This is one of those ideas that sounds clean and helpful until you picture what people will actually do with it.
TinyFish just launched something called BigSet. It’s open-source. The pitch is simple: you write a plain-English description of the data you want, and the system builds a structured, live dataset for you. Not a one-time spreadsheet you babysit, but a dataset that can refresh itself over time.
On paper, that’s a gift. In real life, it’s also an excuse. And that’s where my unease starts.
The obvious win is speed. If you’ve ever tried to pull together a dataset for a project—anything from “all the new job postings in this city” to “every company in this niche with these traits”—you know the grunt work is the bottleneck. It’s not the fancy analysis. It’s the messy collecting, the cleaning, the endless “wait, what counts as the thing we’re counting?”
BigSet tries to erase that friction. You describe what you want in plain sentences. It infers a structure for the data, then uses multiple agents—an orchestrator and other pieces—to go find and fill it. It can also refresh the dataset automatically so it stays current.
If it works well, it changes who can do “data work.” Not just analysts, not just engineers. Anyone who can describe a need clearly gets a dataset that looks like it came out of a serious pipeline.
That democratization story is real. It’s also the story people tell right before they stop asking hard questions.
Because the most dangerous part of data has never been collecting it. It’s the false confidence that comes after. A structured dataset feels official. Columns and rows give people the impression that reality has been captured cleanly. And when the dataset keeps updating itself, it feels alive and trustworthy—like a dashboard you can lean on.
But a dataset built from a plain-English prompt is only as good as the prompt, the inferred schema, and whatever choices the system makes while gathering. Those are judgment calls, even if they’re disguised as automation.
Imagine you’re a small business owner. You type: “Track competitor prices for these products every day.” You get a neat table, refreshed automatically. Great—until the system starts pulling “equivalent” items that aren’t actually equivalent, or mixes sale prices with regular prices, or misses changes because the source pages shift. You’ll still make decisions with confidence. You’ll still change your pricing. You’ll just be doing it with a quieter kind of error.
Or say you’re in a city office trying to monitor “recent building safety complaints.” The dataset refreshes. The table looks clean. The temptation is to treat it as reality and not a messy slice of it. That’s how you end up allocating attention and money based on what was easiest to collect, not what was most important.
The BigSet idea leans hard on “schema inference” and “orchestrator agents.” Translation: the system guesses what fields should exist, and it coordinates how to gather and update them. That’s clever. It’s also where the hidden power is. Whoever controls the defaults controls what people think the world looks like.
Even if it’s open-source, most users won’t read the guts. They’ll copy a setup, run it, and trust the output. That’s not a moral failing. It’s normal behavior. People are busy. They want it to work.
And then there’s the refresh feature. A dataset that updates itself sounds like less work, but it can also mean less noticing. When humans collect data manually, they see the weird cases. They notice when sources change. They feel the gaps. Automation removes that “pain,” and it also removes the little alarms that tell you your data is drifting.
To be fair, there’s a strong argument on the other side: this is exactly how you scale understanding. Plenty of teams don’t collect data at all because it’s too slow. They make decisions from vibes, not from evidence. If BigSet makes it cheap to create a living dataset, maybe we get fewer gut-driven calls and more testable claims.
I buy that. I just don’t buy the idea that the output will be treated as “testable” by default. Most people don’t test datasets. They use them.
So the stakes are pretty clear. If BigSet works, the winners are people who can move fast: founders, product teams, researchers, marketers, anyone trying to track a changing world without hiring a data team. The losers could be the people on the receiving end of those decisions—customers, workers, communities—when the dataset is wrong in a quiet, systematic way.
And the biggest risk isn’t one dramatic mistake. It’s a slow feedback loop where the dataset shapes decisions, decisions shape reality, and nobody can tell where the original assumptions ended and the “facts” began.
I don’t know how BigSet handles disagreements between sources, or how it flags uncertainty, or how often it gets schema guesses wrong, or what it considers “accurate and efficient” in practice. Those details matter more than the demo. They determine whether this becomes a tool that sharpens thinking or one that industrializes bad shortcuts.
If we’re going to hand people the ability to create “live truth tables” from plain English, what should be the line between “easy enough to use” and “hard enough to force you to notice what you don’t know”?