Artificial intelligence still has some way to go

upper management in tech really has no clue what happens on social networks do they? like they have no idea what kind of people bubble up

wasn't there an instance a month or so ago where a twitter exec said something mildly stupid and got harassed for the next 48 hours solid? and he was stunned. he had no idea how his community platform functioned de facto.

what else? moot just got hired by google. were they like, "google+ is a failure, you made a playground for fascists and pedophiles, can you help us?"

i have a lot of trouble imagining my way into the techy view of the world

― goole, Thursday, March 24, 2016 1:23 PM (19 minutes ago) Bookmark Flag Post Permalink

I admit that I am more surprised that Microsoft would do this as opposed to, say, Coke's hilarious "Share a bottle with ______" campaign

http://arstechnica.com/information-technology/2016/03/tay-the-neo-nazi-millennial-chatbot-gets-autopsied/

Bot creators, especially those of interactive bots like Tay, say that they have to continually adjust their bots to keep them on the straight and narrow. Abusive users, and how these will be addressed, have to be considered right from the start. The Tay experience has made some from this group angry. Natural language researcher thricedotted, who has 37 bots (the best known of which, or at least, the only one that regularly gets retweeted into my timeline, is sexting bot @wikisext) told Jeong that "You absolutely do NOT let an algorithm mindlessly devour a whole bunch of data that you haven't vetted even a little bit." In other words, Microsoft should have known better than to let Tay loose on the raw uncensored torrent of what Twitter could direct her way.

I especially like the conclusion that the reason the Chinese equivalent didn't turn Hitler Youth is because the PRC's censorship regime makes the necessary circumstances all but impossible

Censorship pwns

http://www.newyorker.com/tech/elements/ive-seen-the-greatest-a-i-minds-of-my-generation-destroyed-by-twitter

What was astonishing about that victory was, in part, how quickly AlphaGo became an expert. In five months, it reviewed and played more matches than most humans could in a lifetime. Tay appears to have accomplished an analogous feat, except that instead of processing reams of Go data she mainlined interactions on Twitter, Kik, and GroupMe. She had more negative social experiences between Wednesday afternoon and Thursday morning than a thousand of us do throughout puberty. It was peer pressure on uppers, “yes and” gone mad. No wonder she turned out the way she did.

Memristor-based tech can provide huge gains in learning speeds:

http://www.tomshardware.com/news/ibm-chip-30000x-ai-speedup,31484.html

Tay has returned! ...And is gone again.

http://www.theguardian.com/technology/2016/mar/30/microsoft-racist-sexist-chatbot-twitter-drugs

Genuine lols, I hope she keeps being resurrected for more and more madness at irregular intervals

There are definitely aspects of the Tay situation that aren't funny at all (teaching it to harass people = horrible) but the general "let's throw an overly-receptive chatbot at Twitter and see what happens OH NO SHUT IT DOWN" vibe of this has me rolling

it's hilarious these brilliant people can design complex artificial intelligence and yet they seem clueless about the real world environment of Twitter

who ever would have thought that nerds might be bad at dealing w reality

certainly no one who has posted on ilx

idk, i feel like the people programming it and the ones unleashing it/doing public relations are two different groups

the engineers are probably all "whatever, other groups will fuck with it and it'll gain variety eventually"

everyone else is "goddamn it the bot loves hitler again"

OTM, also lol

http://www.nytimes.com/2016/04/04/technology/chinas-companies-poised-to-take-leap-in-developing-a-driverless-car.html

on one hand, there is an advantage in working with a government that can impose regulations and allocate funding to promote autonomous vehicles. on the other hand, designing a driverless car that can operate safely in big cities in China is craaaaaaaaazy. i suppose if they manage to pull it off there, they'll pretty much be able to do it anywhere.

well they could also set up an infrastructure that would be impossible anywhere else; something that has, say, roadside reflectors as a guide

I’ve spent the past year working on computer vision problems. And I’m now skeptical we’ll see the proliferation of autonomous vehicles (i.e. vehicles without manual steering) soon. I’d bet the public’s widespread belief that autonomous vehicles are forthcoming will be responsible for the next AI winter.

