José Macías Silva

The first wide turn was unnerving. Letting a vehicle steer itself — knowing a failure mode meant consequences for my life and others — is the kind of nervous-system event you don’t forget. But as the miles accumulated, I started noticing patterns: where the system was confident, where it hesitated, which lane markings tripped it up, which merges it handled cleanly. The fear didn’t disappear — it converted into something more useful. A working model of when to trust the system and when to take back the wheel.

By the end of several long trips, Autopilot had saved me from a few potential collisions and, just as importantly, reduced the fatigue that makes human failure modes more likely. Two sets of eyes. Two sets of judgment. One overall safer drive. That experience reshaped how I think about AI in production: the right relationship with a powerful system isn’t blind trust or blanket skepticism — it’s pattern recognition. Understand its strengths, respect its weaknesses, design the workflow so the human stays accountable for the consequences.

Two years later, during the COVID lockdown, I was building out Nighthawk and burning long evenings on a question that had been bouncing around since my business-analyst years: what if I had a super-intelligent version of myself that remembered every query, every operational quirk, every piece of institutional context — and could generate analytical work at a fraction of the time? I started researching ML, and that’s where I found Rasa — an open-source Python framework for building conversational AI agents with intent classification, entity extraction, slot-filling, and dialog management. Pre-LLM, this was real ML work: hand-curated training data, custom dialog flows, the whole pipeline. I built an internal NLU agent on it for fun. It was the first time I’d built something that could understand intent rather than just match strings.

When OpenAI released GPT-3 / ChatGPT, I signed up the day it opened. The first hour with it crystallized something I’d been chasing through Rasa: this was the leverage I’d been looking for as an engineer, as a business analyst, and honestly as someone juggling more projects than time. The interaction model had shifted from “describe every rule to the system” to “describe the goal and collaborate.”

The most memorable thing I asked it for early on wasn’t an engineering problem — it was a structured nutrition plan and a training schedule for a marathon. The plan worked. It rekindled my love for running, made me measurably healthier, and convinced me that this collaboration model wasn’t just for code. It was a lever you could pull on any part of life worth investing in.

Today, my daily driver is Claude Code powered by Anthropic’s frontier Claude models. The fidelity of the collaboration keeps improving — it’s closer to working with a sharp colleague than operating a tool. I believe this is the technology of the century.

But the lesson from the Tesla still applies. I can see the patterns. I can see where the model is confident, where it drifts, where the failure modes are subtle enough to slip past a casual reviewer. Leveraging AI well means designing for its strengths and engineering around its weaknesses — and being the one in the room who flags the safety and security implications before they become incidents.