Facebook Messenger Assistant Powered by Humans | Hacker News Reader

Facebook Messenger Assistant Powered by Humans

recode.net

63 pointschris-at13 comments

> “The reason this is exciting is it’s scalable,” he added. “We cannot afford to hire operators for the entire world to be their personal assistant.”

Why all the praise when the goal will look nothing like it looks now. They don't actually know that it's scalable, otherwise we'd be seeing a completely non-human dependent AI with limited functionality (like ordering food) first. The current approach is just getting more data to see if the real version can work at scale.

> Why all the praise when the goal will look nothing like it looks now

Because the media machine needs to keep spinning. Rather than give a balanced statement now, which takes more effort to put together, it's easier to post "Wow, this is amazing" now and then "Wow, that was disappointing" later, and then get double the hits.

Relevant: https://medium.com/backchannel/how-the-tech-press-forces-a-n...

Reminds me of Google's 411 service back in the day. People thought "cool, Google is offering a free service". But Google was just collecting voice samples to train its voice recognition algorithms.

ok Google. Spy on all my activities. Sell them to corporations and governments.

>If a user asks A, you respond B.

A user may never ask A. Or, a user may only ask A once; in which case, did we really need to pre-compute the answer?

It's not pre-computing an answer, it's training a neural network that will be able to answer that and similar questions with minimal or no human interaction.

Whether that works out or not remains to be seen, but it's definitely worth investing research effort in to.

with 100s millions of user, each question will become a common question.

Questions like: what, why, when, how? They are very common, but don't have definite answers.

The problem is, even with 100s of millions (or even billions), there are still an infinite number of questions that can be asked. Each of those questions can't be common.

Classic Pareto setup (https://en.wikipedia.org/wiki/Pareto_principle)

We are still far away from having computer's truly be able to learn on their own, without having a human optimizing the responses.

> We are still far away from having computer's truly be able to learn on their own, without having a human optimizing the responses.

Could say the same about humans, though. We can't truly learn on our own without having others calibrating the process, or otherwise things set up by others.

"Supervised learning" makes me even more excited about M, because pure AI at this point won't be so useful.

AI in large part is supervised learning. I think the more important question here is when that becomes feasible for general purpose stuff for facebook.

Reply on news.ycombinator.com