Satya Nadella inverts Kenneth Arrow's information paradox: in the AI age it's the buyer, not the seller, who risks giving away knowledge, paying once in money and again in the proprietary data a model needs to perform. Every prompt, correction and eval becomes 'exhaust' the provider learns from, skewing the asymmetry toward whoever owns the learning infrastructure. Nadella argues enterprises must own their evals and memory behind a trust boundary.
Jamin Ball reframes Alex Karp's 'own your weights' pitch: a downloaded open model is a melting ice cube, a frozen snapshot that decays in relative terms as frontier models pull ahead. The durable asset is the engine behind the weights, your data flywheel, RL infrastructure and eval harness that keeps producing better task-specific models. Before claiming model sovereignty, check whether you can RL an open model against your own workflows.
Epoch analysed 41 core contributors to OpenAI's public Codex repo, using LLM judges to estimate how long each merged pull request would take an unassisted engineer. In Q2 2026, 8% of contributor-days reflected work judged at over 24 hours of solo effort, up from 2% a year earlier. Treat the figure as an upper bound on time saved, since the judges are imperfect and longer work isn't always more valuable.
Aparna Dhinakaran and Laurie Voss argue that 'the loop,' AI engineering's word of the month, hides four different architectures. They separate execution loops, an agent acting, observing and deciding, task loops like Geoffrey Huntley's Ralph Loop that restart an agent on a fresh context until tests pass, and product loops, the software factories running triage through shipping. When someone sells you a loop, pin down which layer they mean.
Daniel Miessler argues the single biggest shift right now is moving from prompt engineering to intent engineering: stop specifying the steps and describe the outcome you want. He grounds it in Sutton's Bitter Lesson, where hand-written procedures get relatively dumber as models improve, so detailed HOW instructions increasingly poison a model's own approach. Audit your prompts and scaffolding with your strongest model and convert HOW instructions into WHAT ones.
This essay defends engineers working in codebases too big to hold fully in one head, arguing partial understanding is the only realistic mode, not a failure. It pushes back on Naur's 'Programming as Theory Building,' which claims a broken system should be rebuilt from scratch, since large systems carry thousands of quirks no rewrite can reproduce. When you're unsure, take a position, make your best guess, and own the consequences.
Greg Jarboe warns that free AI citations are a temporary window, not a permanent channel. He draws the parallel to July 2013, when Google reclassified optimised anchor text in press releases as unnatural links and erased that ranking tactic overnight; Resonate Labs' Shane Tepper says what's closing is 'the cheap part,' winning position with work rather than budget. Bank earned visibility now and plan for a paid, fenced-off phase.
Responding to the AI 2040 forecast, George Hotz rejects hard-takeoff thinking: intelligence is just today's bottleneck, not a master key. Reality resists tokens, he argues, because chip fabs still take three months and hardware, supply chains and physical failures set the pace no matter how smart the model gets. He also insists AI should run locally and answer to its owner, since a model you can't overrule isn't truly aligned.
