rprenger

Didn't Matt Levine tell a story about Masayoshi Son doing that? https://news.ycombinator.com/item?id=21427688

Also, while we're getting rid of angles, can we please get rid of cross products and just use geometric algebra tools?

And you can do that at any board state, so if it starts with like 16 white squares you can make one or two greedy moves to minimize white squares, then do your memorize trick.

I felt a similar way when I moved from Chuck Yeager's Air Combat to (I think) F15 Strike Eagle 3. Yeager was the perfect mix for me. I remember rarely even seeing the enemy planes I was firing at in newer games.

I think if Kramnik accuses someone of cheating it might actually drop the posterior probability that they cheated.

One thing about horror movies is that even though they usually pretty much follow a three act structure they at least usually have to have a prologue to set the mood before the "regular life" part. But recent ones I've noticed that don't follow the structure closely are

"Skinamarink" and "I Saw the TV Glow"

Yeah as soon as I saw the "answer" I was like "yeah sure blah blah blah, you can make a trick question, but what IS the country that you'll actually hit if you walk East from Seattle". I was guessing England, but the right answer is France.

This is pretty close to how AlphaZero works.

https://medium.com/applied-data-science/alphago-zero-explain...

I think it's like:

Passive: Uh, your code looks great! Aggressive: Your code is garbage! Assertive: Your code is broken. Passive-Aggressive: It's great that you think your garbage code is great!

You can have both! LeanDojo: Theorem Proving with Retrieval-Augmented Language Models https://arxiv.org/abs/2306.15626 (shameless plug)

I believe Zero123++ gives you the ability to generate consistent images from multiple camera views, where as DreamCraft3D seems to output 3D models with geometry and textures if I'm understanding it right. Like I could take the output of DreamCraft3D (if I had the code) and drop the result into Blender for example.

I kind of hoped the long term result of htmx, instead of just being a Javascript library, was to become a standard in a newer version of HTML itself. The constraints of only forms being able to do POSTs, and the only way to handle the response is to rewrite the entire page seems arbitrarily limiting. If htmx were just HTML6 or whatever, it'd outlive all the Javascript frameworks.

The printing press.

I think the "weight" here is weight as in mass and "convolutions" here is as in folds. The size/weight and folds of Einstein's brain have been the subject of studies, and the quote was from 1980.

This Scott Aaronson lecture I really liked is relevant. It's like a "why quantum mechanics probably had to do the weird probability amplitudes (which can be negative and complex) instead of just normal probabilities even without experimental results" lecture: https://www.scottaaronson.com/democritus/lec9.html

I generally agree with the assessment that the limiting factor in quality is the signal from the MRI. Just one slight correction to the description of previous projects. We first built models that predict the voxel responses from images using ML. Then recorded the fMRI response on a new image. Then we ran many many images (not just those previously seen that have an fMRI response) through the voxel models and picked the top N images with predicted responses close to the actual response. Basically we used the big list of images as an approximate "natural image prior". Then we'd often show the average of that top N images so you get one picture (which destroys interesting multi-model stuff, but we didn't have good stable diffusion models of images back then that could do a gradient search on or something).

If you wanna see what the reconstructions look like without fancy image priors (to get a better idea of how little signal there is) check out this figure from one of the early papers: https://pubmed.ncbi.nlm.nih.gov/19778517/#&gid=article-figur...

We put this in our ASR based keyboard app back in 2016. Basically kept the ASR decoders beam search candidates around and could use their differences and relative scores to suggest replacements. I would've assumed newer keyboard apps would have a similar option, but I haven't checked: https://techcrunch.com/2016/10/03/baidus-new-talktype-keyboa...

It should output the result of a GPT-stlye transformer fine-tuned on data set of 1,000,000 instances of "Hello World"

Dr. Sbaitso! https://oneweakness.com/dr-sbaitso-online

I don't think a larger vocab would help. All the individual letters are in the ~50k token vocab already, but the word "alphabet" will still not get tokenized to [a, l, p, h, a, b, e, t]. Using a larger vocab like PaLM's 256k vocab would have the same issue.

Here's the resource I've used that I've liked best. It's more the math, but thinking this way helped me with other puzzles: https://math.berkeley.edu/~hutching/rubik.pdf

Is this a high enough price to be considered a "Bear Hug"? I immediately thought of this scene from Succession: https://www.youtube.com/watch?v=v4QHqjel4kI

For my first search of "GFlowNetworks" (which the search bar suggested) It said: Found 5,887 (or something) results, but showed no results

For my second I searched my name and got a Wikipedia article about a show I've never heard of which didn't have my name anywhere in it.

For my third I searched "GFlowNetworks" again and it said Found 2,656,844 results in 1.61s, but showed no results again

That's an interesting pattern, that seems pretty clear now that you mention it. I would also argue that Ralph Breaks the Internet falls squarely into the "self/internal conflict" class too, as the main antagonist is literally Ralph's own insecurity.

I think what the other commenters are getting at is that PDEs can be used without having a closed form solution (and mostly are used that way as closed form solutions usually only come up in special artificial cases). You start your system in a real known state and then propagate it forward in time using (for example) the finite difference method on the equations to figure out the state at a later time. https://en.wikipedia.org/wiki/Numerical_methods_for_partial_...

Has anyone figured out how to enable "stay sitting on chair during meal"? StackOverflow has totally failed me here.

Do you mean tritiated water?:

https://en.wikipedia.org/wiki/Tritiated_water

This got me thinking. If they did bump it up from $1m to $100m, do we think the Riemann Hypothesis would be solved sooner?

Me too. I had to look it up. This page has pretty good breakdown:

https://itnext.io/a-wild-way-to-check-if-a-number-is-prime-u...

The main trick for me was you first have to convert the number to unary, which was done outside of the regex.

Yeah, figure 4 is more clear. This is the early stopped loss though, so the regularization is more explicit. If you trained the large models to completion on small data sets they would do much worse on test error due to overfitting.

It is interesting that larger models with regularization (early stopping) seems to work better than than training smaller models to convergence though.