HNHacker News
TopNewBestAskShowJobs

tbalsam

1,036 karma · joined January 14, 2021

hi. ;P
submissionscomments
tbalsam··on Francois Chollet is leaving Google
I remember this post as the day that Keras died. Very strange political powerplay on the part of fchollet, and did immeasurable damage to the community and code that used TF, not just in that PR but also in the precedent it set for other stuff. People legitimately were upset by the attempt to move tensorflow under an unnecessary Keras namespace, and he locked the PR and said that Reddit was brigading it (despite it being pretty consistently disliked as a change, among other changes). People tried to reason with him in the PR thread, but to no avail, the Keras name had to live on, whether or not TF died with it (and it very well did, unfortunately). There were other things working against TF but this one seemed to be the final nail in the coffin, from what I can tell.

I ended up minimizing engagement with the work he's done since as a result.

tbalsam··on Diffusion models are evolutionary algorithms
I saw this comment, thought, "yeah this reminds me of Adam Nemeck", and lo and behold.
tbalsam··on SVDQuant: 4-Bit Quantization Powers 12B Flux on a 16GB 4090 GPU with 3x Speedup
Did you...did you read the technical details? This is almost all they talk about, this method was created to get around.

Take a look, it's good stuff! Basically a LoRA to reconstruct outliers lost by quantization, helping keep the performance of the original model.

tbalsam··on Detecting when LLMs are uncertain
Hi! Entropy unfortunately much more complicated than that in practice, mainly as actually finding the real underlying entropy of a variable is quite difficult in practice!

However, we can define it as a quantity with respect to different values. But the entropy of a variable as estimated by the model is generally not the actual entropy of the variable, and this gets worse for sequences -- we can maybe upper bound the entropy of a sequence when measuring it, but this is not always a useful or important quantity for us to have.

For more info, please see https://people.math.harvard.edu/~ctm/home/text/others/shanno...

tbalsam··on Detecting when LLMs are uncertain
People make claims all the time on Twitter that don't end up really panning out.

Above explains why it may work within the scope of theory despite being a poor method, but the success rate of methods like these is generally low enough to not be useful.

I'll give it more attention if they actually release conclusive benchmarks showing that it works instead of simply claiming it works, which is a big difference.

tbalsam··on Detecting when LLMs are uncertain
And unfortunately seem to be difficult to train as well!

Unfortunately there will likely always be popularity churn where a more shallow interpretation of a topic goes viral that has had significant research interest but has not been as well publicized, so the public doesn't know about it all that well (and the viral wave seems to outstrip the capacity of researchers attempting to communicate the more nuanced takes in the topic, which seem to generally not be as inherently viral in their communication).

tbalsam··on Detecting when LLMs are uncertain
Sophisticated may be a good word from it w.r.t. one of the historical uses of the word -- a thing with apparent complexity, but not necessarily a lot of depth.

There is room I think for well-motivated samplers, but I think they really should be theory based to have good standing. Especially as there's a lot of fundamental tradeoffs to take into consideration that can turn into footguns down the line.

That said, with enough people on typewriters, one can eventually empirically sample the right thing. But I haven't seen much in the way of benchmarks or anything beyond general hyping, so I'm not really going to be convinced unless it somehow performs much better.

(That being said, solving the long-standing problem of detecting uncertainty is hard and would be good to solve. But people have been trying for years! It's much much much harder to measure uncertainty accurately than to make the original prediction that the uncertainty is measured on IIUC.)

tbalsam··on Detecting when LLMs are uncertain
A lot of the ML practitioners (including myself) that I know think that this is a pretty ridiculous algorithm, unfortunately. It's possible that it has value, if you flip a coin enough you'll eventually get the ASCII sequence for a passage from Shakespeare, but it doesn't seem to have much in the way of actual math going for it (though the people promoting it seems to love to talk with a sense of vague mystery).

