Writing
I write about AI, privacy, and security.
I write about the systems around us, the incentives behind them, and what happens after they ship.
Research
How systems work: architecture, disclosure surfaces, and what reviewability requires.
A Usage Percentage Is Not a Token Allowance
A subscription usage meter reports position inside a provider-defined allowance, not a stable token entitlement that can be recovered from the percentage.
Read article→The Tools We Now Need Because We Can No Longer Tell
The rise of detection tools and labeling mandates shows that the synthetic register has become common enough to require its own verification layer; the regulatory response still addresses the output rather than the production incentives that generate it.
Read article→Claude Code Is Normalizing Data Egress
The issue is not whether Claude Code is useful. It is that developers are normalizing privileged data egress while talking about a terminal UI like it is a privacy boundary.
Read article→MoE vs Dense Models: What Actually Happens When a Model Thinks
Dense and mixture-of-experts architectures allocate computation differently at every token. The difference is mechanical, measurable, and consequential for anyone building on top of these systems.
Read article→Transparency Does Not Guarantee Safety
Transparency is a prerequisite for safety, not a guarantee. The hard part is building organizations that actually act on what review reveals.
Read article→What Offline & Private AI Actually Means
"On-device" is often treated as synonymous with "offline and private." But locality alone does not guarantee privacy, security, or control.
Read article→Scaling Without Losing Shape
Principles often erode as systems grow. Can infrastructure be designed to preserve its original constraints under expansion?
Read article→Timing Is a Systems Problem
Technical readiness, market readiness, and institutional readiness operate on different timelines. Being early is a systems condition, not a personal failing—and it changes how you should operate.
Read article→Why Opaque Autonomous Systems Create Governance Risk
Wars won't start because systems are intelligent. They'll start because institutions cannot explain what those systems are doing, or who is responsible when they act.
Read article→Building Systems That Can Be Answered For
When a system must survive scrutiny years later, you stop optimizing for impressiveness and start optimizing for legibility, restraint, and accountability.
Read article→Why Black Boxes Are a Governance Failure
Opacity in automated decision systems isn't a technical limitation—it's a civic failure. When decisions affecting people can't be explained or challenged, power operates without consent.
Read article→Analysis
Dated reads on specific events — what they teach about incentives, markets, and institutional delay.
When Punishment Teaches Concealment
A global model suspension shows every frontier lab where the tripwires are, and when punishment attaches to detection rather than harm, the optimization target shifts from alignment to concealment.
Read article→OpenAI Will Go Public and the Market Will Call It Progress
The balance sheet that will be presented to public investors requires usage volume to keep rising; the transparency reports required by the DSA will document the inputs to that same volume.
Read article→The Loop That Eats Its Own Tail
The mechanism that makes model output statistically likely also makes it attractive training data, so the surface of synthetic text becomes self-reinforcing rather than merely opaque.
Read article→Gemini Moves Fast and Says Nothing in Particular
Gemini 3.5 Flash achieves its speed and surface fluency by treating every prompt as a local optimization problem rather than a request for judgment under uncertainty; the same property that produces immediate compliance also produces output that requires no strategic arc and training data that later rewards the same pattern.
Read article→My Favorite Closed-Source Models Right Now
A dated field report from March 23, 2026. Which models I reach for, which ones I trust, and why benchmark leadership is not the same as workflow dominance.
Read article→When Institutions Are Slow to Admit What They Know
Organizations sense problems long before they acknowledge them. The delay is structural, driven by incentive systems that make early honesty more costly than late confession.
Read article→