ZD1908

Releasing Brontes: A modified Wave U-Net architecture for audio super-resolution. This one is trained to operate on NeuCodec outputs. I'm releasing a general 30M checkpoint on a variety of speech. See links in replies. All on MI300X thanks to @HotAisle @AIatAMD

Introducing M²RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling We bring back non-linear recurrence to language modeling and show it's been held back by small state sizes, not by non-linearity itself. 📄 Paper: arxiv.org/abs/2603.14360 💻 Code: github.com/open-lm-engine… 🤗 Models: huggingface.co/collections/op…

If you're cold emailing people, you have to prepend [Not AI slop] to the subject line, and write the whole thing with your own two hands.

Hacker News is such a primitive platform it hurts. Even 4chan has autorefresh.

@Wolvan1 Bitwise Operator's shaking that voluminous derrière.

Post OCs and/or art ideas pls

@CounterStrike The ArmAfication of Counter-Strike is upon us.

For your consideration, an update pertaining to Guns, Guides, and Games: steamcommunity.com/games/CSGO/ann…

@ad0rnai >terminally offline >on X (formerly Twitter) Do you also go to butcher shops looking for vegans?

I am hiring someone who is: - terminally offline - not in any group chats - has covered the curriculum of the great books program - idiosyncratic - slightly off-putting

All of this was iterated start to end on a single AMD Instinct MI300X for ~5 days, thanks to @HotAisle and @AIatAMD. Audio quality isn't the best, but only so much one can do with few parameters. Scaling up will be key.

GitHub: github.com/ZDisket/vits-e… HuggingFace Demo: huggingface.co/spaces/ZDisket… This is a highly upgraded VITS trained on LibriTTS-R + VCTK datasets, both fully open. Speaker encoder is Resemble AI's Resemblyzer.

VITS EVOlution, my TTS model: 1. ~31M TTS model and speaker encoder in ONNX format. Faster than realtime on CPU 2. Natively outputs 48KHz audio 3. Voice cloning, or voice blending--mix two or more speakers to make a new voice! 4. Apache 2.0, use anywhere without worry Links in replies:

My model is working and ready for release tomorrow but I came down with a cold today.

reporting to a middle-aged girl dad will change your life

@giffmana Bigger models, plus switch from mostly convnet-based--most of the time U-Net (locally coherent, globally weak) to DiT (locally and globally strong), although newer conv is starting to emerge. x.com/miru_why/statu…

I have a question about last year's image-generation progress, wonder what y'all think. How did we go from all models consistently getting fingers wrong, to all models consistently getting them right? This "flip" seems to have happened basically across all companies/models at the ~same time. Even "random" non-frontier papers seem to get it right? Or they just cherry-pick the figures?

Using DistilHuBERT features as speaker encoder also failed. Time to throw GPT 5.4 at the Resemblyzer repo and have it modernize the pipeline.

After months of work, today I’m releasing Foundation-1. A SOTA text-to-sample model built specifically for music production workflows. It may also be the most advanced AI sample generator currently available - open or closed. • ~7 GB VRAM • Entirely local • 100% free 😁

@jparkjmc @nikitabier You could drop an email address instead.

got way more inbound than i anticipated tried to go through all DMs and got rate limited 😵‍💫 @nikitabier can you un rate limit me, agi depends on it

Vercel: makes open source project Cloudflare: forks it Vercel:

@JasonBotterill French, dog name, French.

Codex has subagents now but bro what the fuck these agent names are somehow worse than Grok's Benjamin. I don't want Russell going through my family photos

I should get into rocketry, I want to make guided missiles for delivery of medical supplies.

@Alfauz19767861 Peronism is like pizza: there's a flavor for everyone.

My understanding of Peronism

Well... I was going to release a zero-shot modeltoday, but my speaker encoder sucks and collapses every voice into an audiobook reader, as it was trained on a small amount of data. I'm going to use DistilHuBERT features instead.