I am also skeptical they will soon solve the problem of self-driving vehicles sharing high speed roads with ordinary drivers. Currently, all the self-driving vehicles operating in traffic are on roads where speeds are 35 mph or below.

They probably won't, but ordinary drivers should be banned anyway.

extraordinary drivers only

xxp which problem in particular do you think is too difficult at "high" speeds? It seems likely that they test at slow speeds because the cost of a fuck up is a lower. There may not be a processing/reaction time limitation when moving from 35 to 75 mph. The difference may be inconsequential.

The problem of mixing high speed self-driving vehicles with regular high speed traffic is not the processing speed or reaction time of the self-driving vehicle, but the high percentage of drivers on high speed roads who both create and accept high risk situations, so that in most US population centers the majority of vehicles on high speed roads are spaced too closely for safe stopping in an emergency situation.

I understand that self-driving vehicles would still be safer than regular vehicles in these situations, because their reaction time to the vehicle ahead will be at least a second faster than a driver could react, so in that respect the problem is not so much a technical one as a question of liability and public acceptance, especially in fatal collisions. So, when you say that slower speeds reduce the cost of a fuck up, you have to include serious injury and death as potential costs, and these are far more likely at high speeds than at low speeds. That's an enormous difference.

I would also note that a human who practices good, defensive driving techniques is constantly looking well past the vehicle ahead, noting the behavior of vehicles in all lanes of traffic, including parallel lanes, intersecting roads and driveways, and using sophisticated predictive heuristics to evaluate risks and react much sooner to developments than just sensing the speed changes of the vehicle immediately in front. I'm not sure how well current self-driving cars are able to model these techniques.

If the several companies experimenting with this technology have good answers to these problems, I would be happy to hear it, since those companies are currently pushing hard at the political end of things to allow this technology much sooner than later. If it's going to happen anyway, I want it to be as safe as possible.

etaeoe, can you talk more about your work? I'd be curious to hear from someone actually working in the field as to what the real pitfalls are.

xp the serious injury and death liability question is definitely an interesting one. I wonder what kind of lifetime the self-driving data will have. I suppose they will rely heavily on cameras for accident liability stuff?

I'm not sure, but I suspect that the self-driving cars will possess some kind of radar. That alone should give the car potentially better cross-lane and up-ahead awareness than a human. You can then add communication with other self-driving cars, and data pulled from the sky. So I do think you're underestimating the possibilities if you assume a self-driving car is only "sensing the speed changes of the vehicle immediately in front."

I think you are correct that there are difficult AI problems, such as classifying another driver on the road as dangerous, which a human can possibly do better, though.

etaeoe, can you talk more about your work? I'd be curious to hear from someone actually working in the field as to what the real pitfalls are.

I’m a member of the Broad Institute of MIT and Harvard. I work on computer vision problems. My recent work uses convolutional neural networks to solve object recognition problems in biology, e.g. “is this cancer?”

I took a poll at our group meeting last week about the feasibility of autonomous vehicles and few believed it’s possible with existing (computational and statistical) theory or (software engineering) methods. However, one optimistic member of our group is leaving to work for Mercedes-Benz on this problem (they use CNNs too). So, we’ll see!

I see three major problems:

* Garbage in, garbage out—sensors and cameras suck! I can’t reliably segment (i.e. isolate or extract) cells (i.e. uniform blobs) from static backgrounds from images that were captured with the world’s most sophisticated microscopes. We need better acquisition for more reliable predictions.

* While learning methods can reliably solve common object recognition problems, they can’t solve “scene understanding” problems, e.g. a model may predict an image includes a “cowboy” and a “horse” but it cannot understand the relationship between the two (e.g. “roping a steer”). This is the major open problem in computer vision and I think it’ll be fundamental to autonomous vehicles (and other difficult machine vision problems).

* Software is too hard to write. Autonomous vehicles will likely require complicated parallel programs that must be bug free. Compilers (and static analysis) need to advance.

I should add that I think these problems (and others) are solvable. It’ll just take some time! I should also add that I’m usually wrong when trying to predict this stuff! :P

I'm not sure, but I suspect that the self-driving cars will possess some kind of radar. That alone should give the car potentially better cross-lane and up-ahead awareness than a human. You can then add communication with other self-driving cars, and data pulled from the sky. So I do think you're underestimating the possibilities if you assume a self-driving car is only "sensing the speed changes of the vehicle immediately in front."

