The democratization of deep learning has a dark side. Those five lines of PyTorch that fine-tune BERT on your dataset? They’ll happily train a broken model and hand you decent-looking metrics while hiding fundamental flaws. Thomas Wolf, a researcher at Hugging Face, has spent enough time in the trenches to know that the gap between ‘it runs’ and ‘it works’ is where most projects quietly die.
His advice starts with something that sounds almost heretical: put machine learning aside entirely. Before touching a single tensor, he recommends diving into the raw data — checking label balance, questioning whether the gold labels even make sense, tracing how the dataset was collected and where noise crept in. Ask what a rule-based algorithm would score. If you can’t answer that, you have no baseline for judging whether your fancy architecture is actually earning its complexity.
Then comes the humility step. Wolf suggests channeling your inner Andrew Ng student and running embarrassingly simple baselines first: logistic regression on word2vec embeddings, fastText classifiers, anything that takes minutes rather than GPU-days. The uncomfortable question he raises is one most engineers avoid: if BERT only beats a bag-of-words model by two points, is that gap worth the infrastructure cost and debugging surface area? Without knowing what a random predictor scores or what the loss looks like for random guessing, you’re navigating without coordinates.
What makes this perspective valuable is that it comes from someone who builds the very plug-and-play tools that create the illusion of maturity. Wolf admits he’s ‘partially guilty’ of the hype at Hugging Face. That honesty matters because it cuts through the framework marketing. The field isn’t mature — it’s in constant development, and the only reliable defense against fooling yourself is methodical skepticism: understand your data, establish trivial baselines, and never trust a model that trains silently without raising questions.