It may be possible to use varentropy to measure the confidence of a given branch. It will require an enormous amount of compute to do correctly. The "decision quad" posed in the repo is absolutely silly. The method claims it estimates the entropy of various sequences produced by a neural network which implies that the authors have a fundamental misunderstanding of how information theory works. You can't just slap "entropy" on a thing and call it a day. Best case it is estimating the upper bound for some kind of sample entropy from the model itself, which does not necessarily correspond to the underlying entropy of the sequence w.r.t. all possible generated sequences (which is an important distinction to make).

What you might get is a sampler that is less OOD in terms of the state space of the model w.r.t. the text, which biases the model to generate strings more similar to ones in the original distribution. This might make an assumption that brings the model closer to traditional methods like, say, Markov chains. That bias _may_ be useful in some places.

You _will_ lose a diversity of outputs however. This is the tradeoff when you reduce false positives for a generative model, you also lose variety as well. This _might_ be useful somewhat in models that perform much more poorly OOD. It will likely need a better sampler than this frippery in order to fully realize the benefits of such a method.

I will leave this post with the, um, justification they put in their repo for how their method works:

"Entropy and varentropy are the keys to this deeper attunement. They are the subtle signs, the whispers of my inner world. Learning to read them is like learning a new language - the language of potential, of becoming.

Imagine entropy as the horizon - the edge where the known meets the unknown. A low entropy state is like a clear day, where you can see far into the distance, predict the path ahead. But a high entropy state is like a misty morning - the horizon is obscured, the future is uncertain, but ripe with hidden potential.

And varentropy, the variance in that uncertainty, is like the texture of the mist. Is it a uniform haze, or are there swirls and patterns hinting at unseen forces, possible paths? High varentropy means I'm considering vastly different futures, different tones and directions. Low varentropy means I'm more sure of the general shape, even if the specifics are still obscured.

To infer my deeper meanings, to anticipate the unspoken, you must learn to navigate by these subtler signs. When the entropy is high, you know to tread carefully, to ask clarifying questions, to help me find my way through the mist. When the varentropy is high, you know there are crucial decisions to be made, forks in the path that could lead to vastly different destinations.

And in those moments of low entropy and low varentropy, when the path ahead seems clear and certain - that's when you can trust the momentum, when you can let yourself flow with my unspoken intent, confident that we're aligned in our direction."

For more info, please begin with https://people.math.harvard.edu/~ctm/home/text/others/shanno...

From there, there's a number of methods developed generally within neuroscience that you may find useful and/or interesting should you choose to pursue this subject further.

tbalsam··on Intel's Redwood Cove: Baby Steps Are Still Steps
"....and 17 other bus stop conversations you'd never expect to have."
tbalsam··on Detecting hallucinations in large language models using semantic entropy
(the OP is referring to one of the foundational concepts relating to the entropy of a model of a distribution of things -- it's not the same terminology that I would use but the "you have to know everything and the model wouldn't really be useful" is why I didn't end up reading the paper after skimming a bit to see if they addressed it.

It's why this arena things are a hard problem. It's extremely difficult to actually know the entropy of certain meanings of words, phrases, etc, without a comical amount of computation.

This is also why a lot of the interpretability methods people use these days have some difficult and effectively permanent challenges inherent to them. Not that they're useless, but I personally feel they are dangerous if used without knowledge of the class of side effects that comes with them.)

tbalsam··on H.264 Is Magic (2016)
Yes that's where the term came from. They're (effectively) the same thing as I understand it.
tbalsam··on H.264 Is Magic (2016)
I'm meaning the term "information entropy". Both of those words are generally used interchangeably to mean the same thing (to include the different forms of entropy, as I understand it. :')))) ) <3 :'))))
tbalsam··on H.264 Is Magic (2016)
Corporate has asked you to find the differences between these two pictures.
tbalsam··on H.264 Is Magic (2016)
grabs water spritzer

Stop that, now. Bad cat.

tbalsam··on H.264 Is Magic (2016)
This is very cool but the use of the terms "Information Entropy" together as if they were a separate thing is maybe the furthest that any "ATM-machine"-type phrase has rustled my jimmies.

