Hacker News Reader

Comment by seattle_spring | Hacker News Reader

Did anyone honestly think it was anything other than humans manually entering data from the submitted images?

> Did anyone honestly think it was anything other than humans manually entering data from the submitted images?

They can have humans entering data manually but have safeguards in place to protect data - usually done by either hiring people in-house or having a contract with a firm that employs a regular set of people to do this work.

Sending the data to Mechanical Turk implies basically none of that safeguarding. Yeah, it's possible to do it, but from the sounds of it - especially given Expensify's appalling response - it's pretty apparent that they didn't.

lol so this is like the modern version of the actual, original namesake mechanical "Turk" from the 18th centruy?[1]

"Um, you pretty clearly have some human wearing a costume and playing chess."

'No, no, I insist, it's 100% machine that happens to be good at this task!'

Now it's:

"Um, you pretty clearly must be using humans to interpret the receipts."

'No, no, I insist, it's all done by our proprietary algorithms!'

[1] https://en.wikipedia.org/wiki/The_Turk

Can't wait for the Show HN of a mechanical turk backed chess bot.

That would actually be interesting! Humans are good st certain kinds of computational problems like approximate Traveling Salesman, so if you had agood algorithm that assumed access to such an oracle, you could indeed power it with Mechanical Turk workers in a nontrivial way.

> Humans are good st certain kinds of computational problems like approximate Traveling Salesman

Are humans any better at it than computers? I wouldn't think that humans have a good enough heuristic that the extra 1000x processing speed isn't more of an advantage.

Yes, that's why the system is called Mechanical Turk.

The modern version is more depressing. Companies let humans do the work, while using the proceeds to develop AI solutions which eventually put those people out of work.

Nope. I remember when MTurk came out, used the API and tried some of the work units. For the most part, it really is moot though for most tasks now, except for insanely complex ones or new business processes that haven’t been automated yet. This decline of usefulness is due to the fact recent AI/ML/DL, after little training, can complete mundane, repetitive tasks faster and with far better accuracy and precision than all humans.

I disagree. There are still lots of tasks out there that fit MTurk. Take something like "look at a picture and determine the person's race". Costs five cents for a Turk to do that. You can certainly train an RNN plus a Bayesian network to do this task, but a) there's no public corpus of training data for this so you'll need to generate that anyway and b) this solution still isn't thaaaat accessible to startups that have 1000 other things to worry about. How much time would the engineering team have to spend training, testing, tuning, validating, and deploying the model? If you only ever need to tag 1M images, that $50k on Turk is probably well-spent over the ML solution. If you need to tag 1B images, that's a different story, but if you're looking at that scale you probably already have the resources to do it.

There's already a great variety of cheap demographic recognition API's readily available... https://deepai.org/machine-learning-model/demographic-recogn... is just one.

Race is a cultural construct.

Quick, what's Barack Obama's race?

(Hint: your answer will depend on your culture and what you know of his ancestry and upbringing.)

This literally has nothing to do with what OP was saying.

You took one word out of his entire point and decided to make a statement about that.

The problem is not having humans in the mix, it's that access controls were basically zero. Anyone could go to the site and get personal data in moments, and all the entry was being done in insecure environments by unknown random people.

I assumed it was similar to Evernote's OCR...unless...Evernote was/still is also using Mechanical Turk..

Reply on news.ycombinator.com