woodson

6 ms·

61.
by woodson 1y ago
Because he was German Bohemian? ( https://en.wikipedia.org/wiki/Sudeten_Germans#Austria-Hungar... )
62.
by woodson 1y ago
Check out this model based on the same architecture for Japanese: https://github.com/nu-dialogue/j-moshi
63.
by woodson 1y ago
Sounds like the same farce as with the forced Pixel 4a update. Not sure if they learned anything from that, other than to ignore the issues and move ahead anyway.
64.
by woodson 1y ago
As someone who has designed neural network architectures, I disagree.
65.
by woodson 1y ago
I think that, for the purposes of creating high-quality Free audiobooks, the issues are essentially the same with human-generated recordings as with AI generated ones. The recording quality and faithfulness to the original text (both in ter
66.
by woodson 1y ago
The speech data collected by this project has been used for more than a decade to build automatic speech recognition and text-to-speech synthesis systems (see LibriSpeech, LibriTTS, LJSpeech). It definitely has been a benefit to AI.
67.
by woodson 1y ago
How does this differ from SuperBPE, which seems to pursue a similar goal? https://arxiv.org/abs/2503.13423 Looks like parallel invention. (I’m not associated with the paper or its authors.)
68.
by woodson 1y ago
There has been some innovation, e.g. SuperBPE ( https://arxiv.org/abs/2503.13423 ) claims substantial gains on 8B-sized models.
69.
by woodson 1y ago
Slight correction: It worked in llama.cpp via the CLI tools, but not in the llama-server (OpenAI API compatible interface).
70.
by woodson 1y ago
Nowhere was I talking about state actors. Most fingerprinting is done for ad targeting, as far as I know.
71.
by woodson 1y ago
Maybe it includes a list of fonts installed on your system, your screen resolution, etc. You don’t need much to get a “fingerprint” that is anonymous but can be correlated with those collected by other tools’ telemetry.
72.
by woodson 1y ago
Human-to-human conversational patterns are highly specific to cultural and contextual aspects. Sounds like I’m stating the obvious, but developers regularly disregard that and then wonder why things feel unnatural for users. The “median del
73.
by woodson 1y ago
Oh, that’s pretty slow. Have you tried using quantization (int8 or int8_float32)? In my experience that can help speed up CT2 execution. Personally, I haven’t had much luck with small-ish decoder-only models (i.e., typical LLMs) for transla
74.
by woodson 1y ago
Not sure if you tried that already, but ctranslate2 can run BART and MarianNMT models quite efficiently, also without GPUs.
75.
by woodson 1y ago
They’re both lossy. They use a VAE-VQ type architecture trained with a combination of losses/discriminators. The differences are mainly the encoder/decoder architecture, the type of bottleneck quantization (RVQ, FSQ, etc.) and of
76.
by woodson 1y ago
Some of these small models still have their uses, e.g. for summarization. Don’t expect them to fully replace ChatGPT.
77.
by woodson 2y ago
So are CNNs, but I guess their popularity heavily increased at that time, to the point where it made sense to optimize the hardware for them.
78.
by woodson 2y ago
To be fair, that’s pretty far from Vienna. Those seem to be separate incidents.
79.
by woodson 2y ago
The discussion thread on reddit is eye opening. People think it’s normal to train on your test data. EDIT: Ok, seems this sentiment wasn’t shared by other commenters (or Reddit’s UI hid those comments).
80.
by woodson 2y ago
These observations are interpretations of behavior viewed through human eyes. (What’s the concept of a “hat” to an orca?) Perhaps we interpret their behavior this way because it is our behavior.
81.
by woodson 2y ago
Ok, so very different from DARPA (the original) then. Look into EU HORIZON programs.
82.
by woodson 2y ago
There’s plenty of R&D funding in the EU. How would that help get a European cloud?
83.
by woodson 2y ago
It’s often called endpoint detection (in ASR).
84.
by woodson 2y ago
From what I’ve seen, they overheat a lot just from having a browser and VSCode running.
85.
by woodson 2y ago
This project apparently uses MLX, Apple’s ML framework, which doesn’t use Python’s pickle library that’s behind the safety issue. There are several options for storing models/tensors in MLX, none of which I think have such (de-)seriali
86.
by woodson 2y ago
It uses MLX ( https://github.com/ml-explore/mlx ), Apple’s ML framework, for running LLMs.
87.
by woodson 2y ago
I used to have papers read to me via TTS when I had a long commute. This was before the current crop of neural TTS, mind you, so the quality and naturalness wasn’t as good, but it was good enough to tolerate and to get the gist of a paper.
88.
by woodson 2y ago
It was used quite a bit in WFST-based speech recognition decoding, before so-called end-to-end approaches (CTC, RNN-T, AED, etc) took over.
89.
by woodson 2y ago
There’s also k2 ( https://github.com/k2-fsa/k2 ) which implements a lot of FSA and FST algorithms in CUDA, with PyTorch bindings.
90.
by woodson 2y ago
Yesterday’s submission: https://news.ycombinator.com/item?id=42865619

More ›