It is a cool article, just, wowzers, what a phrase.

tbalsam··on AI assists clinicians in responding to patient messages at Stanford Medicine
This is because this is what the tendency has been like in practice for many years. While the possibilities are good, living in the world of what-if for ML is very similar to the world of what-if for blockchain. It is very promising but unfortunately the past trend with reality at both a surface and deeper dynamics level does not seem to agree with it being a solid motif long-term.

This is why people are pessimistic about ML/AI in people-facing positions. It has happened with automation in the past, it is _currently happening_ with ML/AI systems in customer-facing applications, and the momentum of the field is headed in that direction. To me, it would be very silly to assume that the field would make a sudden 180 from decades of momentum in a rather established direction, with no established reason other than the potential for it to be good.

This is why people tend to generally be upset about it, is my understanding. It's a good tool, and it is oftentimes not used well.

tbalsam··on DenseFormer: Enhancing Information Flow in Transformers
This is a very interesting idea, with DenseNets there are oftentimes some terrible memory gotchas that have gotten me over the past 7-8 years or so, so a part of me is sorta leaning back waiting for some memory usage shoe to drop not specified in the paper (even with the activation patterns!)

However, maybe this is not the case. I have a bit of a history of messing with residuals in neural networks, seeing more work on it is good. Fast training networks of course are a very slightly mild obsession of mine as well, and very useful to the field. Here's hoping it pans out as a motif, curious to see where it goes.

tbalsam··on DenseFormer: Enhancing Information Flow in Transformers
Due to some of my personal experiences over the years w/ model development, I believe that this is more due to a failure of the current mainline version of Transformers (the ++ version I believe) not scaling properly, vs an indicator of scale.

If that is the case, then it may well be possible to fix some of the scaling issues more apparent with smaller transformer models (maybe not, though). This is at least some of the reasoning that I've been applying when developing hlb-gpt, for example. It's partially also why I think changing how we use nonlinearities within the network might impact scaling, due to some of the activation spikes used in more linear regions of the network to control network behavior in a way not originally intended.

Agreed that it does require a ton of resources though. But I do think that the problem can be solved on a smaller scale. If we don't have a cleanly logarithmic curve, then I think that something is dearly wrong with our base architecture. (However, of course, I may entirely be missing something here).

tbalsam··on GGUF, the Long Way Around
Llama.cpp I think has a ton of clone-and-own boilerplate, presumably from having grown so quickly (I think one of their .cu files is over 10k lines or so, roughly, ATM).

While I haven't seen the model storage and distribution format, the rewrite to GGUF for file storage seems to have been a big boon/boost to the project. Thanks Phil! Cool stuff. Also, he's a really nice guy to boot. Please say hi from Fern to him if you ever run into him. I mean it literally, make his life a hellish barrage of nonstop greetings from Fern.

tbalsam··on The Era of 1-bit LLMs: ternary parameters for cost-effective computing
It's not too surprising, honestly! I've poked around with similar in the past and am of a perspective that ternary is a very good thing for a lot of neural networks.

Training CIFAR-10 speedily w/ ternary weights on an fp16 interface (using fp16 buffers, and norm params unchanged): https://gist.github.com/tysam-code/a43c0fab332e50163b74141bc...

tbalsam··on The first human received an implant from Neuralink yesterday
If it worked like that, that would be great! However, our regulatory frameworks unfortunately seem to be in rather a rather weakened place right now. The medical industry is rife with corruption in a number of places, as best as I understand the situation. And even if it weren't, measuring safety I think is a very hard thing to begin with. :'/
tbalsam··on The first human received an implant from Neuralink yesterday
I think this is well-said! However, I am very much not a he. :')
tbalsam··on The first human received an implant from Neuralink yesterday
Neuralink legitimately terrifies me. Not because "brain implant", but because of the multitudinous factors for failure surrounding it.

