HNHacker News
TopNewBestAskShowJobs

radarsat1

4,157 karma · joined May 28, 2011

Info: https://sinclairs.gitlab.io

Email: <myusername>@gmail.com

submissionscomments
radarsat1··on Unsupervised speech-to-speech translation from monolingual data
Amazing. And frustratingly light on details about this magical "MUSE loss" which seems to do a lot of the heavy lifting here, will have to read the paper to see how that works. Anyone have a tldr?
radarsat1··on Unsupervised speech-to-speech translation from monolingual data
AudioLM was developed more than a year ago and was an "audio continuation" model which is basically what you describe. However it was not used for conversational purposes afaik. Maybe it just needed some RLHF tuning?

https://google-research.github.io/seanet/audiolm/examples/

radarsat1··on Show HN: Play a pen-and-paper game that is almost unknown in the US and Europe
I was wondering about the architecture because I want to be able to make little games like this, so I'm super happy to see the source here, thanks for posting it! And looks like you put a lot of thought into the architecture too, going to study it.
radarsat1··on Animate Anyone: Image-to-video synthesis for character animation
I'm as interested as anyone in the methods behind this and find deep learning just continues to amaze and I really enjoy working in the AI space. Having said that, seeing that we now essentially have the capability to synthesize videos of people doing things they haven't done and saying things they haven't said, I really do worry about the potential for abuse of this technology once it's really perfected. What does it mean for society when can no longer enter recordings into evidence?
radarsat1··on The Small Website Discoverability Crisis (2021)
Nice to see this attitude from a mod. I rarely have something to contribute to forums but love to read about people's projects. I've been in the position before of actually, finally, having done something I felt was worth sharing, a super rare occurrence for me, and then posted it and just getting instabanned for "self promotion".. it just feels like such a slap in the face from a community that you were enjoying being part of. Then getting into arguments with mods about it and eventually just having to unsubscribe. It hurts.
radarsat1··on Unix Time reaches 1.7 billion
Does that mean unix date calculations do not take into account the leap second?

https://en.m.wikipedia.org/wiki/Leap_second

radarsat1··on Show HN: Open-source digital stylus with six degrees of freedom
I've always found it fascinating that companies that makes specialized hardware devices like this lean so hard into proprietary software, as if that's the thing that gives them the edge on the competition.

If anything the hardware is the hard part to copy and open source drivers could do nothing but benefit them and ensure wide platform compatibility and longevity, and yet so many companies like this insist on closing things up and forcing hackers to reverse engineer everything. It's bizarre logic to me.

radarsat1··on Ask HN: Which sources cover AI developments without falling for the hype?
They may not be "large" in the same sense that GPT4 is "large" but apart from then simulator stuff, every single one of the models you mentioned is transformer-based. Every one of them basically includes encoders to project different modes of information (images and audio) into a "language-like" space so that it can be compared with and mapped to and from text. I think it's fair to say that language models, if not LLMs, unlocked a surprising amount of power.
radarsat1··on Real-time image editing using latent consistency models
I searched for "latent consistency models" out of curiosity, I guess this is the project (with paper link):

https://github.com/luosiallen/latent-consistency-model

radarsat1··on AWS to start charging for IPv4 usage, but critical services don't support IPv6
Oh I see. I installed awscliv2 after discovering that awscli did not have the ec2-instance-connect command, so I could not follow the documentation and went searching to see if there was an updated package and discovered that one. So I guess you are not supposed to install it using pip anymore. Good to know.

Update: and now I've read through the github issue, oh my I've missed quite some drama.

radarsat1··on AWS to start charging for IPv4 usage, but critical services don't support IPv6
I don't really know any details, but all I can tell you is that this is the output of running it in a fresh environment (masked out the instance details and snipped some output..):

  $ docker rmi amazon/aws-cli
  <snip>
  $ cd /tmp
  $ python3 -m venv venv
  <snip>
  $ . venv/bin/activate
  $ pip install awscliv2
  <snip>
  $ ssh -i mykey.pem ec2-user@i-masked     -o ProxyCommand='awsv2 ec2-instance-connect open-tunnel --instance-id i-masked'
  Unable to find image 'amazon/aws-cli:latest' locally
  latest: Pulling from amazon/aws-cli
  6ebddf7084e9: Pull complete 
  eb7d160dbc3b: Pull complete 
  af71d24e41d0: Pull complete 
  cf26adf9ea98: Pull complete 
  2491adb1df28: Pull complete 
  Digest: sha256:a8c8edb4641672d9ef61873594aa02d1c493341ca30f96e257fdc4d3da17ef9c
  Status: Downloaded newer image for amazon/aws-cli:latest
  
