HNHacker News
TopNewBestAskShowJobs

radarsat1

4,157 karma · joined May 28, 2011

Info: https://sinclairs.gitlab.io

Email: <myusername>@gmail.com

submissionscomments
radarsat1··on Inside the proton, the ‘most complicated thing you could possibly imagine’
> if one of the values were even slightly different we wouldn't have anything like the universe we see today

This is a hallmark of a chaotic system. It's not impossible but the chances of sitting exactly on such an unstable point seems very low. It seems more likely that the constants are some optimum in a basin of attraction, a stable point in some higher order dynamic system.

radarsat1··on Neural network training makes beautiful fractals
I'm really curious what effect the common tricks for training have on the smoothness of this landscape: momentum, skip connections, batch/layer/etc normalization, even model size.

I imagine the fractal or chaos is still there, but maybe "smoother" and easier for metalearning to deal with?

radarsat1··on A Git client for simultaneous branches on top of your existing workflow
Not sure this is something I need. One thing I do need though, maybe someone knows a solution:

Often I find myself maintaining a handful of "local" changes. I make some changes that only make sense in my local environment, that I don't want to push.

What I end up doing is maintaining these changes as a commit, committing on top of them, and using `git rebase -i` to periodically move them up. Then before I push, I have to temporarily rewind the branch to remove them, push, then cherry-pick them again.

It's all a bit awkward and I would love a tool that maintains a kind of "virtual branch" that isn't shown but is automatically re-based on top every time I make a commit, maybe letting me resolve conflicts or even telling me ahead of time if I've created one, before committing.

Someone must have already solved this, or am I doing it all wrong?

radarsat1··on The Annotated S4
This was an excellent write up thanks. It'll help me understand the Mamba work a lot more.

I still find it really confusing how a linear model can perform so well.

radarsat1··on Oslo acquires ChatGPT for 110k students and teachers
This reminds me of something more innocuous but funny that happens to me. Innocuous because it's not on purpose.

When I forget to start Slack and someone sends me a message, Slack sends me an email saying "You have a new direct message in <channel>".

Well, it so happens that when I get this, the notification dialog that pops up to tell me about the email is just the right width that it gets cut off as "You have a new dire.."

The first several times this happened I went straight to Slack eager to find out what was so important.

radarsat1··on Gnuplotlib: A gnuplot-based plotting backend for NumPy
Absolutely yes, but you'd have to design that for each application you develop. I'd love an out of the box solution. Then maybe skip the file and use pipes or sockets.
radarsat1··on Matryoshka Representation Learning
Any relationship to residual vector quantized embeddings?
radarsat1··on Gnuplotlib: A gnuplot-based plotting backend for NumPy
I don't mind matplotlib at all but I just find it annoying in one of my most important use cases which is to update the plot during a long calculation.

You gave to call fig.canvas.draw but it doesn't always work and sometimes plt.pause works but sometimes it only partially updates the display and you have to call it multiple times. (At least this in is my experience with TkAgg backend.)

Meanwhile you can't interact with it.

I would love a version of matplotlib which lives in a separate process and gets updated via nonblocking communication.

radarsat1··on Gitlab's ActivityPub architecture blueprint
I've wondered if there might be room for services that use email as their communication channel. Something like this would be perfect for it. The forges would send emails to communicate in issues, information is distributed via mailing lists, but forges parse and display the information in their own way. They become fancy single-purpose email clients basically.

With such a system, people could also contribute by writing the same emails by hand, as long as the format remains accessible. And no special protocols necessary, all existing mailing list infrastructure can be used instead of designing new complicated ways of doing the same thing.

radarsat1··on Gitlab's ActivityPub architecture blueprint
No, I mean a standard for the forge things like issues and comments and pull requests. The things that would be carried over activitypub here.
radarsat1··on Gitlab's ActivityPub architecture blueprint
A tool that automatically scraped github issues etc into a standard format and committed them to the repo could accomplish the same thing, no activitypub needed!

Then collaborating peer forges could clone the info straight from git and submit changes etc. I think git already has all the sharing tools necessary, what is needed is standards and converters.

radarsat1··on I just wanted Emacs to look nice – Using 24-bit color in terminals
I also use emacs graphically, somehow I managed to switch after years of using it in the terminal and now I prefer it. However, I've never managed to settle on any of the terminal emulation modes. Just too many keyboard shortcut conflicts and complicated magic to remember how to copy and paste between different types of buffers.

These days I've pretty much settled on having and emacs window and GNOME Terminal side by side as my two main work windows, and it's pretty comfortable. I like that I can mouse select and copy from the terminal window and then C-y into emacs and vice versa.

