andrewducker (![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png)
andrewducker) wrote2018-01-21 12:03 pm
andrewducker) wrote2018-01-21 12:03 pm
Entry tags:
- academia,
- ai,
- anger,
- animation,
- argument,
- books,
- censorship,
- clothing,
- comic,
- communication,
- disney,
- douglasadams,
- drwho,
- epicfail,
- ethics,
- fraud,
- freespeech,
- funny,
- hacking,
- history,
- holland,
- india,
- language,
- liberal,
- links,
- magnets,
- mars,
- men,
- menstruation,
- money,
- movies,
- myths,
- ohforfuckssake,
- politics,
- republicans,
- review,
- security,
- space,
- uk,
- usa,
- viajennierigg,
- women
Interesting Links for 21-01-2018
- Why King of Bollywood Akshay Kumar Made A Movie About Menstruation
- (tags: India menstruation movies )
- Dutch police are confiscating expensive clothes from kids who look too poor to wear them
- (tags: clothing money Holland OhForFucksSake )
- WE! DO NOT! TALK ABOUT! THE ORANGUTAN!
- (tags: argument academia funny viaJennieRigg )
- Ethics, artificial intelligence and the Lovelace Oath
- (tags: ai ethics )
- School for Sirens
- (tags: women men myths comic )
- What, exactly, is no platforming?
- (tags: censorship freespeech language )
- Where The Shutdown Blame Lies
- (tags: politics usa republicans )
- How to communicate with angry people
- (tags: anger communication )
- Doctor Who and the Krikkitmen
- (tags: drwho review books douglasadams )
- British 15-year-old gained access to intelligence operations in Afghanistan and Iran by pretending to be head of CIA
- (tags: hacking fraud usa uk security epicfail )
- Liberalism, Centrism, and the Difference
- (tags: politics liberal uk )
- East Coast vs West Coast - how Dumbo got its psychedelic sequence
- (tags: animation Disney history )
- Why is only half of Mars magnetized?
- (tags: mars magnets space )
Lovelace Oath
I get the impression from chat around the office that the deep learning tools handling very large data sets can't tell you why they "know" what they know. I think a lack of transparancy might be fundamentally linked to how some of these systems work. So I'm not sure it's entirely possible to apply the Lovelace Oath and always use something that has the required level of transparancy without losing significant amounts of functionality.
Not having that functionallity might be worth having the transparancy. It may be possible to back fill the transaparancy and achieve the goals some other way.
Re: Lovelace Oath
Correct. Deep learning tools work by twiddling thousands of dials to maximise how much the results line up with reality. They have no understanding of the interaction of the various knobs, or why they should lead to accurate results.
So, yes, this will definitely lead to problems. And the only way to try and work around this is to test the models against a really large subset of reality to see whether it does things you don't expect.
Re: Lovelace Oath
You can test in simulation (e.g. run your driving software though several million hours of simulated real conditions and see if its Desired Outcome for Accidents Arbiter is working as you would like) but you have a risk that your simulations aren't very good.
A larger risk is that we are trying to create dumb but powerful Gods using 21st Century capitalism.