  A newer release of "Amazon Linux" is available.
  <snip>
And now I am logged into the machine. You can clearly see that it downloaded an AWS docker image. Exit from ssh, back to local terminal, and now there is an image on my machine:

  $ docker images | grep aws-cli
amazon/aws-cli latest 817d1061df76 3 hours ago 384MB

So I don't know if I agree that aws ec2-instance-connect doesn't use docker. And also, I don't necessarily have a problem that it does, it just surprised me a bit.

radarsat1··on AWS to start charging for IPv4 usage, but critical services don't support IPv6
To be fair I have a bunch of instances with public IP addresses just so I can ssh into them easily. This situation made me look into how I'd get into these instances otherwise and indeed you can set up an appropriate endpoint and then run "aws ec2 instance-connect" to make a tunnel that gets you an ssh connection without needing the public IP, so just like that my needs for public IPs actually goes down pretty drastically.

On the other hand, on my first try I couldn't get "instance-connect" to work and it turned out I needed a different package "awscliv2" which I had no idea existed.. I've been using "awscli" for the longest time and didn't know there was an alternative and more up to date package available. What a mess.

Also, when running the new one apparently it does a bunch of weird docker magic in the background instead of just being a normal Python program so not sure what to think but I guess it works. If anyone knows a leaner solution to opening an instance-connect tunnel I'd love to know.

radarsat1··on Play the first ever text adventure game in your Linux terminal
Couple of things: 1) the writeup says this has been written in Fortran but the files in the gitlab are .c/.h. Was it rewritten in C at some point? 2) no one's ported this to a browser version yet? WASM maybe? Sounds like a fun project.
radarsat1··on SciPy builds for Python 3.12 on Windows are a minor miracle
A lot of these kinds of routines could be translated to other languages but aren't because they are complex and often unmaintained and no one is really around that understands them well enough to port them to C or Rust or whatever.

There is also the issue that often they were published before adding a LICENSE file was a thing. I've found myself in the position before of having to email professors in some random university to ask them if I can get permission to redistribute such a routine while packaging a library that depended on it. In one case I asked them if it would be possible to update their code with a license (which was just a zip file on netlib) and the answer was, "no, but you have my email". So I found myself having to write something like "distributed with permission from the author, private correspondence, pinky swear" in my copyright file. some of this code is so old the authors aren't around and it would get "lost" in terms of being able to get permissions to use it, I mean it's a potential crisis to be honest, if people really cared to check. (Until the copyrights expire I guess, which is what, 70 years after the author's death or some such?)

Anyway, I wonder if a potential solution is to autotranslate some of these libraries using LLMs? Maybe AI will save us in the end. Of course you can't trust LLMs so you still need to understand the code well enough to verify any such translation.

radarsat1··on Firefox Development Is Moving from Mercurial to Git
Speaking for myself, coming from git, I found the branching model of Mercurial hard to understand. Then it turns out that instead of branches I should have been using "bookmarks"? But which to use when?

I tried to work on a Mercurial-based project (not firefox) but eventually gave up and just used git to track my changes and then prepare, very carefully, specific big merged commits to contribute back. It was easier, because I could never figure out how to clean up properly if I made a mistake -- so this way, I could just wipe the local copy and start over instead. But it was annoying, and I was glad to stop using it.

Completely bewildered by every thread about mercurial full of people proclaiming its advantages and "simpler mental model". Simpler than a DAG with pointers? Sure.

If I remember correctly you even needed to install some kind of external plugin just to do simple rebasing / branch cleanup. Now's the part where you tell me I'm doing it wrong, shouldn't be rebasing etc etc, but I'm just saying, it didn't work for me; and the constant claims that it's somehow objectively "better" but misunderstood get really boring.

radarsat1··on Pix2tex: Using a ViT to convert images of equations into LaTeX code
Indeed, synthesizing the data well is an equally or maybe more important part of this kind of project than the particular choice of neural network architecture.
radarsat1··on Starfish bodies aren’t bodies at all, study finds
I agreed with you until I read the article. It actually makes some really interesting observations about the comparison it's making regarding the typical definition of "head" and how body parts in general relate to gene expression patterns. The images in the article are really cool.
radarsat1··on Pix2tex: Using a ViT to convert images of equations into LaTeX code
Nice idea. This is one of those dream problems where you can just synthesize a ton of data and solve the inverse problem. As a student this is a great way to go for a project, but can be hard to think up.
radarsat1··on Audacity 3.4
Sure, but it's not even enabled by default in the upstream repository. Maybe that's a result of all the fuss about it, but nonetheless..

https://github.com/audacity/audacity/blob/6c2e8a2377542d6722...

radarsat1··on How deep is the brain? The shallow brain hypothesis
> every time some neurologist tried to compare brains to neural networks

