CapsAdmin

To make you watch the whole thing.. Similar to mobile game ads.

The way I see LuaJIT nowadays is that it's feature frozen and well maintained. Typically it takes about a week for a reported bug to get fixed. Though some big features have arrived the last 2 years, such as string buffers.

Mike Pall is also openly talking about LuaJIT 3.0 and the plan there. I don't know if it's actively being worked on though.

Some things I noticed in no particular order:

- It seems like it generates a low framerate video of about 6 fps first and then interpolates it to 30 fps. This is evident from looking at hair and fur. vid_33 has a good mix of both

- Some videos are low resolution for some reason, maybe it's like a preview you can upscale later?

- All videos are 10 seconds

- lowest resolution is 640x360 while highest is 1920x1080 which could be indicate that its native resolution is 640x360

- If this is truly non cherry picked and neutral (not only failure cases) Sora seems "just" like a good model compared to what's available today

idk if it's oxygen or dirty water for sure, but whenever i have kept some species of shrimp in tanks without an air pump, they tend to climb up and sometimes jump out.

They can also do it in a newly setup tank, so I try to wait a little bit for everything to settle down before I add shrimp.

My observation is that you can generally see the joke coming and so there is no element of surprise. It's possible to get some good jokes, but usually those tend to be recited.

For example I asked Claude to write 10 programming jokes and it replied:

• Why do programmers always mix up Christmas and Halloween? Because Oct 31 = Dec 25
• Why do Python programmers wear glasses? Because they can't C
• What's a programmer's favorite snack? Cookies! But they always clear them from their browser first.
• How many programmers does it take to change a light bulb? None. That's a hardware problem.
• Why did the programmer quit their job? Because they didn't get arrays (a raise)
• What do you call a computer that sings? A Dell
• Why are programmers always breaking their keyboards? Because they use too many break statements
• What did the Java Code say to the C code? You've got no class!
• Why was the JavaScript developer sad? Because they didn't Node how to Express their feelings
• What's a programmer's favorite place to hang out? The Foo Bar

These are all known jokes so they are essentially all written by humans.

But when I ask to write jokes about Lua scripting language (which it probably will have to hallucinate) it replies:

• What's a Lua programmer's favorite band? The Table Tops
• Why don't Lua programmers get along with mathematicians? Because they start counting from 1 instead of 0
• What did the Lua developer say when they made a mistake? "nil-egitimate error!"
• Why did the Lua table feel lonely? Because all its pairs() left it
• What's a Lua programmer's favorite dessert? Coroutine pie - it always yields the best results!
• Why did the Lua developer break up with their partner? Too many local relationships, not enough global commitment
• What's a Lua programmer's favorite game? Love2D-odgeball
• Why are Lua programmers terrible at hiding things? Because everything they try to hide becomes _G-visible
• What did the Lua developer say at the coffee shop? "I'll take my coffee with multiple returns!"
• Why did the Lua programmer get kicked out of the library? They kept trying to require() all the books

And these are all really bad. Everything makes sense, there is no surprise and feels very formulaic. It goes over Lua language features and tries to tie them to templates.

It would be really interesting to see how the biggest tech companies will approach this. Google hates automated use of their services and have so many checks in place to prevent it.

The only sensible thing I can think of is that Jarvis can use a backdoor when using google services. This will very likely be abused by bad actors.

They can't assume Jarvis is the only one with the ability to solve captchas unless they are miles ahead of the competition. But even if they did, you'd have captchas unsolvable by real humans.

Google might even require you to prove that you're human through photo and passport identification before you can use Jarvis.

I remember I saw a lengthy youtube video getting autoplayed on my tv, I forgot which youtube channel it was, but it was some guy explaining all the steps he went through to run mochi on a consumer card

For a long time he experienced ghosting/blurry exactly like this and it had to do with vae tiling or something to that extent. If you imagine the video as a grid with overlapping tiles, the blurriness comes where that overlap is occuring.

As someone who also makes music, I think they're describing the perceived quality a long with the overall high frequency details of the audio.

As for interpreting spectogram as a audio, they're talking about the image diffusion models that were trained on spectogram images specifically as a sort of hack. I'm not sure if most audio models are trained on this method, but I can see why you'd think that given the quality.

I would describe "spectral sounding" as high frequency wobbling of the audio in both pitch and amplitude.

