On AI

Guardrails are critical

I see firsthand people taking AI slop and trusting it as if it's gospel. AI has propelled ads, spam, and malicious content alongside all of the good we've gotten from it. AI doesn't know if it's "helping" by pen-testing a developer's benign software or if that developer is actually instructing it to exploit a new zero-day to attack a bank. The technology needs thoughtful guardrails.

The knowledge atrophy problem

Having AI do EVERYTHING for you is bound to cause an ultimate loss of knowledge. I've seen students submit AI-written essays or research contributions, and I've seen where hallucinations have caused them to get it wrong, or they can't complete a test or answer basic questions in a field because they didn't do the work. ChatGPT did. We must learn to work WITH AI, not have AI do everything for us.

We're all learning this together

AI is rapidly evolving and still "new" by any stretch. The gap between a learner and teacher is still relatively small. My best advice for anyone wanting to use AI for marketing, product, or sales: just give it a shot, learn through experimentation and collaboration, and refine your process over time. There are no shortcuts on a path that has yet to be created.

Understand what AI learned from

We still need engineers who actually know how to build things. A model trained on developer forums and public repositories is learning from a skewed dataset: roughly 50% broken code with issues, 50% solved problems. The really good stuff: working, secure, scalable, well-architected production systems - almost never exposes its source code publicly. AI learns from visible debris of software development, instead of the invisible foundations that the best software operates on today. If you don't understand what the AI learned from, you can't predict where it will fail you. Rather than replacing engineers, I believe the future will be led by skilled engineers who understand both the systems being built as well as the AI's blindspots. These engineers become 10x more valuable, and the speed of software and product development will skyrocket.

Embodied context is the next frontier

We often talk about the need for more data in AI development, but overlook the need for humanity and context from the real world to guide it properly. We often assume "digital context" equals full context, but Internet-only data is noisy: keyboard warriors, sarcasm, and the echo chambers of social media can only tell AI about Uncle Frank's political leanings and religious preferences, not the facts and systems needed to build true intelligence. When AI observes not just text, but how humans act, move, and interact in physical space, it gains materially richer context. Wearables, AR glasses, robotics, and handheld devices capture motion, gesture, and environment (instead of just keyboard inputs).

The AI Box Model

Think about how we started - with AI in a chat window. That chat modality is "AI in the box": limited reach to external systems and trained on irrelevant data from non-applicable domains, constrained to simple question-and-response patterns. Then we connect the boxes - an LLM linked to a CAD system via MCP server, or a simulation engine that doesn't just answer questions but actually acts on your behalf. Finally, we move to "out of the box": AI everywhere, with connected systems and platforms generating the specific context you need. Vertical training creates specialist-level outputs versus the generalist responses you'd get from search.

The Antagonist agent

Many of today's AI coding assistants have a fundamental flaw: They treat almost every interaction as an isolated request, driving context rot and mistakes that might never happen through a human perspective. Rather than using a single AI agent, I orchestrate teams of specialized agents with persistent memory and adversarial quality control. The most critical innovation is the Antagonist agent: one specifically tuned to never provide positive reinforcement, and trained on the shortcomings of its own process. It identifies vulnerabilities, questions assumptions, and refuses to acknowledge progress until issues are resolved. This inverts the typical AI tendency toward affirmation and creates a forcing function for quality. If AI systems perform better with formalized adversarial roles, perhaps we should reconsider how we treat critical evaluation in our own processes.

Multi-agent orchestration in practice

Below is a screen capture of four Windsurf AI coding agents working simultaneously on different git worktrees for the same codebase. A fifth window (not shown) serves as the orchestration layer, directing tasks, resolving conflicts, and ensuring coherent progress across all workstreams. This is how I actually build.

Four AI agents working in parallel across git worktrees, orchestrated by a fifth

Making agentic AI accessible

I've always believed that the promise of AI is its ability to elevate our own capabilities and improve us societally. But for many, this technology remains locked behind a wall of complexity. My goal is to make agentic AI workflows accessible to everyone, regardless of technical background - teaching the principles of automation, not just a single tool.

Never copy/paste - always refine

I am strongly against copy/pasted AI output. It must be refined, nurtured, and verified. Context and data must be appropriate to ensure factually and directionally correct outputs. We must be mindful of blind spots that actually hinder us - both the AI's limitations and our own tendency to over-rely on its outputs without critical evaluation.