4 ms·
For me, the "smoke under the door" moment was Karpathy and Li's Deep Visual-Semantic Alignments for Generating Image Descriptions [1]. The almost perfectly gram
by Tossrock 9y ago
For me, the "smoke under the door" moment was Karpathy and Li's Deep Visual-Semantic Alignments for Generating Image Descriptions [1]. The almost perfectly grammatical machine-generated captions of photos were unnerving to me in a way that simple categorization was not. It somehow called to mind the image of a blank-eyed person speaking in a monotone while images flashed in front of them. What if they wake up?
[1]: http://cs.stanford.edu/people/karpathy/deepimagesent/ http://cs.stanford.edu/people/karpathy/deepimagesent/
- psyc 9y agoMicrosoft's Seeing AI app does this fairly well.
- Retra 9y agoWhat is "wake up"? How can a machine "wake up" in a way that we can't shutdown with a trivial disconnect? Our computer systems are exceedingly fragile and bottlenecked. No runaway bandwidth-hogging super intelligence is going to be able to 'wake up' without pissing off someone sitting next to a power switch.
- UncleMeat 9y agoYet Andrej himself thinks that RNNs are not really making meaningful progress towards AGI.
- Tossrock 9y agoWell, if I were to play the Yudkowsky's Advocate here (which in general, I'm not), I would say that this is precisely what he's talking about in the article. Because Karpathy knows how hard he's had to work, and how flawed and dumb in particular ways the current techniques are, he may overestimate the distance to AGI. Now, generally I disagree with Yudkowsky on a lot of points, but I do think he raised some decent ones here.
- rspeer 9y agoThese systems are lining up common phrases that appear in corpora of image captions, with visual patterns that they can be trained to recognize such as color, textures, and shapes. Nothing is "waking up". Your astonishment at what these systems can do tells me that you may have looked at cherry-picked positive results. So here's an article I found that cherry-picks negative results instead: [1] [1] https://gizmodo.com/this-neural-networks-hilariously-bad-image-descriptions-1730844528 https://gizmodo.com/this-neural-networks-hilariously-bad-ima... Now of course this article is exaggerated too. Ideally, if a system is 95% accurate, you'd be looking at representative output from the system, with 95% good results and 5% bad ones, perhaps by running such a system yourself on a different set of images.