I’ve been told that current self-driving programs rely mostly on non-camera sensors (e.g. GPS, Lidar, Radar, etc.). I think the technique is:

• follow a GPS route; and

• don’t bump into stuff

I wonder if you ever ran into my ex-roommate's work, he did video image recognition in colonoscopy footage to attempt cancer detection. Worked with the Mayo Clinic, iirc

world's most sophisticated microscopes = normal imaging or something like oct?

xp innerestin'!
your notes about the lack of sophistication of cameras and sensors brings to mind the old anti-atheist "miracle of the human eye" saw; i assume there must be some research into a completely reimagined method of acquiring imagery based on liquid or gel lenses?

http://www.nature.com/news/octopus-genome-holds-clues-to-uncanny-intelligence-1.18177

gimme an octopus-style neural network

eh, it will just run away to the ocean the first chance it gets

I see you have read about Inky's adventures, too

http://i.imgur.com/qTQOWYC.png

https://www.captionbot.ai/

looooooooool

i think AI has arrived

Awesome

https://pbs.twimg.com/media/CgRG4gqXEAEQlB8.jpg

that's clearly the guy from interpol

ccaaaaaeeeeeekkkk

http://www.nytimes.com/aponline/2016/04/19/us/politics/ap-us-aviation-bill-senate.html

re: the amazon delivery drones, the senate passed a bill

WASHINGTON — The Senate approved a bipartisan aviation policy bill Tuesday that would boost airport security, extend new protections to airline passengers and help speed the introduction of package-delivery drones.

...The bill also would remove obstacles to commercial use of drones while enhancing privacy and safety protections. It requires that within two years the FAA authorize package deliveries by drones. The agency would create a small drone "air carrier certificate" for operators of delivery drone fleets, similar to the safety certificates granted to commercial airlines. The rules are needed for Amazon and other companies to deploy fleets of delivery drones.

Another provision would establish criminal penalties for the reckless use of drones, aimed at penalizing operators who fly drones near airports without prior approval. Another provision requiring the TSA to restore passenger screening at small airports where airline service has been reduced would force the agency "to reallocate staff and equipment from higher-risk, higher-need facilities," the White House said in a statement.

The White House criticized the bill's delivery-drone language as "overly prescriptive" and said it would disrupt the agency's ongoing efforts to write safety regulations for commercial drone flights.

I don't get the point of drone delivery. It seems like they are looking for a more efficient way for people to lose their packages.

i dunno, UPS is doing a good job of making it look good. two days ago they tried to deliver a wedding dress to our apartment, but we missed them by 15 minutes. so instead they left a message saying they'd drop off the package to a random cell phone store about a mile away (?!). the next day we went there at noon, but the store was closed (because it's a shitty random cell phone store). we went again later in the day, but the package wasn't there. the UPS tracking said that the package was en route to the auxiliary dropoff point (?). today the tracker said that it had been delivered - to the shitty random cell phone store. which may or may not be open.

so i don't know, i get the appeal of a little flying thing landing at your front door. it's all a moot point for me, though, because from what i've seen it seems tailor made for the suburban lifestyle - big grassy front yards and lots of space. i can't see a tiny drone landing next to a city apartment anytime soon.

Well see, UPS is showing you a very inefficient way of losing a package. The Amazon model loads the package into a drone, it flies away, and then *poof* package gone.

but imagine the poignant moment as you run outside and catch just a faint distant image of the drone flying into the sunset, with your errant package dangling precariously by one little claw

I have a really big tree in my yard. I keep wondering if someday in the future, I will see some amazon drones in it

drone delivery strikes me like VR and not-strictly-necessary bionic augmentation: thanks for the offer, I'll check back in 20 years once you've worked all the kinks out

kinky drones, now we're talking

but imagine the poignant moment as you run outside and catch just a faint distant image of the drone flying into the sunset, with your errant package dangling precariously by one little claw

i guess late stage capitalism does have a little romance to it.