No smilies, no avatars, no flashing gifs. Just discuss the issues of the day, from last night's telly via football to science or philosophy.
Started by JDW70Wellwhynot on Sep 29, 2017 9:41:48 PM
Artificial Intelligence

A book reviewed in last Sat Guardian:

Life 3.0: Being Human in the age of AI by Max Tagmark

The reviewer writes. "Long before the appearance of super-intelligent computers, our society will be completely transformed by rather crude and dumb AI that is nevertheless good enough to hack humans, predict their feeling, make choices on their behalf, and manipulate their desires".

What is meant by AI and should we be worried?

upgoerfive - 29 Sep 2017 22:33:53 (#1 of 69)

He could have a point there - there are those who would argue that an army of rather crude Twitter posting bots helped to determine the outcome of the U.S. Presidential Election.

Bonusy - 29 Sep 2017 22:56:40 (#2 of 69)

Only by amplifying what other human operatives were doing. That's like attributing ability to a Marshall amp.

thisonehasalittlehat - 29 Sep 2017 23:08:13 (#3 of 69)

It'll be fine.

SirensOfAcreLane - 30 Sep 2017 00:07:52 (#4 of 69)

I think it depends on what the endpoint for AI is.

The differences I've read of previously, between a 'crude and dumb AI' and a more sophisticated AI suggests it won't be fine. It seems like most people adjust to a virtual butler/concierge/PA type of AI readily enough, Siri et al, but will they adjust to the kind of true AI superminds that could manage the global economy, law and justice, politics (!), transport, independent AIs plugged into the internet of things, and so on?

Stuff like this is a precursor, I guess:

Tripos - 30 Sep 2017 10:40:37 (#5 of 69)

No doubt the future will involve things like this:

HouseOfLammo - 30 Sep 2017 11:27:04 (#6 of 69)

Only by amplifying what other human operatives were doing. That's like attributing ability to a Marshall amp.

"The Edge" there.

carterbrandon - 30 Sep 2017 16:38:05 (#7 of 69)

Winces at irony of thread title.

inandout - 30 Sep 2017 17:01:03 (#8 of 69)

Computerphile has some interesting clips, Concrete Problems in AI safety.

Check out the Stop Button Problem and how you test for AI corrigibility (You can't, which is an issue...)

He also has some videos on his own channel discussing Reward Hacking and Wire Heading, and how General Artificial Intelligence might destroy us even whilst we enjoy it.

tasselhoff - 30 Sep 2017 17:08:11 (#9 of 69)

It's all logistic regression and neural networks these days. Blah blah sigmoid activation function blah blah loss function blah blah training sets blah blah

brooklyn - 30 Sep 2017 17:11:12 (#10 of 69)

heh. I remember "Magnus, Robot Fighter" from my boyhood. we'll never let them conquer us!,_Robot_Fighte

RosyLovelady - 30 Sep 2017 17:45:21 (#11 of 69)

I thought it kinder to say nothing about that, Carter.

Dementor - 30 Sep 2017 17:51:53 (#12 of 69)

A remake of Colossus: The Forbin Project has been on the cards for ages.

Old one's trailer:

RosyLovelady - 30 Sep 2017 17:53:17 (#13 of 69)

Colussus: The Corbyn Project

SirensOfAcreLane - 01 Oct 2017 12:18:36 (#14 of 69)

#9 Logistic regression, neural networks, and fuzzy logic were what compsci grads and others talked of in the 90s, don't recall sigmoid activation functions, had to look it up, hmm, is the intent to develop AI along human lines desirable?

Which begs a question; intelligent, or sentient AI?

JohnIlly - 01 Oct 2017 12:20:22 (#15 of 69)

Artificial Intelligence will at least be able to spell "intelligence".

Policywatcher - 01 Oct 2017 12:27:57 (#16 of 69)

Is it possible to have artificial intelligence without artificial intransigence?

SirensOfAcreLane - 01 Oct 2017 12:28:22 (#17 of 69)

#15 I've sent the mods a request to correct, should do the same for the missing r on my latest thread title, but time's up, oh well!

JohnIlly - 01 Oct 2017 12:31:17 (#18 of 69)

They've done it. The problem with thread title errors is that you can't change them after the first post.

tasselhoff - 01 Oct 2017 12:36:01 (#19 of 69)

There are other functions even better than Sigmoid that really help gradient descent adapting to training sets (such as ReLU) depending on the requirements.

tasselhoff - 01 Oct 2017 15:11:50 (#20 of 69)

Nice AI cheatsheets

Check Subscriptions
Home » IT & Computers