[jareklupinski]

it's this part:

> latency matters more than raw accuracy – think industrial inspection

it (rightfully) raises red flags in anyone when you hear someone confidently claim raw accuracy is _not_ important in things like _inspection_

[IanCal]

They didn't say it wasn't important they said latency was more important, and they're right for many use cases. Once you can't run at realtime where you're operating, you need to move to batching or offloading the work to a pool of workers and handling more async issues. You can no longer have something that shunts the component off to another track where your camera is, you need to have the camera somewhere else then 40s later pull it out of another location. You need good networking so you can fire off images to get processed elsewhere. That's also a bunch more systems to maintain.

These things aren't impossible of course but it's additional management over "place the device here".

Here's how you know that accuracy isn't the be all and end all of the discussion - we already deploy systems with less than human accuracy to monitor things, and when we use humans we very rarely inspect every single item. So there must be a tradeoff we're happy making in lots of industries.

Even if you're focussed on not missing anything, lower accuracy that comes at the cost of more false positives can be massively useful as you can then do a two step process (even with humans as the second step if you need). The goal of the first step is to ignore the 99% of totally fine items so you spend the costly process on just 1% of the items.

[jareklupinski]

totally

but i wouldnt stand on a soap box and yell that to the world without all that ^ nuance

but by the time i'm done with all that, i'm only preaching to the choir

[ctparker]

That is definitely the right part. The dash isn't a symbol on a normal keyboard, and "think blah blah blah" occurs frequently in LLM chat sessions (for me, at least). I suspect those easy-to-spot indicators won't be around forever, which will make AI posts much more difficult to spot. But I think the thinly veiled advertisement that follows in that clause will be the bigger tell in future models. If we feel like we're being marketed at, we can almost guarantee there isn't a human on the other end. This isn't the internet I signed up for.

[WithinReason]

they might be referring to using a quantised version which gives them high performance and the accuracy drop is less important