Why Aspiring AI Engineers Are Wasting Years of Study

There's a particular kind of hell reserved for developers trying to break into AI engineering: six months deep in calculus tutorials, drowning in backpropagation derivations, convinced that somewhere in these equations lies the key to getting hired. Marina Wyss, a Senior Applied Scientist at Twitch/Amazon, wants you to know you're optimizing for the wrong role.

In a recent video, Wyss maps out five mistakes that keep aspiring AI engineers spinning their wheels. What's interesting isn't just the advice itself—it's the underlying tension between what the industry claims it wants and what it actually needs. The role of "AI engineer" is both desperately in-demand and widely misunderstood, even by people trying to hire for it.

The Math Trap

The first mistake, according to Wyss, will irritate the traditionalists: spending months on mathematical foundations before building anything. "After 6 years working on production AI and machine learning systems, I've never once needed these skills," she says, referring to hand-calculating derivatives and deriving backpropagation from scratch.

This is where definitions matter. Wyss draws a clear line between AI engineers and traditional ML engineers or data scientists. AI engineers, as she defines them, "build applications on top of pre-trained foundation models like GPT-5 or Claude." They're working with prompt engineering, retrieval augmented generation, fine-tuning, and AI agent systems—not training models from the ground up.

The argument is pragmatic rather than anti-intellectual: you need conceptual understanding of supervised versus unsupervised learning, model evaluation metrics, and neural network intuition. But "every hour you spend on hand calculations is an hour you're not spending building the skills that will actually get you hired."

What Wyss doesn't address is how hiring managers actually filter candidates. Do job descriptions reflect this reality? Or are companies still requiring mathematical rigor they don't actually need, creating a mismatch between stated requirements and actual work? The shortage of AI engineers she mentions suggests the latter—that credential creep is pricing out capable builders who could do the job.

The Model-Only Mindset

Even when people understand they're building with pre-trained models rather than training from scratch, they hit another wall: treating the model as the entire project rather than one component.

"Your chatbot might work great on your laptop with three test users, but what happens when a million people try to use it simultaneously?" Wyss asks. Real AI engineering involves evaluation, performance optimization, security, cost management, and deployment reliability. The model is just the part that generates text or embeddings—the system is everything around it that makes that model actually useful to humans at scale.

This mirrors a pattern I've seen in open source: the tendency to confuse "working demo" with "production system." GitHub is full of impressive prototypes that would collapse under real load or leak API keys within hours of public deployment. The gap between proof-of-concept and production-ready is where most projects die, and it's also where most learning materials stop.

Wyss points to resources that cover LLMops—deployment hygiene, rate limiting, monitoring, security—as particularly valuable because they're "often completely missed from other resources." That absence itself is revealing. It suggests the education ecosystem is optimized for getting people excited about AI capabilities, not for getting them ready to maintain AI systems that companies actually depend on.

The Portfolio Problem

Portfolio projects are standard advice for breaking into any tech role, but Wyss highlights a specific failure mode: building what everyone else builds. "If your portfolio consists of toy projects and ChatGPT wrappers, this is the AI engineering equivalent of the Titanic dataset in data science."

The alternative she proposes: "Build something actually useful that solves real problems for real people." Projects should demonstrate the full AI engineering lifecycle—evaluation, observability, deployment, security. They should show hiring managers you can do what they need, not just that you can follow tutorials.

But here's the catch: building production-quality systems requires understanding production-quality concerns, which requires either work experience or exceptionally good learning resources. It's a bootstrapping problem. The people who need portfolio projects most—those trying to break in—are least equipped to build the kind of portfolio projects that actually demonstrate production readiness.

Framework Churn

The fourth mistake addresses a problem specific to fast-moving fields: trying to keep up with every new tool release. "It's like as soon as you've learned one tool, there's five new ones that have come out," Wyss notes.

Her solution: focus on underlying concepts rather than specific frameworks. Develop mental models of architectural patterns rather than memorizing API documentation for tools that might be obsolete within a year.

This is solid advice, but it also reveals something uncomfortable about the current AI engineering landscape. If the tools change so rapidly that mastering them is counterproductive, what does that say about the maturity of the ecosystem? Traditional software engineering has stable patterns precisely because the community has converged on solutions that work. AI engineering's constant churn might indicate innovation—or it might indicate we're still searching for approaches that actually solve the problems at hand.

The LinkedIn Black Hole

The final mistake is probably the most honest: relying on online job applications when you have no experience. "It's you versus a few thousand other people applying to that same job on LinkedIn," Wyss points out. "Even though there are a lot of AI engineering jobs right now, there are also a lot of applicants."

Her prescription is networking: connecting with people at target companies, engaging with their content, asking for informational interviews, contributing to open source projects where people might notice, building in public. "None of this is rocket science. It's just work that most people won't do because it feels awkward or time-consuming."

This is where the shortage narrative gets complicated. Yes, companies need AI engineers. But they need AI engineers with demonstrable experience, shipped projects, professional networks—all the things that people trying to break in by definition lack. The "shortage" exists at the experienced level, not the entry level. Breaking in requires the same networking grind that's always existed, just with different technical skills as table stakes.

What Wyss is really mapping is the distance between marketed opportunity and practical reality. AI engineering is a genuine career path with good compensation. But the path itself hasn't been paved yet. The education is inconsistent, the job requirements are often misaligned with actual needs, and breaking in still depends more on who you know than what you know—though you need both.

The video is sponsored by DataCamp, which shapes the framing but doesn't invalidate the core observations. These mistakes are real. The question is whether the solutions scale beyond individual hustle—whether the industry will build better onramps, or whether we'll keep telling people to network harder while complaining about talent shortages.

—Dev Kapoor