Engineering Unblocked

There are genuine operational insights buried here - AI tools saving one hour/week at a net negative ROI, coding not being the bottleneck, developer SLOs as a measurement framework - but they're heavily diluted by extended career autobiography (Borland, Netscape, Pixar punch-card nostalgia) that delivers no actionable value for a B2B operator.

“we figured out that we had scientific proof that we had pre and, um, post users of AI and we were able to measure that at, at, on average, they saved an hour a week on the amount of time they spent coding and had zero impact on the amount of overall product velocity”

“Coding is not actually the problem. Right. We've now proven it. Not just us.”

The framing of a self-described AI skeptic running MongoDB's internal AI productivity program and arriving at real negative-ROI data is a somewhat fresh angle, but most of the underlying arguments (coding isn't the bottleneck, junior-engineer-replacement hype is overblown, AI costs are unsustainable) have become increasingly common takes rather than genuinely contrarian ones.

“The amount we saved was less than the amount we were spending for them to save that hour. So the roi, therefore, scientifically no good”

“You think, oh, I'll fire all my junior engineers and have all these agents and all of a sudden we're a billion dollar ARR. And I'm like, uh huh. Notice people aren't saying that as much anymore”

Tara Hernandez is a genuine operator with 30+ years of hands-on infrastructure and dev-productivity experience across Netscape, Pixar, Google, and now VP-level ownership of CI, build systems, security infrastructure, and AI at MongoDB - a real practitioner who has measurably run these programs at scale, not a career thought leader.

“MongoDB has a proprietary Evergreen. Proprietary CI system called Evergreen, don't judge us. It predates a lot of commercial, uh, uh, solutions and it is purpose built to build and test a distributed documentdb. We run about 3 to 400,000 compute hours worth of tests for this thing a day”

“one of my, uh, directors actually wrote a new procurement policy because having too many of the same type of tool causes all kinds of flags to go for finance”

There are some credible concrete numbers - 300-400K compute hours/day, one hour/week AI saving, 50% initial Slack bot accuracy - but the episode is substantially weakened by repeated vendor anonymization ('I don't want to get sued'), vague cost comparisons with no actual dollar figures, and high-level assertions about ROI without quantified thresholds.

“we run about 3 to 400,000 compute hours worth of tests for this thing a day”

“we figured out that we had scientific proof that we had pre and, um, post users of AI and we were able to measure that at, at, on average, they saved an hour a week”

The host sets up context reasonably and occasionally lands a good reactive follow-up ('This is 18 months ago?'), but consistently fails to press on vagueness - never asking for the actual dollar ROI delta, which specific signals changed their vendor choice, or what percentage of engineers are now on agentic tooling - and allows long autobiographical detours to run unchecked.

“This is 18 months ago?”

“Speaking of those thousand different things. So at Stripe, I ran unsuccessfully, um, I'll be honest, a program called Papercuts”