Ideas, experiments, and the occasional useful rabbit hole.
A first-person reflection on what Todoist's Enlightened milestone taught the author about outcome-first system design, how AI now handles much of his workflow maintenance, and why context, monitoring, verification, judgment, and recovery remain his responsibility.
Read the article →A practical guide to verification-first and troubleshoot-as-you-go coding workflows, choosing proportionate checks, and stating what each result does and does not establish.
Read the article →A frustrating Dotloop search and YouTube interruption lead to a task-first product argument: advertising incentives can make revenue-generating features compete with what people came to do, while the keyboard analogy, open-source reflection, and bounded AI case explain why the author puts people before profits.
Read the article →A first-person reflection on why a human-sounding dot can feel relational across varied work, how context, memory, tools, and voice shape that feeling, and what the experience cannot prove about intelligence or consciousness.
Read the article →A fit-first guide to Research Runner 1.0.1: the host capabilities it requires, the records and review boundaries that make its process inspectable and resumable, the public fiction examples that illustrate its output form, and the documented steps for making it available in a compatible host.
Read the article →AI can help new engineers finish today’s change, but output alone does not build the understanding and judgment needed to own tomorrow’s problem. Teams should preserve consequential work, explanation, independent diagnosis, feedback, revision, and progressively greater responsibility.
Read the article →What GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna add to a frozen-task experiment, how to inspect each saved profile, and why one response per profile establishes neither a universal ranking nor a causal setting effect.
Read the article →Why The Model Release Experiment keeps one frozen story-and-math task, where readers can inspect model cards and run pages, and why receipt checks cover arithmetic rather than overall response quality.
Read the article →Observed: All 24 trials in this small synthetic study met the full intended specification on the first pass. Spelling-error mean output stayed within 2% of clear; reduced-grammar runs used 5% more output for Task A and 18% more for Task B; ambiguous-rule runs used 9% more and 27% more. Interpretation: Ambiguous rules transferred rule selection to the agent, so specify governing rules, exceptions, invariants, and success checks first. The one-model, two-task sample, conventional intended answers, non-causal design, and all-success ceiling do not show that ambiguity is safe or that surface form never matters.
Read the article →Why I expect AI alignment to improve, what benchmark results can establish, and whether refusing unsafe instructions belongs in the definition.
Read the article →A practical two-reviewer cycle for pruning an AI-assisted codebase: reconcile independent evidence, make one reversible change, and audit the behavior it must preserve.
Read the article →Restricted and uncertain compute access can reduce experiments while making efficiency more valuable. Chinese model disclosures show adaptation, but not that chip controls caused innovation or improved the odds of breakthroughs, including any path to AGI.
Read the article →An inherited product default can constrain how model capability reaches users. Review it as a versioned product contract with an accountable owner, task-specific evaluation, decision evidence, a canary against a control, and a trigger for reconsideration.
Read the article →A prompt can name a change but cannot complete it alone; context, tools, permission, feedback, and claim-matched verification turn an instruction into trustworthy computer work.
Read the article →Apple's September Siri AI beta is meaningful progress, but its documented access limits leave a dated concern: Apple's integrated operating model may still slow AI delivery until broad, reliable product results prove otherwise.
Read the article →Why I built a longitudinal record to challenge my forgiving memory, guide decisions about standards, rework, subscription economics, and project attention, without treating activity as achievement.
Read the article →A seeded optimizer gamed its score, one live agent repaired an apparent intent gap, and another obeyed a false criterion. These small historical demonstrations lead to a practical contract test for concrete AI goals.
Read the article →Build an inspectable AI-assisted article workflow with observable artifacts, evidence-based voice controls, scoped publication, live verification, and a clear boundary between verified publication and search indexing.
Read the article →