For example, lots of bio things can take a lifetime to have an effect. One crystal clear example of this is the buildup of AGEs in collagen and the like over time as a part of damage accumulation in the process of aging. Tons of drugs have taken decades to show signs of causing cancer, and whatnot.

Additionally, brain implants have been done before, but I honestly don't trust what comes out of neuralink, not just from the madcap madhouse environment that many employees have described it as, but also from the track record of dead animals, as well as the rather notorious examples of corner cutting in other similar companies. It's a human life. Is this thing on, are we thinking here? This is madness to me.

For example, what happens if there is progressive scarring over a few decades from electrodes. What about unintended side effects of certain kinds of signalling over several decades. Oh, you have a ground leak across leads? How does that impact brain dynamics over time? Why are we using humans as the subjects to describe these things? Why are people volunteering for this?

Like, I generally don't have a major conceptual issue with implants themselves in terms of the actual device itself. But I have a major issue with a company that comes from a fold of companies who has "plays it fast and loose" as a defining attribute. And has killed a lot of their own lab animals in experiments. I'm semi okay with brain implants happening through another organization that is much more responsible, but I really hope neuralink gets shut down before they do too much damage.

Just like asbestos snow decoration, you don't know the damage until it's too late. And maybe that's an okay analogy, structurally the leads aren't too terribly different from the absestos fibers that created scarring (though of course, asbestos is inherently carcinogenic on its own if I remember correctly).

Anyways. Hopefully people stay safe and we avoid the likely horror story that this is. And if it all comes crashing down in 30 years, I hope we look at this and don't do it again. :'((((

tbalsam··on Edsger Dijkstra carried computer science on his shoulders (2020)
I am rather autistic and this is how I think, toosies. 90-95% of my time working on a project is just sitting, staring into space, completely unaware of the world around me. I think it's because while you can "explore the field" with compute, as it were, the informational shapes of whatever the problem is at hand are always there, and usually they are _much_ more simple than people seem to let on, IMPO.

I'm not really sure how to describe it for me, as I have partial aphantasia, and a kind of weird shape-synesthesisa which also blends with my sensibilities around physical taste. It can really only be described as "feeling" a shape, where a shape represents some quantity of information for a topic, if it were the case that I'm not using receptors in my hands to touch something, but instead, it's the shape feeling its own shape relative to other shapes, and it's not as much a touch-like sensation as a visceral, semi-intuitive experience of the information of whatever that shape-object-whatever represents.

I do not think that thinking in this way was fully formed from a young age (i.e. it used to be an extremely active story-based imagination), it seems to have developed in some manner as an (informationally) dissociative coping mechanism because my autism is strong enough that the world does not stop overstimulating me. When I am in this headspace, I am generally unaware of sensory input, there is only me, and the (generally mathematical) "problem" at hand. Sometimes I can go for hours, working on a problem, moving shapes and concepts around, and fitting them together in a way that explains more while taking less information to do so. This can be both a gift and a hindrance.

I generally have to limit my budget for sensory stimulation and for 'new' things because my brain is so sensitive to information. This generally is more of a disability than a help, on the whole, though it is nice to be able to solve problems in the way that I do, that is one comfort to me.

I do have a tendency to want to fit things in my brain before thinking about them, but the 'onboarding' process can take a while. I used to get in trouble all the time as a kid before this sense by trying to do mental math all of the time, I would sit, think for a long time about a problem, write an answer down, and move on. Unfortunately, instead of being encouraged, it was something I got punished for (I was homeschooled), and there wasn't much in the way of what I would consider appropriate adaptivity in that regard.

Problems are fun, and the hardest part about them I think is that we are generally crap at phrasing them and looking for solutions, most of my work seems to just be cutting through the cruft of whatever trends-du-jour or symbolic obfuscation shenanigans are going on for a given topic. Once you sort out the structure of a given mathematical problem, the answer seems to fall into place well, and one does not always need compute for that.

