lunixbochs

13 ms·

91.
by lunixbochs 5y ago
`spctl -v --assess -t execute` is crucial. My app layout is fairly complicated, so I'm sure I'm exercising a lot of the corner cases: https://news.ycombinator.com/item?id=26996223 I check that executables don&#x27
92.
by lunixbochs 5y ago
I can't remember, because I did this a few years ago, but I think there was some other code signing benefit to not putting all of Python in Frameworks as well.
93.
by lunixbochs 5y ago
I auto notarize my app when I push to a release-candidate branch even if I don't deploy it. I also have a general code signing CI test that catches stuff before notarization would. I believe I've never, in thousands of pushes to t
94.
by lunixbochs 5y ago
That seems to work if you put the entire Python stdlib under Frameworks but not if it's somewhere else. I do the `find` thing because I pre-sign my libraries when I build them to save a bit of time during app build.
95.
by lunixbochs 5y ago
I rearranged my CI graph to run notarization in parallel with my unit tests, and turned on GitLab's DAG to allow jobs from different stages to run at the same time as long as their dependencies are satisfied, so the impact of the notar
96.
by lunixbochs 5y ago
This is my notary script for a DMG: https://gist.github.com/lunixbochs/0ceeb23be5c3d5c6748165e61... ./notary.sh notarize app.dmg ./notary.sh staple app.dmg I store a password as a keychain entry na
97.
by lunixbochs 5y ago
In python there are a handful of .so files in subdirectories you need to sign. You can basically `find . -name \*.dylib -o -name \*.so` and codesign those.
98.
by lunixbochs 5y ago
I ship my app as universal2 with latest Python (including modules and complex shared objects preinstalled to site packages such as Pytorch, numpy, opencv), xpc services, a launch daemon, a handful of frameworks, and ~137 shared objects in d
99.
by lunixbochs 6y ago
I ship an app (Talon) which runs many kinds of wav2letter ASR models on consumer CPUs. It has real-time inference pipelines for both pytorch and flashlight. (I wrote the pytorch code). Performance of both inference pipelines is fine and com
100.
by lunixbochs 6y ago
I'm specifically talking about the new Conformer model, available in early access as of ten days ago. What you tried was likely the previous (circa 2018) model, which is much less accurate than Conformer. This is a demo of Conformer in
101.
by lunixbochs 6y ago
What are you building?
102.
by lunixbochs 6y ago
I have users who have intentionally switched their speech engine from the latest version of Dragon to Talon, for both dictation and commands. Talon is cross platform and directly targets accessibility use cases (far more than just speech in
103.
by lunixbochs 6y ago
I'm in the on-device space, here's a recent demo of my engine for a niche use case (it also does large vocabulary dictation very well): https://twitter.com/lunixbochs/status/1378159234861264896
104.
by lunixbochs 6y ago
Here's a quick zero shot text compression attempt with GPT3, compressing then decompressing some of my sentences picked from upthread. Looks pretty impressive imo. Notes: These results include every attempt I made and are not curated i
105.
by lunixbochs 6y ago
I implemented a "conceptual" lossy text compression a few months ago by using word2vec (glove) to convert each word to a list of floats, laid the result out in a square, then jpeg compressed it at different quality levels. This en
106.
by lunixbochs 6y ago
Very interesting! Well, I'm playing with the concept a bit now. Currently writing a skia client to generate glyph SDFs as images so I can just use pytorch.
107.
by lunixbochs 6y ago
Something like mixup augmentation might work here as well for training - blend two SDFs and say "this is 20% A and 80% B". Someone could also train a model that given some glyphs of a font predicts the rest of the glyphs. Then we
108.
by lunixbochs 6y ago
It's more than a CFG, it simulates line editing, so you need to port some code either way.
109.
by lunixbochs 6y ago
Hm, but I won't know if the text/terminal is utf8 or not going in.
110.
by lunixbochs 6y ago
Interesting: https://man7.org/linux/man-pages/man4/console_codes.4.html I'll definitely need to add support for that. Possibly annoying to strip that with utf8 involved. Any idea how many terminals suppo
111.
by lunixbochs 6y ago
My vtclean project [1] cleans terminal escape sequences. It handles the full ECMA-48 spec, and every behavior described by all of the vt100 documentation I could find, and observed behavior of many command line programs and terminals. It ha
112.
by lunixbochs 6y ago
The demonstrated commands in this article aren't best practices in Talon, it's more like: funky hello slap funk print string hello
113.
by lunixbochs 6y ago
They're possibly something like a dosdude patcher or modified bootloader to run an OS like Catalina on it.
114.
by lunixbochs 6y ago
This was exploited by one team during an ML challenge (ai-han-solo) at Defcon CTF Finals 2019. Dropped a meterpreter every time you loaded that team's tensorflow model.
115.
by lunixbochs 6y ago
It's also common for papers with code to not generalize well at all.
116.
by lunixbochs 6y ago
I'll check, thanks! Or you can tweet at me, as I see you found that thread as well.
117.
by lunixbochs 6y ago
Can you post a quick summary of what's running and the amount of ram allocated to the vm? Is there one process claiming to have written terabytes of disk in Activity Monitor? Would be curious to repro as my M1 is nowhere near this and
118.
by lunixbochs 6y ago
Wait do you have an easy repro for this that doesn't go over ~16GB total memory used? I'm not actually convinced it's caused by swap.
119.
by lunixbochs 6y ago
The neural engine for simplified AI models is separate silicon from the GPU.
120.
by lunixbochs 6y ago
The main speech engine is English. Due to the large amount of open data available for English model training, even if you have a strong accent, you'll probably get better command accuracy by tweaking Englishy commands rather than tryin

More ›