Frontier
Field Notes
Essays and experiments at the frontier of AI and complex systems. From the circuits inside language models to the evidence behind blockchain forensics: how things work, what we can know, and where our assumptions break.
Archive
All writing
I Found the Circuit Behind In-Context Learning. Then It Refused to Die.
Seven mechanistic interpretability experiments across four language models — and what a self-repairing circuit taught me about the hardest problem in AI safety.
The Next Two Halvings
A Falsifiable Forecast for Bitcoin’s Security Through 2036
The Block Space Market
Are Bitcoin's new tenants paying enough rent to keep the building standing?
Are Bitcoin's new tenants paying enough rent to keep the building standing? Examining whether inscription and token demand can actually solve the security budget problem.
The New Tenants
How Ordinals, BRC-20, and Runes turned Bitcoin into contested real estate
How Ordinals, BRC-20, and Runes turned Bitcoin's block space into contested real estate, generating unprecedented fee spikes and reigniting the debate over what Bitcoin is for.
The Energy Displacement
What happens when AI outbids Bitcoin for electricity
AI data centers are outbidding Bitcoin miners for electricity, and the displacement is reshaping mining economics in real time.
The Security Budget Cliff
What happens when the revenue that funds Bitcoin's security starts to disappear
The declining block subsidy, the transaction fee market that's supposed to replace it, and the formal academic literature that says the transition may not work.
The Thermostat
Why the difficulty adjustment — not proof-of-work — is Bitcoin's real innovation
Why the difficulty adjustment — not proof-of-work — is Bitcoin's real innovation. A thought experiment about unlimited energy reveals what actually makes Bitcoin scarce.
Privacy Through Economics, Not Cryptography
When tracing Bitcoin becomes too expensive to bother
Sometimes the analysis is technically possible but economically pointless. The cost of performing it exceeds the value of the information it produces.
Change Address Detection
Guessing Which Output Comes Back to You
Change address detection is probably the second most relied-upon technique in Bitcoin forensics. Get it wrong, and your entire trace goes sideways.
Blockchain Transparency and AI Interpretability
Both fields face the same fundamental problem: reverse-engineering emergent behavior in systems that weren't designed to be understood.
Both fields face the same fundamental problem: reverse-engineering emergent behavior in systems that weren't designed to be understood.
Taint Analysis
How Dirty is Your Crypto?
When 2.0 BTC from a hacked exchange mixes with 3.0 BTC of legitimately purchased coins, how tainted is the result? It depends entirely on which model you use—and there's no industry standard.
Peel Chains
The Complete Investigator's Guide to Bitcoin's Most Common Laundering Pattern
The peel chain is deceptively simple. One input becomes two outputs, with most of the value flowing in one direction and a small amount diverted to the other. Repeat this fifty times, and you have a classic fund extraction pattern.
Following Bitcoin Through the Maze
Flow Tracing Methodology
Flow tracing sits at the heart of blockchain forensics. It is the methodology that connects a known starting point to downstream destinations where funds might be recovered or attributed.
The Common-Input-Ownership Heuristic
Foundation and Fractures in Bitcoin Forensics
Reid and Harrigan's 2011 paper first formalized what remains blockchain analysis's most powerful—and most vulnerable—assumption.
The Crypto Compliance Landscape at the End of 2025
Get new issues by email
Subscribe on Substack or follow on LinkedIn to receive new issues.
Prefer to listen?
Explore the companion audio archive on Bitcoin and blockchain forensics.
Listen to the podcast