Stephanie Kirmer's reflection on the $200 billion investment bubble and the shifting role of the ML engineer is exactly the kind of grounded perspective this moment needs. We agree with her central point: trust in AI companies has eroded, and rebuilding it will require more than flashy funding rounds or promises of general intelligence. For the ML engineer, this means the job is no longer about chasing the next frontier model, it's about making the tools we already have work reliably, transparently, and for real people.
Kirmer's day-to-day work has changed with the rise of LLMs, and that change is instructive. She isn't spending her time training massive models from scratch; she's integrating, evaluating, and adapting existing systems to solve specific problems. That's a practical shift away from the hype cycle and toward the kind of engineering that actually delivers value. For anyone in a similar role, the lesson is clear: your expertise in data pipelines, evaluation frameworks, and deployment realities is more critical than ever. The companies that will rebuild trust are the ones that prioritize these fundamentals over grand narratives.
What this means for readers who work with spreadsheets or data tools is equally direct. The same principle applies at every scale. When an AI feature promises to automate your workflow, the question isn't whether it's "cutting-edge", it's whether it produces results you can verify and rely on. Kirmer's experience underscores that the most trustworthy AI is the kind that fits into existing processes without requiring a leap of faith. Your job, whether you're an ML engineer or a business analyst, is to demand that transparency.
The investment bubble Kirmer references isn't a reason to dismiss AI; it's a reason to get specific about what works. The engineers and teams that will thrive are those who focus on measurable outcomes, not on being the first to deploy the latest model. For our readers, the takeaway is to treat every new AI tool the same way you would a new spreadsheet function: test it, understand its limits, and only adopt it when it makes your work more reliable, not more complicated. That's how trust gets rebuilt, one verifiable result at a time.