DRµGS: Deep Random Micro-Glitch Sampling | Hacker News Reader

DRµGS: Deep Random Micro-Glitch Sampling

github.com

169 pointstosh41 comments

I appreciate the README etc, it's very funny and cute.

As a bit of a layperson / just an AI integrator, I want to get some clarification of how this interfaces with a generative AI. I haven't dug too deep in and I have a mostly elementary understanding of neural nets. No need to _entirely_ dumb it down though, just needing a smarter person to confirm or correct my understanding. :)

Is this influencing the probabilities of token output? My understanding is that currently it is a static value that is essentially "how wild do you want it to be" where each generation has a different consistent static value of "wildness".

So rather than using a static value for an entire generation, this dynamically injects randomness in so rather than the output being 'monotone', it is more dynamic rather than by-the-book that AI tends to be?

Typically the output of an LLM is a classifier that gives scores reflecting the "best" next token, for some measure of best. You can deterministically just take the highest score or you can randomly sample in a way that makes it more likely to get a token with a higher score, according to different schemes. The point is that this all rests in the performance of the output classifier layer.

From my skim, this proposes to inject randomness throughout the model (as opposed to just during sampling the classifier output) which according to the author (I'm paraphrasing) lets the full power of the model be used to make the most of that randomness, instead of just jittering the output. So the output is still random but it's a function of random noise being added through the model layers instead of just at the output, supposedly demonstrating better properties. I have no opinion about it as I haven't studied it.

> You can deterministically just take the highest score or you can randomly sample in a way that makes it more likely to get a token with a higher score, according to different schemes.

Ah ok, so it's like the difference between greedily always taking the shortest / highest value vs using A* and exploring alternatives in hopes of finding something better? That probably doesn't line up as a metaphor as it isn't "exploring" or running multiple parallel generations.

And it employs the method in the layers so that the benefit can be amplified rather than only employing the method in the final layer / output when it selects the output to take

The difference is between something that makes sense (the way the model is intended to be used) and something that makes almost no sense (this repository).

> benefit

In an intellectually honest way, it is not conclusive if there is a benefit to the approach in this repository. However, as a betting man: there is no benefit.

There are a lot of surprises! Riffusion and LCMs for me were super surprising. With no examples or meaningful investigation, I don’t think this is going to yield anything.

It's funny that you think the way transformer models are "intended" to be used is the result of intentional engineering and not an accident that ML academics are no closer to giving some causal explanation for why it works today than they were when Attention is All You Need was published.

It's so funny how little love that work on the decoder/sampler side with LLMs gets. You can take bad LLM and make them decent, or good LLMs and make them garbage with intelligent sampling settings. It's cool to see further innovation in this space.

I appreciate the commitment to the metaphor. Bravo.

it is excellent copy, at the very least

Great Readme. Strikes the perfect balance between quickly informative and enjoyably cute.

So instead of the mathematically sound approach of sampling from the categorical output distribution (that the model parameters were fitted to with maximum likelihood) you just inject some noise in the activations, and then pick the highest probability output?

Injecting noise has many mathematically-sound interpretations, like the Bayesian interpretations of dropout for ensembling or posterior sampling.

It’s a biologically inspired approach, hence the name. Biological neurons don’t work based on sampling functions, they have inherent non-determinism distributed throughout the network, and receptors to enable runtime modulation of the network’s function.

This dynamic modulation is seemingly at odds with mathematical soundness. Whether mathematical soundness should be a goal or not is a question to be answered empirically.

Thank you for the explaination.

I had to chew on this for a minute or two to put it together. This explaination helped a lot.

I wonder if this works for Stable Diffusion as well...

There are transformer-based approaches to vision models. It should work for those.

From what I understand, Stable Diffusion starts out with random noise that is then step by step made into the final image. Randomness is there at the beginning.

Current LLMs first find the n most likely next token and only then gets randomness on the choice among the top-n. This injects randomness to the initial search for the top-n.

So I think, from my ignorance, that Stable Diffusion is already doing this in some sense

cute :P

What does cute mean ?

kawaii

Isn’t this the GamerGate guy?

Apparently so:

https://ggwiki.deepfreeze.it/index.php/Eron_Gjoni

Unclear how "the gamergate guy" he was Vs just "accidentally triggered gamergate". I'm not going down that rabbit hole though - where's Internet Historian when you need him?

Probably plagiarizing something.

Internet Historian recently got cancelled for being a Nazi and hiding "Nazi symbolism" in his videos is what a quick google told me.

I'm still not really sure what gamer gate was or who the Internet historian is

Where did you find that Internet Historian is a nazi? I only found something when searching "Internet Historian nazi", a reddit post that also claims that Jontron is a white supremacist. This post: https://www.reddit.com/r/196/comments/18erlt0/internet_histo...

If you actually googled "Internet Historian nazi" for a guy you don't even know who he is, that would be pretty funny.

I have seen an Internet Historian video before. I just googled "Internet Historian" and a cancellation style reddit post was the third result. I even mentioned that I only did a quick google; I encourage you to do your own research here and form your own opinion on this. I have heard of Jontron but I am not familiar nor have I read anything about that.

JonTron actually said white supremacist stuff on video, for which he later apologized (why he's not part of Game Grumps anymore).

What do you mean? He left Game Grumps in 2013, way before he said anything controversial.

I think to be "cancelled" you have to actually get cancelled from somewhere... His Youtube channel seems to be going fine.

It's very obvious from his videos that he is a 4chan troll. Didn't know he was full on Nazi though.

"Despite GamerGate’s positive changes on the industry, it has been met with heavy criticism from the journalists and publications under fire and their supporters, as well as third-wave feminists involved who have attributed the harsh criticism towards female offenders to misogyny and sexism."

That summary seems a bit tendencious!

Nothing like good old monk humor about sex.. Unless druggy humor by the notoriously drug free..

Yes, because HN crowd or AI/software crowd in general is famous for being drug free and never indulging in various psychedelics out in the desert.

If they would, they would be burning man. Nothing to ignite some passion than a deep dive in the desert ;)

It's kind of funny, I don't really see psychedelics as "drugs" in the same way that you wouldn't view coffee as a drug. I don't think SW engineers are, for example, known for taking meth or heroin, so I'd definitely call the group fairly "drug free".

I'd think (based on nothing but my experience) that your average nurse might have higher likelihood of taking hard drugs than an average SWE :)

> I don't think SW engineers are, for example, known for taking meth

Some people think meth means “amphetamine” (although it usually methamphetamine), and there’s a lot of software engineers who take different types of prescribed amphetamines like Adderal and Ritalin.

And don’t even get me started on weed and millenials.

Methylphenidate is not an amphetamine but your points stand.

From Wikipedia: > A drug is any chemical substance that when consumed causes a change in an organism's physiology, including its psychology, if applicable.

Coffee, psychedelics and alcohol are drugs just like heroin. Whether you believe they are useful to consume is a different matter.

That's exactly what I meant. In a literal sense, pretty much any substance/chemical is a drug (like sugar). Yet, since the 80s, the word "drug" means something very different.

In that sense, I wouldn't call SWEs a "drug using group."

We all know the pedantic meaning of the word "drug" - in that case, the whole humanity is a drug-taking species, since we prefer drug-altered consciousness to our natural one (e.g. by caffeine).

Out in the desert is more extraordinary than microdosing while WFH I guess

Ahh yes, the old drmgs.

i was just wondering yesterday what happened to this guy

Reply on news.ycombinator.com