Value of this comment aside, it kind of makes me chuckle how casually it (and other comments in this thread) just drops the word "artificial" from neural networks here, specifically when comparing with neurology. The irony is funny. Like, somehow we've forgotten why we call them that in the first place, exactly when talking about the thing that inspired the approach.

radarsat1··on A Generation Lost in the Bazaar (2012)
The way the build system imposes on a C/C++ project is just too much. I've maintained multiple projects now that have ended up supporting multiple build systems just because different people prefer different things. I always end up find myself wishing I could break off the build systems into separate projects so that other people could maintain them but that feels a bit silly. But honestly for a smallish C project you can easily end up spending more time on and having more LoC in the build system(s) than in the actual codebase, it's ridiculous.
radarsat1··on A Generation Lost in the Bazaar (2012)
> It disturbs me that Helm templates became a defacto standard for k8s. Something went very wrong there.

I don't know much about k8s, but this reaction reminds me very much of how I felt when I first learned CMake. "Wait we're replacing all this crap with.. this crap?"

radarsat1··on Why is text of sumissions in low-contrast grey on HN?
I just tried using Firefox Mobile's reader mode and it seems to work really well. All the comments get included, although oddly the usernames do not. But maybe that is not a bad thing ;)
radarsat1··on MemGPT – LLMs with self-editing memory for unbounded context
I don't see why you'd need to stop at "grammar". If you have something like intellisense that can tell you all legal completions (eg. list of member functions of the class of the last-referenced object) then you could use the same approach to limit sampling to function names that actually exist.
radarsat1··on Kalman filter from the ground up
The thing is, to train a NN to estimate your output from your input, you need input-output pairs. KFs are a way of measuring that output in the first place. So they are not even the same class of solutions.
radarsat1··on Kalman filter from the ground up
You could probably call a Kalman filter an ML model if you wish. One major difference is that you usually have to prescribe the prediction model. It doesn't learn to predict the next value like an LLM, instead it learns to find an optimal weighted sum of its prediction and the measured values. So it's a prediction-correction loop that requires an interpretable model, that has the side effect of allowing you to estimate system states. This is quite different than having an arbitrary hidden state with learned structure.

In other words it performs well for certain applications specifically because it allows you to bring in domain knowledge in the form of the process model and known uncertainties. Whereas deep learning models try to generalize the model and learn implicit structure from data.

radarsat1··on The Quest to Recreate a Lost and 'Terrifying' Medieval Mead
Where the hell can you find raw honey though? Anything in the store seems to be pasteurized.

Although, I wonder, if you're cooking it like this anyway, does it matter if it's raw?

radarsat1··on First word discovered in unopened Herculaneum scroll by CS student
> if the method becomes scalable

the machine learning stuff is cool, but it's important not to discount the apparently pretty manual labour still involved in the virtual unwrapping:

> Early in the summer, a small team of annotators (the “segmentation team”) joined our effort. They began mapping the 3D structure of the scroll using tools initially built by EduceLab and improved by our community. By July we had segmented and “virtually flattened” hundreds of cm2 of papyrus.

So, it sounds like it was about a month or two of work, for a single scroll. Although, it probably could be partially or fully automated too, with some effort. Already they developed some tools to help, and I guess it's the kind of task that gets easier after you do it the first time.

radarsat1··on Prophetic Perfect Tense
The idea of prophecy is kind of curious when you think about it. These days if you said you had a vision of the future that you literally saw and were so sure about, people would tell you you were dreaming, and if you kept insisting, delusional.

Plenty of fiction uses "future seeing" as an interesting plot device, (Dune for instance, and currently the Foundation tv series), so we clearly find the concept intriguing, but it is relegated to fiction of course, so likely we generally find it preposterous, and perhaps it's all just biblical references after all.

But it seems that there was a time that people could get away with prophecy as a real claim. It's hard to put your mind sometimes into a pre-scientific mindset, but obviously enough people found these claims compelling. (I realize there is some survivorship bias going on here.). How did they get away with it, I wonder? Just make it vague enough and far enough into the future to be unverifiable? Just make it scary enough that people don't want to take the risk that you are wrong?

When you think about it, it's amazing that some of these stories made it so far and that people even today still believe them. I mean this from a kind of anthropological point of view I guess. Makes you wonder what we commonly believe today that will seem silly 1000 years from now.

radarsat1··on Neurons in Large Language Models: Dead, N-Gram, Positional
Absolutely. Pruning can decrease overall accuracy. The challenge is to prune as aggressively as possible while keeping accuracy as high as possible. It's certainly not an operation that is expected to have no effect.

It should be noted that there are methods for training the network to encourage these prunable neurons, eg sparsity penalties.

← PreviousPage 10 of 34Next →