psb217

12 ms·

91.
by psb217 2y ago
If the generative model/simulator can run at 20FPS, then obviously in principle a human could play the game in simulation at 20 FPS. However, they do no evaluation of human play in the paper. My guess is that they limited human evals t
92.
by psb217 2y ago
The agent never interacts with the simulator during training or evaluation. There is no user, there is only an agent which trained to play the real game and which produced the sequences of game frames and actions that were used to train the
93.
by psb217 2y ago
Trouble sneaks in when the pattern matching is only correct most of the time. Eg, if some code for regex-based search missed anywhere from 0.1% to 10% of matches, with the miss rate depending on the regex and no obvious way to know which re
94.
by psb217 2y ago
Your counterargument is invalid. The most advanced human intelligence invented (or discovered) concepts like multiplication, pi, etc., and created tools to work around the ways in which these concepts aren't well handled by their biolo
95.
by psb217 2y ago
The Rock would be bad at persistence hunting, but it would be weird if he was so susceptible to persistence hunting. What incentive does he have to fatigue himself in this scenario? When it's you vs animal, the animal runs from you sin
96.
by psb217 2y ago
In a sense, the model _is_ simply applying a finite and known set of axioms and manipulations. What makes this hard in practice is that the number of possible ways in which to perform multiple steps of this sort of axiomatic reasoning grows
97.
by psb217 2y ago
And to have a very well-tuned sense up vs down when the hill is almost flat...
98.
by psb217 2y ago
Well, the answer is probably between 1 and 10, so if you try enough prompts I'm sure you'll find one that "works"...
99.
by psb217 2y ago
How can I know whether any particular question will test a model on its tokenization? If a model makes a boneheaded error, how can I know whether it was due to lack of intelligence or due to tokenization? I think finding places where models
100.
by psb217 2y ago
People also often forget "orderless autoregression", which was introduced a while back and has been reinvented many times since. See Sec 4 (pg 8) of "Neural Autoregressive Distribution Estimation" [ https://arx
101.
by psb217 3y ago
She's also wearing a different jacket at the end of the video. Continuity is not maintained when the video zooms back out to a wider shot after the close-up on her face. See, e.g., no zipper on end jacket and obvious zipper on jacket e
102.
by psb217 3y ago
New AIdea... A "fast-forward or mute when person X is talking" app/plugin/whatever. You'll never have to hear X again! The core tech would actually be pretty simple, considering what's available open source.
103.
by psb217 3y ago
The Transformer paper, ie "Attention is All You Need", was a Google Brain/Research paper, not a Deepmind paper.
104.
by psb217 3y ago
There's a subtle difference here between the translation scenario and what you observed. In translation, the reversal only applies to the second sentence which will tend to present information in the same order as the first sentence (f
105.
by psb217 3y ago
Due to the particular form of the recurrent update of the hidden state, there's actually a parallel algorithm for computing the recurrence over length N in log(N) time via dynamic programming. Note, you don't save FLOPs, you just
106.
by psb217 3y ago
I've got a Ti double-walled mug from Snow Peak that I use a lot around the house. The big strengths are light weight, near indestructability, and a cool Ti functional aesthetic. It's double-walled and holds heat well, but I prefer
107.
by psb217 4y ago
Funnily, these folks at MSR also don't know what's in the black box. They just got early access and permission from OpenAI to poke it with a stick.
108.
by psb217 4y ago
How is the request that someone provide a clear set of definitions and some empirically falsifiable hypotheses a "dead end for scientific theories"? It seems more like the foundation of the scientific method.
109.
by psb217 4y ago
With MCMC, depending on application, it seems risky to just toss out the NaN/inf results. I'd guess these numerical issues are more likely to occur in certain regions of the state space you're sampling from, so your resulting
110.
by psb217 4y ago
To train the inverse diffusion model, we take a clean image x0 and generate a noisy sample xt which is from the distribution over points that x0 would visit following t steps of forward diffusion. For any value of t, any xt which is visited
111.
by psb217 4y ago
The reason why big steps produce worse results, when using current architectures and loss functions, is precisely because the least squares prediction error and simple "predict the mean" approach used to train the inverse model do
112.
by psb217 4y ago
In the reverse diffusion process, the reason we can't directly jump from a noisy image at step t to a clean image at step 0 is that each possible noisy image at step t may be visited by potentially many real images during the forward d
113.
by psb217 5y ago
In this metaphor you're the buildings, not the builders.
114.
by psb217 6y ago
I've added a brief explanation in a reply to a sibling comment.
115.
by psb217 6y ago
The bound is completely invalid, as are the NLL/PPL numbers they report with the MELBO. Look at the equation. If they optimized it directly, it would be trivially driven to 0 by the identity function if we used a latent space equivalen
116.
by psb217 6y ago
FYI, the MELBO bound in that paper is invalid. Their perplexity numbers using the MELBO bound are also invalid.
117.
by psb217 6y ago
Information is actually about _reduction_ in entropy. Roughly speaking, entropy measures the amount of uncertainty about some event you might try to predict. Now, if you observe some new fact that has high (mutual) information with the even
118.
by psb217 6y ago
If you know which direction the released news will move the stock and roughly when the news will be released, you could sit waiting to pull the trigger faster than most people who were not similarly advantaged. It would be silly to do somet
119.
by psb217 7y ago
With data augmentation, we're effectively injecting additional information about what sorts of transformations of the data the model should be insensitive to. The additional information comes from our (hopefully) well-informed human d
120.
by psb217 7y ago
Invest conservatively and pull out 2-3% per year. This should provide you with, conservatively, 50k-100k per year in post-tax cash (inflation adjusted) to spend as you like. Historically, you should be able to withdraw at this rate "in

More ›