Someone is reselling the AI you paid for.
Today: the new economics of running models — who pays the bill, who steals the compute, who deploys the agents, and what counts as real work when the cheap output floods in.
Someone is reselling the AI you paid for.
If you ship an AI feature, your inference endpoint is a wallet. Vercel walks through inference theft — bots hammering your paid model calls and reselling the output as their own product — and shows how to price it out per request before the invoice does.
THREAT MODEL // attacker scripts your /api/generate endpoint, resells completions through a cheaper-looking wrapper. you absorb the token cost; they keep the margin. mitigation: bind every request to a verified caller, reject the headless traffic before it reaches the model, and watch the cost-per-caller line, not just total spend.
Close the laptop. The agents keep coding.
Conductor moved its fleet of coding agents off developer laptops and into Vercel Sandbox, so the work continues after you walk away. The piece is a practical map of running many agents in parallel as cloud jobs — the same pattern Notion, Linear, and Ramp now lean on.
From the ticket to the diff.
Braintrust wired Codex (on GPT-5.5) into the path between a customer request and shipped code — running experiments and opening changes faster than a human queue could. A concrete look at what an agentic dev loop does when it is plumbed into real support traffic.
A rubric for the evals you trust.
OpenAI published a playbook for third-party model evaluations: what makes a capability test actually valid, how to probe a model's safeguards instead of its headline score, and how to write up a method so someone else can reproduce it. Read it before you commission, run, or cite an eval.
- 1Validity first — decide what the test can and cannot claim before you run it.
- 2Probe the safeguards, not only the capability score.
- 3Publish the method so the result can be reproduced and challenged.
The engineer ships beside the customer now.
A role hardening into a title: the AI Forward Deployed Engineer, embedded inside a client's org to bend agentic workflows around how that business actually runs.
The pattern borrows from elite consulting and pairs it with shipping access: sit with the customer, learn the workflow, then wire the agents directly into it rather than handing over a generic SDK.
For founders, it reframes go-to-market — the deploy is the sale. For engineers, it is a career lane that rewards taste in problems as much as code, and proximity to the people whose work you are automating.
The same issue clocks two other signals worth a glance: agent traffic is now a measurable slice of the open web, and AI-Act timelines are slipping again.
THE HUMAN SIGNAL
Bullas's manifesto for a web going grey with generated sameness: re-assert the things a model can't source — lived specifics, a real point of view, a voice — and treat them as the scarce signal worth paying for. Scaling output without them just feeds the slop you are competing against.
The cheaper it is to publish, the more a human signal is worth.
That's today.
Six picks on the cost, plumbing, and authenticity of running AI — 30 May 2026.