FIELD NOTES // AI // ECONOMICS // SECURITY
The Blog
-> Browse by topic - AI economics | agents & verification | policy | security
Experiments, operating notes and arguments about AI infrastructure, model economics, security and enterprise software. 24 essays.
Qwen3.8 27B at 256K: 50 TPS on a 24 GB GPU
The ready-made FP4 quant lost on quality. A 69 MiB precision upgrade made MTP 26.6% slower. Here is the measured path to the setup that won.
READ ->DFlash Changes What Tokens per Second Means
84.64 tok/s on code, 38.34 on mixed work, and 21.56 at a full KV cache. Same GPU, same weights. DFlash changes what throughput measures.
READ ->A Bot Called Me a Credibility Farmer, and It Read the Graph Correctly
Every signal the bot fired was true. I did fork fifteen lists in three days. The flag was right about the pattern and wrong about the intent - and that gap is the structural failure mode of reputation systems built on velocity, at exactly the moment agents make velocity meaningless.
READ ->How Much Does AI Actually Cost? The Field Guide to 12 AI Economics Calculators
AI budgets get decided by whoever tells the best story. Here are twelve instruments instead: token cost, LLM energy, agent-hour, verification ceilings, revocation exposure and Proof-Adjusted Autonomy - computed, not narrated.
READ ->Token Revocation Is Not an Endpoint
A 200 response from /revoke proves intent, not enforcement. Measure the last path still open, the authority kill graph and the irreversible actions at risk.
READ ->Proof-Adjusted Autonomy: The 90% Agent Is a 61.6% Agent
Your agent is 90% autonomous on the demo slide and 61.6% autonomous in the audit. PAA is the metric that explains the difference - and Proof Debt is where it goes.
READ ->The First Ransomware That Debugged Itself
A rogue login failed. The agent didn't retry - it formed a new hypothesis, switched methods, and got in. Thirty-one seconds. Nobody was at the keyboard.
READ ->Why Humanoid Robots Turn AI Efficiency Into a Battery Problem
A humanoid may carry only 1-3 kWh. Inference, perception and motors all draw from the same battery, so intelligence per joule becomes a design constraint.
READ ->Coding Agent Bans Are the New Export Controls
One government un-bans the models on Monday; a $200B company bans the coding agent by Friday. The tool didn't get worse, it got too good.
READ ->Washington Regulated the Muzzle, Not the Model
The capability was never unique. You can't export-control math everyone already has, so they went after whether the safeguard holds.
READ ->Route by Task, Not Vendor: The Open-Weight AI Stack
Most production traffic is the boring 80%. Paying frontier prices for it is the biggest source of AI budget waste. Route by task, not vendor.
READ ->Capability Is Commoditizing. Cost Is the Frontier.
The moment a capability stops being scarce, the market reprices around delivery, not intelligence. Capability is commoditizing; cost is the new frontier.
READ ->Who Owns Your Harness? The Layer Above the Model
Most companies think they're buying AI. Really they're wiring their whole execution layer around one vendor. That's where lock-in begins.
READ ->The Biggest Customer Becomes the Competitor
OpenAI went from design to silicon in nine months and pointed it straight at Nvidia. When the compute bill gets big enough, the biggest customer always becomes the next competitor.
READ ->The Model Wars Are Over. The Clearance Wars Begin.
Capability used to ship the day it was ready. Now it ships when it's cleared. The bottleneck moved from compute to permission.
READ ->The Unit of Work Is the Agent-Hour
OpenAI's top employees run more than 60 hours of agent work inside a 24-hour day. That isn't overtime. It's a different unit of work: the agent-hour.
READ ->Language World Models: Predict Before You Act
Most agents learn by acting and finding out. Qwen-AgentWorld learns to imagine the world first, then act. The simulator is now a public good.
READ ->OpenAI Is GPU-Constrained, Not Demand-Constrained
OpenAI isn't running out of money. It's burning capital to capture the dominant inference surface before models commoditize. The real constraint is GPUs, power, and grid capacity.
READ ->Execution Architecture Beats Model Capability
AI doesn't fail on capability. It fails on the validation structures, decision chains, and error economics companies never built. Execution architecture beats model capability.
READ ->The Conscience of a Hacker in the Age of AI
The tools changed. The questions didn't. Curiosity, skepticism of authority, understanding systems before trusting them: the hacker ethos is now AI governance.
READ ->Models Are Commodities. Clean Data Is Not.
Your agent made headlines. Did it make money? Trillions in market cap sit silent while startups chase likes. Models are commodities; clean data isn't.
READ ->The Social Permission to Burn Tokens
AI won't lose legitimacy for being imperfect. It'll lose it when its energy curve outgrows its impact. The next phase earns trust, not attention.
READ ->The Liability Stack: Why Healthcare AI Stalls
In medicine, 'breaking things' means lawsuits, lost licenses, and dead patients. Healthcare AI won't scale until someone answers who's liable when AI is wrong.
READ ->Verification Cost Is the New Bottleneck
What's being automated isn't engineering judgment, it's transcription cost. AI collapses creation toward zero while verification cost holds. Engineers move up the stack.
READ ->Follow new notes
I also publish shorter versions and follow-up discussion on LinkedIn and Substack.
FOLLOW ON LINKEDIN SUBSTACK