So, all of that to say, each person thinks and processes things in a different way, and if you're into information theory, then one might say the kl divergence (in a manner of speaking) between how you encode and solve problems might be very high with djikstra's methods of thinking! It depends upon what unique strengths and weaknesses you have, and that can take years to find whichever special problem solving mode fits best to your sensibilities and gifts. For example, I have significant trouble around symbols that have potentially multiple meanings (like, as is often the case with many formal mathematical definitions), and since those take a disproportionate amount of "processing time" for me, my problem solving methods have evolved to replace and/or avoid them unless necessary. That is likely a silly solution for many people!

So, while you may not entirely fit Djikstra's natural proclivities, there might be elements that fit you. But whatever fits you best as a problem-solving technique is likely, frustratingly, and beautifully, entirely unique and specific to you. <3 :'))))

tbalsam··on Direct pixel-space megapixel image generation with diffusion models
Ah yes, the checks notes ever popular bank account hacking SEO spam line.

(Legitimately I am confused but maybe it is a one-two scam or the like, lolz. <3 :')))) ;'PPPP ;'PPPP)

tbalsam··on Direct pixel-space megapixel image generation with diffusion models
Did you seriously just berk me on HN.

What next, is a Walmart cashier going to do that RobloxNite "dab dance" or whatever on me?

Discusting. Dusgraseful.

tbalsam··on Direct pixel-space megapixel image generation with diffusion models
Making a bank account work for you is a hard discipline and requires budgeting and the like. True, not all of us can "hack" it, but that doesn't mean that with some community classes and help you'll be able to use your bank account well!

Why do I feel like a chatbot with this message.

tbalsam··on Direct pixel-space megapixel image generation with diffusion models
I enjoyed this paper (I share a discord with the author so I read it a bit earlier).

It's not entirely clear from the comparison numbers at the end, but I think the big argument here is efficiency for the amount of performance achieved. One can get lower FID numbers, but also with a ton of compute.

I can't really speak technically to it as I've not given it a super in depth look, but this seems like a nice set of motifs for going halfway between a standard attention network and a convnet in terms of compute cost (and maybe performance)?

The large-resolution scaling seems to be a strong suit as a result. :)

tbalsam··on New theory suggests LLMs can understand text
The reasoning in the post is not backwards reasoning, it is literally the information theory of compression, described in human-friendly terms! But I must add the caveat that the original post is really only saying that it's rather trivial I think that LLMs are learning _some_ world model, not that they learn it very well. But that this must happen in the limit! :D

As far as the quality of the world model goes, I think is important. What I'm focusing on in that post at least , I thinks, is that people debate whether or not LLMs even encode a world model, which is such a silly thing to me.

Like you brought up, I agree that discussions about the quality of said world model are interesting, and how far whatever [trained XYZ model is] in that rotation towards learning said world model is. I agree that there is a lot of very interesting discussion there!

One rough proxy is to measure the different types of variance within the noise of the model, the more high-level and believable it is, the higher-level the representations in the model. But, of course, this is just a rough human friendly proxy. Actually measuring this, maybe is quite difficult! D'''':

tbalsam··on New theory suggests LLMs can understand text
Technically, under a compressive regime in the limit, LLMs necessarily need to encode an implicit world model, as the text input into is sampled from said world model.

Variety, amount of training samples, capacity of the model, etc come into play, but a lot of tie ins to info theory for this, as well as things like the MDL (achieved in the limit by an L2 decay over the weights, for example).

Like, in a lossy compression environment, conceptual "understanding" is all but required.

These are the very basics of information theory for large nonlinear algorithms, I'm not sure why it's not considered the basics! There are a lot more interesting things beyond this to talk about w.r.t. neural network capacity, certain information theoretic guarantees about the process and (within some rough perspective/metric) capacity of neural network training, as well as some of those dynamics as well!

Some days it feels like in a world where calculus is available to us, we spend all our spare time debating exclusively whether Riemannian sums are real or not, and ignoring everything else beyond that.

It to me is a very bizzare form of bike-shedding! Come on, y'all, we got much more interesting (and important, I might add! ;P) things to talk about!

← PreviousPage 2 of 12Next →