Artificial intelligence still has some way to go

Artificial Intelligence still has some way to go

http://arstechnica.co.uk/the-multiverse/2016/06/sunspring-movie-watch-written-by-ai-details-interview/

It says something that the pop song written by that program was vapid, but almost passable, while the script, where humans must talk to one another and act like humans, sounds barely coherent.

https://www.teslamotors.com/blog/misfortune

but it's so much more fun to paint this as a robocop boardroom demo and draw snide conclusions about how AI still has a long way to go

http://yaroslav.ganin.net/static/deepwarp/

lol, a literal RMDE program

The first program that can roll its eyes at how bad it is.

paging gr8080

http://i.giphy.com/3o6ZsSi98Ugr8mUk9y.gif

This work seems to be aimed pretty directly at creating photo-realistic-animation software. This is the sort of 'artificial intelligence' that has a viable path to its goal, but it is hardly the 'strong ai' that gets all the think-pieces.

also totally pointless/of no actual use to humanity

[finds pic of shakey on wdyll thread, runs it through rmde engine, posts it to thread]

good luck with that

xxxp I don't know. If you add 'artificial intelligence' to your google news right now, for example, you'll get two think-pieces about how emerging interest in AI is an opportunity to bridge the STEM gender gap, a few articles about reasonable AI applications, and maybe one article about someone writing a movie script with the assist of AI.

guys i'm worried this has no use to humanity

that's what they said about ice

Just to be clear, I posted this on this thread because it uses neural networks. Rmde at the ai police.

tbh a "best and creepiest siggraph demos" thread would be pretty grand

https://twitter.com/FioraAeterna/status/758329754386694144

feel like a grad student in...anthropology? might have a field day with the whole "underpaid mechanical turks do stuff while devs work on the AI" thing

or, like, a jacobin thinkpiece

all that stuff is mostly smoke and mirrors that the VC community is mysteriously (well not that mysteriously, they're idiots) buying into

non of these concierge type startups in w'ell-turk-it-for-now-and-use-that-to-train-machine-learning-in-parallel are making any progress whatsover on the training, afaict.

if any companies get anywhere it will be because of structural advantages in training data access, i.e. it will be google or facebook or whoever.

but if you can exit one of those chatbot/turk nonsense startups in the next year or two before the penny drops, good for you.

in the meantime of course the turkers are treated like garbage, but they're organizing (cc hoos) http://wiki.wearedynamo.org/index.php/Guidelines_for_Academic_Requesters

eye manipulation needs a lot of improvement

there are a lot more muscles that move when we move our eyes

fwiw that is very significantly better than what's gone before, but the point of that study is to demonstrate a method that involves knowing nothing about physiology. there are no hard-coded heuristics about how the human body works in there. it's all learned by a machine from examples.

i didnt read the documentation and just scanned the text

thats pretty cool

i guess correcting googly eyes on photos in post-shots could yield interesting results

When I was a tech writer I remember a colleague of mine (who knew I dabbled in writing chatbots) asked me what I thought about AI and I remarked offhand that I was convinced that most of the intelligence was embedded in the data structures.

I still think this remark has validity and that neural networks and machine learning are mainly about giving machines the ability to evolve and refine their own data structures. If the newer structure fits the given task better than the unevolved structure, then the machine has gained in intelligence. Of course, real AI researchers are creating applications several orders of magnitude more sophisticated than my poor little chatbots were and could talk rings around me on this subject.

yeah it's interesting that "AI" in the sense that it's discussed in, e.g. the norvig book, has been pretty much abandoned, or at least has a much diminished profile.

this is to the extent that "AI" has almost become one of those magic phrases (like "big data") whose use allows you to quickly spot people in the tech industry who don't know what they're talking about

https://techcrunch.com/2016/08/24/why-ai-consolidation-will-create-the-worst-monopoly-in-us-history/

Interesting topic, but: "monopoly" usage fail; and the idea that "Alphabet, Amazon, Apple, Facebook and Microsoft" owning everything in IT is bad for the rest of us is, hmm, how do you say, not news

yeah. i think the article doesn't do a good job of explaining why a concentration of "AI" (which for now means machine learning) power would be different from previous examples. i think it would though.

