Things I Believe That Most People in AI Would Disagree With
Contrarian Essay
Stop Reading Papers for Six Months
Most ML engineers would be more effective if they stopped reading papers for six months. The field has a consumption addiction disguised as professional development. People read about RLHF advances and mixture-of-experts architectures the way others scroll Instagram: compulsively, passively, and with diminishing returns.
The engineers I've seen ship the most impactful systems aren't the ones with the deepest arxiv habits. They're the ones who spent that time understanding the business domain they're serving.
Where Top Performers Spend Their Time
At OptiU, my biggest forecasting accuracy gains didn't come from architectural innovation. They came from learning how procurement managers actually think about safety stock. Domain knowledge is the most underrated competitive advantage in applied ML, and paper consumption actively displaces it.
Agents Are Being Dramatically Over-Architected
The multi-agent hype cycle has people building elaborate orchestration frameworks for problems that a well-structured script could solve. I say this as someone who builds multi-agent systems professionally.
Most "agent" deployments I've seen in the wild are glorified prompt chains wearing a trench coat. The industry is reaching for autonomy and tool use as defaults when the actual bottleneck is almost always data quality, task decomposition, or simply understanding what the user needs.
The Agent Gap
The unsexy version that works will beat the elegant version that doesn't, every single time. We keep forgetting this, and the agent wave will re-teach it painfully.
The Best Preparation for AI Engineering Isn't Computer Science
It's running something. Managing a real operation with real stakes — whether that's a business, an event, a team — teaches you constraint satisfaction, failure recovery, and user empathy in ways that no curriculum replicates.
The engineers who've only ever operated inside sandboxes build systems that work in sandboxes. I've watched my experience running fencing tournaments directly shape how I design fault-tolerant pipelines.
Skills That Actually Matter in Production AI
I'd hire someone who's managed a restaurant over someone with a perfect LeetCode score if the role involved production systems.
Eval-Driven Development Matters More Than Model Selection
The industry obsesses over which foundation model to use while treating evaluation as an afterthought. I've seen teams spend weeks benchmarking GPT-4 against Claude against Gemini on vibes, then deploy whichever "felt better" with no systematic eval framework.
The teams that win are the ones that define what good looks like before they write a single line of model code. Evals are the product specification. Everything else is implementation detail.
Be Skeptical of Anyone Whose Entire Career Has Been in AI
The field rewards insularity. Conferences cite conferences. Benchmarks reference benchmarks. The most dangerous blind spot in the industry is the assumption that all the important problems and solutions live inside the AI ecosystem. They don't.
The best ideas I've encountered came from supply chain theory, operations research, and competitive fencing logistics. The field needs more people who've solved hard problems in other domains and fewer people who can recite the attention mechanism from memory.
Where the Best Ideas Actually Come From
Constraint optimization, safety stock models, demand sensing
Scheduling, resource allocation, combinatorial optimization
Real-time systems, failure recovery, user empathy under pressure
If none of these made you uncomfortable, I didn't do my job.
If all of them did, you might be exactly the kind of engineer the industry needs fewer of.