bytesandbits

the authors of a 2003 article?

What about Ed Boyden!

This is just bait for regulatory capture. It’s the only way in which you can find alignment of Elon, Sam and Dario within the same 24h.

Zoom is malware, has always been.

No Small Actors.

I imagine it is based off World Labs

absolutely crazy level of detail... in awe

I thank Microsoft deeply for the forced copilot crap, almost impossible to remove, that they have put into vs code. Finally after 5 years I have deleted vs code in my Mac! that was the last piece of windows software I still had around. VS code was great years ago, until Microsoft started to push crap into it – and afaik they also made the fully open-source, telemetry-free fork difficult to use with many extensions.

Really, thanks for forcing me into deleting it. turns out vim + Claude Code or codex was much better all along, it really works well for me.

I know a 10x engineer when i see one.

fine. I will port it myself. Real-time, sub 100ms latency. Here

https://testflight.apple.com/join/myNP5XvU

Handy with parakeet is pretty awesome by the way!

here the original NASA photos at high resolution without unnecessary ads.

https://www.nasa.gov/gallery/journey-to-the-moon/

why is a mail client needed in an onboard space computer at all?

anything that doesn't touch the model parameters at all once it has been compiled. for example, in streaming ASR of an encoder-decoder you can get gains in accuracy just by enhancing the encoder-decoder orchestration and ratio, frequency of fwd passes, dynamically adjusting the length of rolling windows (if using full attention). Prompting would be part of this too, including few-shot examples. Decoding strategy is also part of this (top-k, nucleus, speculative decoding, greedy or anything else). Applying signal processing or any kind of processing to the input before getting it into the model, or to the output. There are a lot of things you can do.

we constantly underestimate the power of inference scaffolding. I have seen it in all domains: coding, ASR, ARC-AGI benchmarks you name it. Scaffolding can do a lot! And post-training too. I am confident our currently pre-trained models can beat this benchmark over 80% with the right post-training and scaffolding. That being said I don't think ARC-AGI proves much. It is not a useful task at all in the wild. it is just a game; a strange and confusing one. For me this is just a pointless pseudo-academic exercise. Good to have, but by no means measures intelligence and even less utility of a model.

Apple

4x faster PREFILL not decode. Decode is bandwidth-bounded. Prefill is flops-constrained.

do you run two eSIMs when traveling and if so how is stability / battery life?

incredible work

I am not sure what "edge" device you want to run this on, but you can compress parakeet to under 500MB on RAM / disk with dynamic quants on-the-fly dequantization (GGUF or CoreML centroid palettization style). And retain essentially almost all accuracy.

And just to be clear, 500MB is even enough for a raspberry Pi. Then your problem is not memory, is FLOPS. It might run real-time in a RPi 5, since it has around 50 GFLOPS of FP32, i.e. 100 GFLOPS of FP16. So about 20-50 times less than a modern iPhone. I don't think it will be able to keep it real time, TBF, but close.

regardless, this model with such quantization strategy runs real time at +10x real-time factor even in 6-year old iPhones (which you can acquire for under $200) and offline at a reasonable speed, essentially anywhere.

You get the best of both worlds: the accuracy of a whisper transformer at the speed and footprint of a small model.

sensei karpathy has done it again

parakeet v3 has a much better RTFx than moonshine, it's not just about parameter numbers. Runs faster.

https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

maybe a deepseek v4 distill. give it a few days

its cause of a chain of events.

Next week Chinese New year -> Chinese labs release all the models at once before it starts -> US labs respond with what they have already prepared

also note that even in US labs a large proportion of researchers and engineers are chinese and many celebrate the Chinese New Year too.

TLDR: Chinese New Year. Happy Horse year everybody!

Not trained in Ascend that is BS. Hopper GPU cluster. Please remove that.

wow Mistral really cooked

how are you so fast man

John Giannandrea leaves and things start to work well in the Apple AI department. Maybe just a coincidence :)

I will never understand why when you expect to sell a few hundred units of a product at most, you would decide to go and build something in china. There are instances where building in china might make sense. this is not one of them! you can easily hand assemble everything yourself in America, since this was a 1-year long time-frame. There is phenomenal CNC shops in the east coast (and also some in the mid-west) if you don't mind the lead times. You don't have to deal with tariffs, imports, and more importantly THEY SPEAK YOUR LANGUAGE. Yes this sounds dumb. Like, with Google Translate you shouldn't have a problem no? Think again! From my experience with Chinese metallurgic, they are clumsy and often mess up even straight forward designs. Yes, they are fast, but at what cost! They will redo it as many times as you need. But the headache, man the headache of dealing with it is not worth it. On the other hand some of the issues here are expected and fun! well maybe not fun, but part of the process, like the swapped components in the board. it happens to the best of us and that's okay! but it would always be easier to debug when you can drive yourself to the CNC factory, or at least have a video call with the tech in your own language. Also, regulations in china is whole different world and not at all straightforward to navigate for westerners, at all. I really don't like that people turn directly to china cause "that's what everyone does anyway". nu-uh. also, China is not the only option! Eastern Europe, such as Poland, can get you there too for reasonable prices, and have tons of phenomenal electronics and manufacturing engineers. Tons.