Can't wait for SAP and Oracle to buy AI

There's a scenario where a DeepMind-style AI takes over *every industry* but so far there are vast swathes of business where Google/Microsoft/Apple/etc have no presence besides search engines and mobile phones. Small/medium/large AI companies can and do thrive without heading towards an inevitable acquisition. Finance is one example - quant trading is basically ML.

I want to teach a robot LEAN manufacturing principles and really let em rip

xp also there is a ton of AI value in proprietary (and in some cases pre-internet) data

Personally I'm waiting for expert systems to really take off

Also the whole article completely fails to mention Watson or IBM which continue to plague my browsing with their ads for cognitive this and that

re: watson https://twitter.com/minethatdata/status/559177526740652035

I tried reading Nick Bostrom's Superintelligence the other week and it was like reading an academic paper composed in the style of an instruction manual. I'm glad I read the New Yorker article on him and AI months ago, otherwise I wouldn't have even been able to make it as far as I did!

yeah i've picked at it a bit over the last few months and it's slow going

It's not quite a monopoly if five different companies are competing...

AI/ML seems like a perfect storm of all the things that existing IP law can't deal with. Decades-long shitstorm of lawsuits on the horizon.

I interviewed for a job at a start-up the other day, and at some point one of the co-founders asked me if I "believe in machine learning." I correctly inferred that he wanted me to answer enthusiastically in the affirmative, so I resisted trotting out my usual dinner party line, that most of the cutting-edge shit that gets discovered by computer scientists ends up being shown to be equivalent to boring old statistics.

i have had that question

In a conference call with reporters, the head of Tesla, Elon Musk, said he believed the upgrades would have prevented the accident on May 7.

“These things cannot be said with absolute certainty, but we believe it is very likely that, yes, it would have,” he said. The new version of Autopilot, with its improved radar, “would see a large metal object across the road” and be able to determine that the object is not an overpass or overhead road sign that poses no threat, he said. “Impact probability would be assessed as high and it would probably brake.”

One major change is that improved radar becomes Autopilot’s main system for scanning the road. Once the update is made, Autopilot will use images from cameras to supplement the radar system. The current system uses cameras as its primary source of images, and relies on radar to help confirm what the cameras see.

http://www.nytimes.com/2016/09/12/business/elon-musk-says-pending-tesla-updates-could-have-prevented-fatal-crash.html

"probably". lol

horseshoes, hand grenades and self driving cars

i'm giving a talk to a hedge fund about recurrent neural networks tomorrow :-(

sounds like a good working definition of hell

so....why were they not using cameras in the first place?

this seems like a bad idea:

https://www.washingtonpost.com/business/economy/for-some-safety-experts-ubers-self-driving-taxi-test-isnt-something-to-hail/2016/09/11/375f980a-769a-11e6-be4f-3f42f2e5a49e_story.html

good luck Pittsburgh

my local university's engineering dept has a self-driving car project which they've been working on for a while now, since before I heard of any others iirc (or maybe about the same time as I first heard of Google's)

I've seen it a couple of times, going forwards and backwards in a straight line in an empty car park, and then stopping while people with clipboards take lots of notes

if the competition is way ahead of where they are I feel sorry for them, getting in early but just not moving fast enough

if the competition is not way ahead of where they are I feel sorry for humanity that a whole bunch of companies are now going "sure fine, just put 'em on the roads already". good luck everyone

i think the public reaction to the tesla death not making a bunch of people cancel their pre-orders or demolish stock price has provoked a better to ask forgiveness than beg permission attitude. First time one of these things go onto a sidewalk and wipe out a kindergarten class, I expect congressional hearings and a 10 year pause in development.

traditional automakers have spent a lot of time trying to determine the best way to market smart safety features to consumers as things that help you avoid accidents and keep you safer if something dangerous happens. it's a balancing act between telling people that driving is an inherently dangerous act (which people do not want to hear or acknowledge) and introducing things in ways that make them seem relatively natural and not disruptive to normal driving

tesla hasn't necessarily marketed their vehicles this way, but the general direction of press is "this thing practically drives itself, but sometimes you have to take control" which has resulted in some knuckleheads letting the car do most of the work, and when they need to take action, they're not paying attention. the self-driving tech is good, but... not that good