If I really have some need I might still open a shell or ansi-term session but it's pretty rare now. Sometimes it's convenient when I find myself in a remote shell session and want to have a text editor and shell vertically tiled, I'll just run emacs with ansiterm because for the life of me I can't ever remember the screen keyboard shortcuts for splitting windows.

radarsat1··on On the dimensionality of spacetime (1997) [pdf]
I think what he means is that chips are 2D, so heat is radiated orthogonal to the plane of the circuit -- up or down.
radarsat1··on The failure of self-checkout technology
The main problem with the way they do the check is that it isn't going to catch any theft. They make you unpack your bag and they start asking for items to scan. They literally let you choose which ones they scan. And they want to scan like 10 items lately so it really makes you unpack the whole thing which is annoying. But never have I had them actually trying to find an item in my bag that is not something I declared. They seem pretty focused on finding things I did scan and confirming them. Mostly because it's done by a teenager who just wants to get the scan done and move on.
radarsat1··on The failure of self-checkout technology
That's exactly how it works in the Netherlands, you just scan your items and go. There are no weight checks and you can scan items more than once and edit and remove things from the list. Apart from getting a poorly executed random check sometimes (which is far from effective for the purpose of avoiding theft imho) it actually works really well.
radarsat1··on SQLite 3.45 released with JSONB support
I find it really odd that the decision was made to store ints and floats as text in JSONB. It seems to defeat a lot of use cases for it as far as I can tell. There are few solutions for storing and retrieving/querying unstructured numerical data.
radarsat1··on Coding Self-Attention, Multi-Head Attention, Cross-Attention, Causal-Attention
It seems that there are some popular attention methods such as relative embeddings and rotary embeddings (rope embeddings?) that are not possible to implement using pytorch's implementation, if I understand correctly. Do these then require the "slow path" versions that can be more easily modified?
radarsat1··on The Internet Is Full of AI Dogshit
Opinion pieces "against AI" are getting just as boring and unimaginative as AI-generated content. I feel like all the arguments have been made at this point.
radarsat1··on The Curious Case of MD5
Also reminds me of MODULE_LICENSE and EXPORT_SYMBOL_GPL in the Linux kernel.

https://www.kernel.org/doc/html/latest/process/license-rules...

radarsat1··on Possible Meissner effect near room temperature: copper-substituted lead apatite
Isn't the whole point that it's something we might not predict from what we understand about the materials so far? Why would it be likely that a simulation would do better than theory at predicting unknown experimental results?

You might hit on some interesting interactions between known properties that haven't been investigated but I would assume the real interesting results are from things we just don't know to model, or how to model.

radarsat1··on Termdbms – A TUI for viewing and editing database files
Use that all the time actually! But the main argument for CSV that I most often hear is "at least I can just load it into any text editor.." I figured having a pretty good TUI program might be just as good as using a text editor, but we'll see.
radarsat1··on Termdbms – A TUI for viewing and editing database files
Nice, a tool like this might actually get my team to finally move from csv to sqlite. I was thinking parquet or feather might be interesting alternatives to csv for us though, any plans to add support?
radarsat1··on Clusters of brain cells grown in the lab have potential as a hybrid bio-computer
I would buy this as interesting if it's compared to a control which I guess would be the confidence interval of a series of similar tests again random projections.
radarsat1··on AI’s big rift is like a religious schism
> Things become incomprehensible without machines doing most of the work — including explanation. Eventually you, or your kids, won’t even understand what’s going on 90% of the time…then less and less.

This also reminds me of the Eloi in the Time Machine, a book written in 1895!

radarsat1··on Show HN: I Remade the Fake Google Gemini Demo, Except Using GPT-4 and It's Real
I think if the same LLM were trained on audio and video input instead of text, and produced audio output, including silence tokens, then the notion of time would get "built in". Audio continuation without translation to text has been shown to work. Mixing it with text is also possible. But all this would require a massive network that maybe even be difficult for the world's biggest companies to train and serve at any kind of scale. So it's more of an engineering problem than a theoretical one imho.

Also imho, I think until the context/memory problem is fully solved we won't really see the AI as having any kind of agency. But continuous, low latency interaction would certainly feel like a step towards that.

radarsat1··on Chipsynth C64 is an emulation of the SID so good, it can replace hardware
Ah I see that's pretty insane. So the character of the sound doesn't just come from the circuit but is really intrinsic to the chip design. That's some really clever engineering.
radarsat1··on Chipsynth C64 is an emulation of the SID so good, it can replace hardware
Yeah, exactly. So I was wondering of the same analog circuit has been built outside an IC (ie. discrete components) but according to the other response, the IC specifics are integral to the result, which is crazy and really cool.
radarsat1··on Chipsynth C64 is an emulation of the SID so good, it can replace hardware
Although its sound is legendary, I can imagine it's not that complicated a chip. Had anyone ever tried to replicate it with discrete components?
radarsat1··on Researchers have discovered magnetic monopole quasi-particles
Stupid question. If you develop a quasi-particle monopole, which I guess is something that is not a monopole but behaves like one, can you use it in such a configuration and get the behaviour you are describing?
radarsat1··on JSONB has landed
I have some cases where sometimes I want to store a vector of floats along with a data item. I usually don't need to match against the column, just store it. I know I could use BLOB but often I just use JSON for this kind of thing so that I don't have to deal with data types and conversion. It's wasteful and imprecise though due to the string conversion. Is JSONB a good middle ground option for these cases?

Edit: Sorry, just saw the comment below by o11c,

> Sqlite's JSONB keeps numbers as strings

which means the answer to my question is basically "no".

← PreviousPage 9 of 34Next →