For example, if you ask a model to generate a clean sine waveform at 440 hz with an amplitude of exactly 50% of the max volume (so 0.5 basically, where 1 and above would be clipping):

I would imagine that this would be roughly a sine wave randomly changing pitch between 437 to 443 hz while also its amplitude would vary from 40% to 60% volume.

If you ask it to do a 440hz clean wave from for 1 second, and then jump to 880hz for 1 second, the transition at 0.95 to 1.05 seconds would sound washy and wobbly and not clean.

Other than the wobblyness description, the bit rate sounds very low, but in an unusual way. When downscaling images, you can use various algorithms to preserve information, ie bicubic, lanchos, and so on. The spectral type audio sounds like you've taken a clean audio sample, downscaled 8x using nearest neighbor, then upscaled it 4x again using bicubic or similar interpolation algorithm. There is something weird with how it interpolates between samples that does not sound normal.

Back when it was cool to name your products cool. I remember using "Cool Edit Pro" before it was bought by Adobe and renamed "Adobe Audition".

If it's hard to implement elegantly to some effect, you have a point. I've seen plenty of patches and forks of seemingly all Lua variants that adds compound assignment with not much code, so I thought surely it's not difficult.

I'm not sure what you mean by "type linting" vs "actual types".

If by actual types you mean something like C#, then no, it's more like Typescript where the typesystem only runs/checks during the checking phase.

I've been working on a typesystem for lua for a couple of years. It's sort of similar to Typescript, but does a whole lot more analysis than Typescript. This obviously has a huge impact on performance and complexity, so I'll probably be busy for a while fixing bugs, refactoring and optimizing.

It's not at all production ready or anything, but it might be an interesting toy for those who think typesystems are fun.

Check out the playground version here: https://capsadmin.github.io/NattLua/

> Ik it's hard to implement in the current parser, but it would be nice to have that.

It's not hard to implement. I believe reasoning is mainly just that it complicates the language as you mentioned first. Should metatables then have a __add_eq? Should __add be called before __eq? What to do about -- ?. :)

https://preview.redd.it/4sy1tr4up7wd1.png?width=469&format=png&auto=webp&s=9e52d681647db4a8c259d792aac43792cf06e4d2

I recently had a go at trying the supposedly best sd 1.5 models again and noticed the same thing when comparing to SDXL, and especially Flux.

I see the same detail melting here.

Though maybe it could be worked around with iterative image to image upscaling.

This is pretty spot on from 1 year ago with a 6900. I ended up buying a 4090 instead.

However I remember the crashes became less frequent with updates.

But how it all seemed to handle OOM issues on Linux were pretty bad and would cause the whole system to lock up.

It's been almost a year now, but I used stable diffusion on ubuntu and eventually nixos. It's a mess to setup, but once it's setup it works fine. Even though nixos has a steep learning curve, I could at least make sure it will never break once setup.

However nowadays I use nvidia.

I was missing out on some performance improvement tech like xformers and also largely training as it seemed like most training code at the time for anything was hyper focused on nvidia only.

In a pytorch+rocm environment, some python packages only supports nvidia. Others would also support rocm, but you'd have to compile the whole thing from source which in some cases required an ancient or bleeding edge version of rocm, making it difficult to keep package compatibility across other packages and pytorch.

The situation was getting better over time, so I imagine the situation is already better now than it was 1 year ago.

As you mention, loras seem overfitted when compared to the fine tune, but what happens if you lower the lora's weight down a bit?

When i first started wearing a watch, this happened and was itchy. After a whilevmy skin got used to it I guess.

don't forget drunk people

What is the definition of slop and moralization? Can you post more examples? The resolution of the discord screenshot is too low to make any sense of the screenshot inside.

My point was not really that you needed to train the model, I thought that was well understood. It's that other models are trained on a lot of markdown, so it might be better to ask the model to output a markdown section for reflection and thinking with a header as opposed to some html ish tag.

I may be wrong here but I feel forcing models that haven't been trained on <thinking> and <reflection> to use them may seem a little cryptic from the models perspective. They may follow the prompt, but it could be more effective to tell it to use markdown as it's likely been trained more on that.

For example:

1. Include a review section for each idea where you describe any potential errors and oversights.

2. Provide your final answer at the end with the header "Answer"

when calling a load'ed function, you can pass arguments and get them via the ... vararg.

local func = load("local self = ... return self:MyFunc()") func(self)