This article takes a less alarmist tone, including excerpts of the alleged "language" which it describes as "simply modified human language"
> Bob: "I can can I I everything else"
> Alice: "Balls have zero to me to me to me to me to me to me to me to me to"
It also claims "it was shut down because it was doing something the team wasn't interested in studying - not because they thought they had stumbled on an existential threat to mankind."
(Epistemological status: I've only read this article and the submitted article on the topic.)
Most likely it was a bug, and since they used DNNs they could not correctly explain why it's happening. Since they could not fix it, they had to shut down the service.
AI is not taking over - nothing to see, please move along.
This story has really gotten legs somehow. The research is related to work with AI agents developing a trade/negotiation logic with shorthand language [0], but the media seems to have spun it into a tale of rogue computers being shut down in the nick of time. There have been some articles attempting to roll back the sensationalism [1] [2], but most of the headlines are still working the "scary AI needs to be shut down" angle. The actual paper in [0] is a very interesting read, and recommended if you want to get some idea of what the kernel of this story actually is.
This whole situation is almost completely falsified and fabricated. It truly is a great example of how misinformation can spread when a lack of journalistic integrity is coupled with irrational fear and bias. There are clear risks to using A.I. at the moment, but they all stem from pre-existing prejudice in training data. To really show how misleading this story is, here's the reaction of a respected and brilliant AI researcher: https://twitter.com/egrefen/status/891362804074033152
Lets not forget how much Facebook boosted PHP's reputation by sending patches at all levels of the stack and developing excellent patterns of discipline, code conventions and style that manage to avoid the worst attributes of the language.
For anyone interested, this is a sad running joke of a headline in the DL community.
All that happened was that the models created an intermediate representation (think a latent space like an encoder decoder model) as intended by the researcher and then for whatever reason shut it down, presumably because it failed to get good results.
Then some jackass reporter decided to get some clicks by making a click bait title.
13 comments
[ 3.0 ms ] story [ 40.3 ms ] threadThis article takes a less alarmist tone, including excerpts of the alleged "language" which it describes as "simply modified human language"
> Bob: "I can can I I everything else" > Alice: "Balls have zero to me to me to me to me to me to me to me to me to"
It also claims "it was shut down because it was doing something the team wasn't interested in studying - not because they thought they had stumbled on an existential threat to mankind."
(Epistemological status: I've only read this article and the submitted article on the topic.)
AI is not taking over - nothing to see, please move along.
For more perspective on what the engineers intended: https://code.facebook.com/posts/1686672014972296/deal-or-no-...
[0] https://arxiv.org/abs/1706.05125
[1] http://gizmodo.com/no-facebook-did-not-panic-and-shut-down-a...
[2] http://www.bbc.com/news/technology-40790258
All that happened was that the models created an intermediate representation (think a latent space like an encoder decoder model) as intended by the researcher and then for whatever reason shut it down, presumably because it failed to get good results.
Then some jackass reporter decided to get some clicks by making a click bait title.