Alex

i. do not have a math background yet alot of finetuning models for ANE optimization is intuition maxing thanks to coding agents.

alot of metalearning is related to understanding whats the fundamental unit a certain field use to measure or benchmark "progress". like how you can't have an economy without an agreed upon currency

tried to fuse mel spectrograms and encoder and decoder models for Parakeet RNNT and CTC before btu the enigmatic ANE Scheduler doesn't make this easy

Weight matrices and Upsampling is the deduction of additional information from limited information. i.e 6*6 matrix input into a 8*8 upsample matrix based on the position and values of the individual coordinates

been using claude code to grill me on coding concepts. man it is brutal in its feedbacks. it is able to politely call you a dumbassss in the most formal way possible

🥳  FluidAudio's Parakeet TDT 0.6B v3 CoreML has just crossed 1,000,000 total downloads and it has 472,749 monthly downloads, ranking #648 out of 2+ million huggingface.co/FluidInference…

In less than a year, FluidAudio's Parakeet TDT v3 CoreML model has crossed over 200,000 monthly downloads, ranking # 1,079 out of 2M+ models on HuggingFace (top 0.054%). huggingface.co/FluidInference…

@y_earu Hi Andrea, Glad to see you are enjoying parakeet eou. I've added your demo to our Video Demos section lmk if you have any other demos/projects using FA you'd like featured! #video-demos" target="_blank" rel="nofollow noopener">github.com/FluidInference…

insane @Alex_tra_memory

@y_earu @sach1n @kyutai_labs we recently added some improvements to the model to support long term sessions released in v0.13.5 github.com/FluidInference…

@sach1n @Alex_tra_memory @kyutai_labs super cool, hows this going?

got claude code rated limited, codex & gemini3 still aren't that good for backend coding as a well. github.com/FluidInference…

@yomlapp we support that as well, huggingface.co/FluidInference…

@Alex_tra_memory No streaming api?

FluidInference/parakeet-tdt-0.6b-v3-coreml just crossed 133K monthly downloads on HuggingFace placing it at #1,458 out of 2.7 million models (top 0.054%). huggingface.co/FluidInference…

devin has been surprisingly great at code reviewing. i should not be surprised that an ai trained by competitive programming code would be so good at debugging. it does add incentives to do some leetcoding to be better at reading & debugging code.

There’s been a lot of bots making issues on GitHub lately. Frankly it’s abit demoralizing. Like why not code up your own solution. Not to mention Responding to a vibecoded issue without a response back. Like what did you even make an issue in the first place.

kimi2 is pretty great at being a teacher, i have learned alot from moonshot vs claude or chatgpt. the issue with western models is their wordceling. just verbose nonsense with nothing useful. it felt more like western models were trying to convince me i understood how the models work when i know i don't. but moonshot is so blunt and to the point. it works. i am able to ask genuine follow up questions and expand my knowledge no issue.

@BHolmesDev building a second brain on the mac

Also working on an open source skill to teach Claude Code, OpenCode, etc how to use Oz

Cloud hosting for agents is great, but the fact that everything runs through a CLI, down to the environment config, is so huge. You can ask any coding agent to make sandboxes, set up agents to run on a schedule, write a TypeScript server to trigger agents from discord events...

@hubeiqiao heavenly reward for all the hard work

After your laptop ran out of battery and you randomly found a sushi restaurant nearby... one of the most amazing dining experiences!

Vibe travelling alone Kamo River in Kyoto.

alot of reading research papers is just learning how to shape rotate with very confusing variable names & sometimes vague unclear blackboxes like trained embedding models. alot of mist in understanding how models work.

@jenzhuscott china has data and scale, america has the chips

Fully Seedance 2.0 generated movies are within reach. Incredible quality. Remarkable progress. But as I always say, plastic flowers won’t make real flowers obsolete no matter how real they look. Brilliant human acting connects you at a different level.

@cherry_cc12 this i can attest with building my own oss community

Strong models grow with strong communities

without coding agents, fluidaudio might have taken years to complete. there was so little we know about coreml. the documentations were very lacking & reading the model pytorch codes were incredibly cryptic. not to mention reading the research papers took the most effort. from years to months to now even weeks, it became alot easier once i had accumulated enough coding examples, documents for our coding agents to grasp the conversion process while also being able to explain back to us how they work