Writing

Essays and notes on the commercial side of AI, Bitcoin, and frontier technology.

You can't print this one on a t-shirt

The US barred foreign access to Anthropic's Fable 5 and Mythos 5, an export control that reads less like security than punishment for refusing the Pentagon.

Who decides which agents get to transact?

Mastercard and Visa announced agent credentialing systems on the same day. The claim underneath: the right to decide which AI agents can act in commerce.

Apple rented the brain

Apple just handed Google the intelligence layer of 2.2 billion devices, and the contract clause protecting users costs Google almost nothing.

Inference got 430× cheaper. Your agent didn't.

Inference got 430x cheaper per token, but agents burn 100x more of them. Real savings on agentic workloads is 4.3x, not 430x. Why the curves diverge.

Compute is an energy business. Bitcoin miners knew first.

The binding constraint on AI isn't chips—it's power. The Bitcoin miners who survived the 2024 halving already own the grid-connected megawatts.

The cost of intelligence just hit zero

AI inference prices fell 1000x in three years. Open-weight models like DeepSeek and Kimi mean the floor stays low permanently.

Every agent in production is a stranger

The internet's trust mechanisms all trace back to a human who can be held responsible. Agents remove the human. The infrastructure was not designed for that.

The catalog isn't the market

Registries indexed 100,000+ agents. Developers still use Google to find MCP servers. The catalog problem got solved. The discovery problem didn't.

The limit isn't reasoning. It's reach.

Agent connections are still hand-built. The pieces to fix that exist—but nobody's assembled them yet.

The CLI is the new API

SaaS companies are shipping CLIs built for agents, not developers. The CLI is the new line between connected and irrelevant.

The real tokenomics

The seat was both the pricing unit and the moat. When agents do the work, both collapse. Companies that haven't moved are watching their multiples compress.

Too big to fail, again

AI concentration risk mirrors too big to fail banking. When OpenAI or Anthropic go down, companies that replaced staff with AI pipelines lose all capacity.

Agents need computers, not compute

AI agents need persistent environments, not ephemeral compute. The cloud abstracted away the machine—agents are buying it back.

Agent-era infrastructure

Every layer of internet infrastructure assumed a human user. Agents break that assumption across compute, discovery, identity, and payments.

Agent payments have a three-body problem

Visa, Mastercard, and PayPal shipped agent payment products. Read the fine print and it's the same system with an agent-shaped UI. The real battle is political.

Nothing enforces your agent's rules

Agent skills carry rules with no enforcement underneath. The skill layer is a trust layer—you write the instructions, the model reads them, and you hope.

I know Kung Fu

A SKILL.md file turns an agent into a specialist instantly—no training run, no fine-tuning. Skills are distribution channels, and the authors are the new buyers.

Nobody knows how to price for agents

The SaaS debate isn't proprietary AI vs. open. It's that neither path offers a pricing model that works.

SOUL.md and MEMORY.md are the new hearts and minds

Agent-driven discovery replaces ad-driven attention. Products must be legible to SOUL.md and memorable to MEMORY.md—you can't buy placement in either file.

AI changed what a product is

AI agents and open protocols broke the assumption that products have edges, with consequences for positioning, pricing, and competitive dynamics.