Model Hub

Aisles split by what the model does

Walk into a hub and the first split is by job. A text aisle, an image-making aisle, a sound-recognition aisle, and within the text aisle, models tuned to summarize well sit apart from ones tuned to hold a conversation well.

Filter by what you're trying to do instead of by name. Set conditions one at a time — language, file format, terms of use, size — and the candidates thin out fast. Sorting by most downloaded isn't a bad place to start either.

A hub isn't only models. Bundles of training material and demo screens you can press right in the browser often sit in the same aisle.

The same job comes in several sizes

Within one aisle, boxes with the same name sit stacked by size. A bigger box gives smoother answers but demands a bigger machine. A smaller box runs on a laptop or a handheld device, but wobbles on hard problems.

A version squeezed down to shrink its footprint sits nearby too. Shrink the weight and it fits on your device and picks up speed, but the finest details blur a little.

Before you load one up, check the room on your own device. Skip checking the file size and the memory it needs, and you only find out it won't open after the whole download is done.

Read the label before you load it

The label on each box says what it learned from, what it's good and bad at, and what terms you can use it under. A box with no label, or just a couple of lines, is a sign that less care went into it.

Check who posted it too. Anyone can stock this warehouse, so something from a well-known source sits right next to something from an account created yesterday. Download count, the last time it was touched up, and comments people left all help you judge.

Worth knowing too: the file itself can be risky. Some formats let executable code hide inside a model file, so the trend these days is toward picking formats without that risk.

After you take one home

The simplest path is using it as is. Load it and call it, and most jobs stop right there.

Next is lightly reworking it with your own material. Tune it to your company's tone or your field's vocabulary, then post the result back to the hub for someone else to use.

Last is pressing the buttons before you take anything home. A few sentences typed into a demo screen reveal a model's personality fast — like handling a sample before downloading a file that runs into gigabytes.

Don't decide off one box either. Load two or three similar candidates and run the exact same question through each — differences the descriptions never showed jump out immediately. If you plan to work in a language other than English, test it with sentences in that language specifically.