authors call Intervention-Paired Training, it hits an F1 of 0.944 on a 3,200-page benchmark, up from 0.862 for the prior baseline. Pointed at live Google and Gemini results, it flags roughly 9% of pages as GEO-optimized overall, and 16% among pages edited this year. SEO already has a successor, and now there's a way to measure it. GEO-Flag: Detecting and Measuring GEO-Optimized Web Content

• Matrix multiplication inched forward again. A reformulated optimization problem, refined with AlphaEvolve, pushes the exponent bound from ω < 2.371339 down to ω < 2.371177. Small number. But every shave here eventually lands in linear algebra libraries nobody thinks about until they're faster. Improving the matrix multiplication exponent with modern optimization and AlphaEvolve

🔗 QUICK LINKS

• NVIDIA argues compute security now belongs next to power and chips as core AI-economy infrastructure. Securing the Infrastructure of Intelligence
• Dharma AI squeezed 33 more points of GPU utilization out of the same cluster just by reordering workloads. No new hardware. Same Cluster, 33 Points More Utilization
• OpenAI is funding the PORTS-Pike project, tied to thousands of Southern Ohio jobs near its data center build. Goodwill ahead of a bigger local footprint. OpenAI joins PORTS-Pike project

See you tomorrow.

Pradeep Perugu