Jack000

13 ms·

91.
by Jack000 4y ago
Sometimes there is just no reductive theory for emergent phenomena. For example, in the case of ant routing we have a good understanding of the behavior of individual ants, and we can observe the intelligent behavior of the colony as a whol
92.
by Jack000 4y ago
This is mainly because DALLE-2 generates at 64x64, then uses two separate upscalers to reach 256px then 1024px. The problem is that the upscalers are much less capable than the main model, and each upscaling stage tends to compound the mist
93.
by Jack000 4y ago
data isn't necessarily a problem for training agents. A sufficiently complex, stochastic environment is effectively a data generator - eg. alphago zero
94.
by Jack000 4y ago
it's not a given that this will happen on a human timescale. The carboniferous period lasted 60M years before bacteria evolved the ability to break down cellulose.
95.
by Jack000 4y ago
Single-task ANNs like convnets are really quite different from the human brain as a whole. Without the "rest of the brain" it fundamentally doesn't understand that the pixels it sees are representations of a 3d world, with di
96.
by Jack000 4y ago
the brain as a whole doesn't work the same way as ANNs, but the visual cortex in particular is extremely similar to a convnet. We even know which layers are responsible for which features, on a coarse scale: https://www.scie
97.
by Jack000 4y ago
ML benchmarks and scores (eg. FID, BLEU) correlate with model ability, but it's problematic to compare their absolute performance against humans. Convnets for example, are directly analogous to the visual cortex of humans. To get an ap
98.
by Jack000 4y ago
Most of these stem from the way DALLE-2 is designed - in particular the composition, counting, negation and concept blending examples. In the paper they explain that these aspects were deliberately sacrificed in order to improve image quali
99.
by Jack000 4y ago
2022: Deepmind releases paper on bootstrapped meta-learning and scaling RL agents 2023: RL agent trained for multi-task learning solves majority of perfect information games. It's a scaled up decision transformer. Scaling laws for RL a
100.
by Jack000 5y ago
hm.. I wonder which clip model they'll use. A big part of what makes DALLE-2 so good is the unreleased huge clip model. To train the diffusion prior they may need to first replicate this clip model.
101.
by Jack000 5y ago
Just tried glid-3 with a batch size of one and I'm getting 4781MiB. The latent diffusion model peaks at 8403MiB These are fp16 numbers though, you might need a recent nvidia card to run it.
102.
by Jack000 5y ago
glid-3 is trained specifically on photographic-style images, and is a bit better at generalization compared to the latent diffusion model. eg. prompt: half human half Eiffel tower. A human Eiffel tower hybrid (I get mostly normal Eiffel tow
103.
by Jack000 5y ago
This is exactly what they demo - they lock a scene and add a flamingo in three different locations. In another one they lock the scene and add a corgi.
104.
by Jack000 5y ago
With glide I think we've reached something of a plateau in terms of architecture on the "text to image generator S curve". DALL-E-2 is a very similar architecture to glide and has some notable downsides (poorer language under
105.
by Jack000 5y ago
This is all true, but it doesn't preclude the possibility of generating DNA. Human share a lot of DNA sequences with other animals, and the genetic differences between individual humans are even smaller. You might have trouble generati
106.
by Jack000 5y ago
The first quote is talking about prompting the model with images instead of text. The second quote is using "mix up" in the sense that the model is confused about the prompt, not that it mixes up existing images. ML models can out
107.
by Jack000 5y ago
I don't know anything about quantum mechanics, but if I'm reading the article right the experiments are meant to disprove this? > Perhaps it’s most natural to represent each complex dimension by two real dimensions, but real-va
108.
by Jack000 5y ago
I did the 3d point cloud just to visualize the data at first. Then after I thought it made a pretty cool landing page.
109.
by Jack000 5y ago
you can do this by setting the background to black, then lock it. When you click generate again, the subsequent palettes will then take the dark background into account. there's also a dark mode preset in the (gear icon) menu. I conced
110.
by Jack000 5y ago
yeah that should be possible, just need to quantize the resulting palettes to the closest available thread color. You could possibly do it with the api (there's some instructions at the bottom of the about page)
111.
by Jack000 5y ago
each template on the left is a separate thing basically. When you lock a color, then click "generate", the next palettes take your locked colors into account - if you lock the background to white, then click generate, the foregrou
112.
by Jack000 5y ago
Try increasing the "creativity" slider in the options. Setting it to a higher value will increase diversity at the cost of accuracy. (This controls the sampling temperature like in language models) When I use the tool myself, I us
113.
by Jack000 5y ago
thanks! it's supposed to be like an easter egg.
114.
by Jack000 5y ago
Also, here's a writeup on how the machine learning system works: https://huemint.com/about/#machine-learning You can guide the ML model by locking one or more colors, then clicking generate again. (click the circu
115.
Show HN: Huemint – Machine learning for color design(huemint.com)
335 points by Jack000 5y ago|52 comments
116.
by Jack000 5y ago
Michelin restaurants can be good value from a certain perspective. I went to a 3 star place in Japan that cost $500 a seat, but there were only 8 seats and two rounds a night, with maybe 6 people behind the counter. You do the math and they
117.
by Jack000 5y ago
In photography it's neither the camera or the camera manufacturer that's considered the artist, I think the same logic would apply.
118.
by Jack000 5y ago
I think this type of model will have a massive impact on the software industry. 99% of programming tasks in the wild don't involve any kind of algorithmic design, but are more like making a CRUD pattern, writing SQL queries etc. This k
119.
by Jack000 5y ago
Tech is always changing, so your current experience and qualifications are always on a timer regardless. Personally, I enjoy solving problems with computers and rather dislike the tedium of actually writing and debugging software. If a ML m
120.
by Jack000 5y ago
An analog radio is probably the best case scenario for lesion study. The main purpose of the radio is signal processing, which is done in serial with not many inter-communicating parts. In fact I think it should be possible for biologists t

More ›