Underdog shipped a 32MB speech model inside a browser game you can play. The question that tests for over and above.
On November 6, 2026, TechCrunch reported that Underdog, a startup building models small enough to run on edge devices, had released a game built with its new text-to-speech model, Woof. The model is 32MB. According to TechCrunch, the game is a room with a retro tech theme, a typewriter, and a CRT monitor. You type on the typewriter with your keyboard and the game speaks it back, and you can rotate a tuner to generate different kinds of music. There is a desk lamp that makes a sound when you click its switch. All of it runs in the browser, and the company says the model can run on most devices without any internet connection. TechCrunch reported that the model has 31 million parameters and is open for everyone to download, and it carried a caveat worth keeping: a model this small likely won't replace bigger models.
Read what was shipped against what was expected. The customary form of a model release is a weights file, a README, and a benchmark table, and anyone who wants to know whether the thing actually works has to go and find out for themselves. Underdog's founder, Sigil Wen, told TechCrunch he wanted to build his own system with models that run on users' devices and let them own their data, and the demo makes that case without a slide. The game's intro warns about "intrusive" cloud providers in the town. The lamp clicks. The typewriter talks back. The release answers the question a customer would have asked next, before they asked it. That habit, seeing the unstated need and closing it inside the delivery itself, is what an interviewer is probing with a question that sounds much softer than it is.
Why they ask it
The question reads as a customer-focus check, and at the junior level it is. At the senior level it is a judgment check wearing a customer-focus costume. Going over and above is cheap to claim and expensive to prove, so the interviewer is listening for three things. Did the candidate understand what the customer actually needed, as distinct from what they asked for? Did they decide the extra was worth its cost, rather than simply doing more? And did they know when to stop? A manager who routinely exceeds expectations without a model of cost is not a customer champion. They are a scope risk.
The interviewer is also listening for whether the customer is a specific person with a specific problem. An answer that never names the need is an answer about effort, and effort is not the skill being tested.
The trap
The first failure mode is the heroics story. The candidate worked the weekend, added four features nobody asked for, and the customer was delighted. This proves stamina and nothing else, and a careful interviewer hears gold-plating. The second is the silent-extra story: the candidate delivered more than the brief without ever checking that the customer wanted it, which is a negative signal dressed as a positive one. The third is the vague-customer story, where stakeholders were pleased but the answer cannot say what they expected, what they got, or what the difference was.
A strong answer does three things differently. It states the expectation precisely enough that the listener could have written the acceptance criteria. It names the gap the customer had not articulated, and says how the candidate found it. And it shows that the extra was either cheap relative to its value or chosen deliberately, with the cost visible to someone who could have objected.
Applying STAR-T
Situation. An internal analytics team asked for a nightly export of billing events into a warehouse table. The stated expectation was clear: twelve fields from their spec, refreshed by six in the morning, starting in three weeks. The team had been burned by a previous feed that silently dropped rows, and the candidate learned this in a kickoff conversation rather than from the ticket.
Task. Own the delivery end to end, on the date promised, for a consumer who would not trust the data until they had reconciled it themselves.
Action. The candidate shipped the table on schedule, as specified. Alongside it, and after telling the requesting lead what it would cost, they added two things the spec did not ask for: a daily row count and checksum written to a side table the analysts could compare against the billing source, and a freshness alert that fired if the export had not landed by the deadline. The addition took two engineering days, taken from a lower-priority backlog item that the candidate named and deferred in writing.
Result. The analytics team stopped writing their own reconciliation scripts in the first week, which they had budgeted two weeks for, and the side table was picked up by a second consumer the following quarter. The strongest version of this story includes the moment the lead was told about the extra work before it happened, because that is where going over and above separates from going off-brief.
Trade-off. The deferred backlog item slipped by a sprint, and the candidate did not build the dashboard the analysts had hinted at, because a dashboard was a want and reconciliation was a need. Saying so out loud is the point. The question rewards an answer that can name what it chose not to do.
The follow-up that breaks weak answers
How did you know they wanted that? Or, in its sharper form, What did it cost, and who agreed to pay it? The heroics story collapses here because the extra was a guess, and the silent-extra story collapses because nobody agreed to anything. A strong answer has evidence that the need was real, usually a sentence the customer said or an incident they had lived through, and an account of where the cost landed. If the candidate cannot answer the second question, the first answer was never about the customer.
Score your answer against the director’s bar
Q: Tell me about a time when you delivered over and above customer expectations.
Bank the story where you saw the unstated need and paid for it openly, then rehearse it until the cost is as clear as the result. Try it free →
