Every few weeks, a new model climbs to the top of a public leaderboard, and the announcement makes the rounds: state of the art on MMLU, a record on HumanEval, another point squeezed out of GSM8K. For most of the industry, that number is the headline. For Olivia Chen, a data scientist and machine learning engineer with six years of shipping models into production, it is closer to background noise.
In a recent essay on her site, Olivia Chen makes a claim that is unusually blunt for a field that tends to speak in hedged probabilities. Most public benchmarks, she writes, are “vibes, not science.” Her comparison is memorable: leaderboard scores are the SATs of artificial intelligence, impressive on a transcript and largely meaningless once the real work begins. The provocation lands because it comes from someone who has spent years watching high-scoring models behave badly the moment they meet an actual user.
What makes Chen’s argument worth taking seriously is not the contrarianism. It is that the rest of the discipline is quietly arriving at the same conclusion.
The uncomfortable truth underneath Olivia Chen’s critique is that many benchmark scores are inflated for reasons that have nothing to do with intelligence. Because leading models are trained on enormous slices of the public internet, the test questions frequently end up inside the training data. Researchers at Yale and Georgia Tech documented this contamination directly, showing that commercial models could correctly guess masked answers from the MMLU test set at rates far higher than chance, with one system reproducing hidden options 57 percent of the time. A model that has effectively seen the exam before it sits down is not demonstrating reasoning. It is demonstrating recall.
Chen’s phrase for this is “test prep,” and it captures the dynamic precisely. When a team advertises that it has beaten a rival on a public benchmark, what that often signals is careful optimization toward the benchmark itself rather than a leap in general capability. The score goes up. The underlying usefulness does not necessarily follow.
The gap between a strong score and a trustworthy product is not academic. It shows up in public, sometimes spectacularly. When one of the most resourced AI systems in the world began advising users to add glue to their pizza and to eat rocks for their health, the failure had nothing to do with a low benchmark result. The model was capable. It simply could not tell an earnest question from internet sarcasm, and the mistake was the kind that gets screenshotted, shared, and remembered. That, Chen argues, is the category of failure a leaderboard will never warn you about.
The more valuable half of Chen’s essay is not the critique but the alternative. Rather than tracking leaderboard position, she evaluates candidate models against a set of signals drawn from the realities of production. Read together, they amount to a practical philosophy of model selection that any data team could adopt.
The first is failure mode diversity. Olivia Chen is far less interested in a model’s average performance than in its worst five percent of outputs. A system that is reliably good, she notes, is worth more than one that is brilliant most of the time and catastrophically wrong on occasion, because the occasional catastrophe is exactly what users notice and publicize. Average accuracy hides the tail. The tail is where reputations are made or lost.
The second is latency under real load. A model that answers in eight seconds once concurrency climbs is, in her words, a loading spinner with extra steps. She runs load tests that mimic real traffic rather than admiring single query benchmarks on an idle machine, because responsiveness at scale is a feature users feel even when they cannot name it.
The third is prompt sensitivity. If changing a single word in a system prompt can swing accuracy by twenty points, the model is too fragile to ship. Chen wants systems that are, in her framing, boring and reliable rather than exciting and unpredictable, a preference that sounds modest until you have maintained something built on the alternative.
The fourth is real task evaluation. Instead of asking whether a model can solve a contrived math problem, she asks whether it can turn a messy customer support conversation into a ticket her team can actually act on. The distance between benchmark tasks and the work a model is hired to do, she argues, is where most systems quietly fail.
The fifth is human preference alignment. Chen runs blind tests with real users clicking to say whether an output helped or hurt, rather than leaning on other language models to grade the work. It is slower and messier than an automated eval, and in her experience considerably more honest.
None of this is idiosyncratic. It mirrors where the research community has been heading. Stanford’s Center for Research on Foundation Models built an entire evaluation framework called HELM on the same premise Chen argues from experience: that a single accuracy figure flattens too much, and that models should be measured across many dimensions at once, including efficiency, robustness, and the trade offs that only appear under real conditions. When a leading academic lab and a working practitioner independently reach the same conclusion, it stops looking like an opinion and starts looking like a standard.
Olivia Chen’s most useful move is to reject the question the leaderboard is built to answer. There is no single best model, she argues, only a best fit for a particular job, a particular set of users, and a particular tolerance for the ways a system can go wrong.
In her own practice, she describes reaching for different tools for different reasons. She favors models whose failure modes are predictable and whose behavior over long contexts stays graceful, values speed and creative range where a task rewards it, and turns to enterprise-oriented stacks when a workflow demands tight integration over open-ended flexibility. The specific preferences matter less than the underlying discipline: she chooses based on the shape of the problem, not the ranking of the tool. Benchmarks compress all of that judgment into one number, and Olivia Chen has simply stopped pretending the number carries the weight the industry assigns it.
That posture, equal parts skeptical and constructive, is what distinguishes her writing. Chen is not an AI pessimist. She is an optimist about the technology who insists on evaluating it honestly, and her essays and long-form field notes consistently return to the unglamorous craft that separates a demo from a durable product. In a discourse dominated by launch day superlatives, that is a rarer and more useful voice than the leaderboard rewards.
Her closing warning to teams still selecting models by ranking is characteristically direct. Chase leaderboard position, she suggests, and you are almost certainly leaving real performance on the table while shipping things that break in ways you have not yet imagined. It is the kind of advice that only sounds obvious after someone with the scars to back it up has said it out loud.