Candace

Candace

Human, authenticated.

Most of us building with AI are constantly wondering if we're solving the wrong problem. We started by optimizing individual tasks, but quickly realized we should be writing agentic workflows. Then we realized that we should be focusing on writing semi-autonomous agents. No wait, but now we need to think about *insert latest substack article here.*

The reality is a paradigm shift: we are redesigning how humans and systems collaborate. We haven't suffered the consequences of the newest risk surfaces. But the truth is, we are just getting started. As individuals, we're best served focusing on how the virtual machines in our own heads must evolve to address the seismic changes we're living through. As organizations, we must put in place flexible structures designed to evolve and mature, not anchored on any given stopping point.

This is a generational inflection point (all puns intended), and not one I can watch from a distance. The map below is how I think about this work, organized around the human, technical, and organizational structures that give meaning to movement.

Humans

Building teams that think clearly about systems—hiring for judgment, designing roles that evolve.

Notes in progress
Senior engineers as AI researchers

Hands-on technical staff are your best bet to become your next AI technical staff. They're positioned for it. But more than that: they have the mindset.

AI tsunami: people

AI interaction is intense. Two people with different predilections will set up evals, constitutions, and frameworks very differently—and end up with very different results.

Operations

Designing the workflows and decision architectures that make human–AI collaboration actually work under uncertainty.

Notes in progress
Operational velocity vs. safety

The tension between moving fast and not breaking things is now an AI question, not just a DevOps one.

Models are stateless

AI models are fundamentally stateless. It's the harnessing—workflow, context, evals—around them that makes them useful.

Evals: unit tests for AI

Without evals you don't have engineering, you have vibes. The discipline travels.

Industry

The industry is bifurcating between AI-native builds and legacy retrofits. I have opinions about which survives.

Notes in progress
AI tsunami: technical

Code is being generated at 10x, but the software factory that safely moves intent to production wasn't built for it.

Frontier & open source models

Google's open source model and what it signals about the next 18 months of the stack.

What I consume