Imnimo

53 ms·

751.
by Imnimo 4y ago
I think this is a step in the right direction, but the performance on most tasks is only mediocre. The conversation and image captioning examples in the paper are pretty bad, and even on some relatively simple control tasks it performs sur
752.
by Imnimo 4y ago
Sure, I just mean that the line is a lot fuzzier than most of the other examples. Like if you showed me just the modern Polish map in the first one, I could clearly draw a pretty accurate approximation of the historical German border. But
753.
by Imnimo 4y ago
Not sure I totally buy the Mexico one. The correspondence just isn't as strong as in some of the other ones, and in some places is confounded by modern state borders.
754.
by Imnimo 4y ago
I'm somewhat skeptical of attempts to play out the game theory of what the leaker's motivations would have been. While I generally agree that the author has identified the correct incentives, I'm not so convinced that we can
755.
by Imnimo 4y ago
>They could instead have said that. "Most Salvadorans have switched to a different bitcoin wallet." Is that actually true? The article only says that "many living in tourist hot spots on the Pacific coast". "Ma
756.
by Imnimo 4y ago
Agree, I felt like I knew what to do for some of them because of programming experience. I remember taking similar tests as a child, and always assumed that the pattern must be of the form "each object in the first square moves to som
757.
by Imnimo 4y ago
It would be an interesting experiment to see where along the path some of these details are lost. For example, in the experiment where they ask for a pear cut into 7 slices, and Dall-E often gives the wrong number. First check if you can
758.
by Imnimo 4y ago
Yeah, I didn't realize there was that additional prompt. It certainly makes more sense that if the prompt includes the description that the agent is playing the role of an AI, it would be able to deduce that it would not be affected.
759.
by Imnimo 4y ago
Another great historical fan-fiction genre I enjoy is the Phantom Time Hypothesis: https://en.wikipedia.org/wiki/Phantom_time_hypothesis It's obviously completely fake, but I still enjoy reading about it.
760.
by Imnimo 4y ago
I feel extremely confident that this offer would not encourage me to take any action on Elon's behalf. I would likely just stop using Twitter if my account got transformed into a scam coin.
761.
by Imnimo 4y ago
The last answer on the Stroop test video (very bottom of the page) is interesting. The system is asked how humans perform on the test, and (kinda correctly) replies that they are slower "when the color of the word and the color of the
762.
by Imnimo 5y ago
Yeah, I'm not sure it would work the same as current SEO spam, but I just feel like some portion of this idea is predicated on mis-aligning incentives. Right now, SEO has no incentive to go after Reddit comments, so Reddit comments ar
763.
by Imnimo 5y ago
How concerned should we be that this only works because currently SEO has no incentive to target Reddit or StackOverflow, and if this takes off, those communities will be overrun by SEO spam?
764.
by Imnimo 5y ago
This is equally an app that allows those countries to search people's social media presence to find reasons to arrest them! :P
765.
by Imnimo 5y ago
I believe those things would be within Twitter's rights, but would object to them as being bad ideas that would make me not want to use the platform. That isn't the same as saying I think Twitter doesn't have the right to do
766.
by Imnimo 5y ago
I think that in CLIP+VQGAN, the VQGAN model is frozen, and what you do is start from a random latent code, generate an image, pass it to CLIP, and the backprop through CLIP and through the VQGAN generator to figure out how you should move t
767.
by Imnimo 5y ago
Yeah, my comment didn't really do a good job of making clear that distinction. Obviously the details are pretty technical, but maybe I can give a high-level explanation. The previous systems I was talking about work something like thi
768.
by Imnimo 5y ago
I recommend this blog post: https://lilianweng.github.io/posts/2021-07-11-diffusion-mode... Personally, I find the core diffusion papers pretty dense and difficult to follow, so the blog post is where I'd begin.
769.
by Imnimo 5y ago
Here is my extremely rough ELI-15. It uses some building blocks like "train a neural network", which probably warrant explanations of their own. The system consists of a few components. First, CLIP. CLIP is essentially a pair o
770.
by Imnimo 5y ago
Maybe one day there will a job for people who are masters of the art of prompt hacking - they know all the special phrases and terms to get Dall-E to output the most aesthetically pleasing images. They guard their magic words like a mediev
771.
by Imnimo 5y ago
I meant to say that Dall-E 2's approach is closer to "teaching a three year old to paint" than the alternative methods. Instead of trying to maximize agreement to a text embedding like other methods, Dall-E 2 first predicts
772.
by Imnimo 5y ago
Yeah, I think the trick is that when you ask for "a picture of a sunset", you're really asking for "a picture of a sunset that looks like a realistic natural image and obeys the laws of reality and is consistent with all
773.
by Imnimo 5y ago
This is actual image generation - the 'decoder' takes as input a latent code (representing the encoding of the text query), and synthesizes an image. It's not compositing or querying a reference library. The only time tha
774.
by Imnimo 5y ago
I'd say Dall-E 2 is a little more unified - they do have multiple networks, but they're trained to work together. The previous approaches I was talking about are a lot more like the microservices analogy. Someone published a mod
775.
by Imnimo 5y ago
Yeah, I mean you're right that ultimately the proof is in the pudding. But I do think we could have guessed that this sort of approach would be better (at least at a high level - I'm not claiming I could have predicted all the tec
776.
by Imnimo 5y ago
I'm only part way through the paper, but what struck me as interesting so far is this: In other text-to-image algorithms I'm familiar with (the ones you'll typically see passed around as colab notebooks that people post outpu
777.
by Imnimo 5y ago
>Newspapers change titles and content of existing posts all the time. Sure, but I trust a newspaper to not swap their article out for something repulsive to troll people. I don't have that same trust in a random Twitter user.
778.
by Imnimo 5y ago
I guess it depends on whether you consider "The Next Google" to imply that it becomes a the dominant company in the space, used by every "average user", or if it's enough to be a niche solution for highly technical
779.
by Imnimo 5y ago
>But right after that, we’re greeted by an “Interesting Finds” section, which has a fun blog post by Derek Sivers, an article filled with stories of Steve Jobs in Japan, and some other cool things you can’t find on Google. You can defini
780.
by Imnimo 5y ago
>But you essentially can't overcome the network effects. There's no way to gradually build a social network when everyone is already on a different one. How did Facebook overtake Myspace then? I'm not really convinced that

More ›