Farhan Syah

Founder/Creator of and ML/AI engineer focused on Rust-powered high-performance computing and distributed systems.

[Farhan Syah @farhansyah90]

Hey, quick intro 👋

I’m the founder of NodeDB, currently running a SaaS + IT business. Most of what I do now revolves around building systems.

What I share here will mainly be things I’m actually working on:

ML / AI framework and infrastructure, databases (NodeDB), edge to cloud distributed system.

Follow if you’re interested in these stuff. 🚀

[Farhan Syah @farhansyah90]

Did you hit this too, and which way did you go?

I picked reserve-before, but I am curious whether anyone regrets it later.

[Farhan Syah @farhansyah90]

The decision I spent longest on had nothing to do with the model.

It was where to put control tokens in the vocabulary.

[Farhan Syah @farhansyah90]

Append them after audio and control ids split across two disjoint ranges.

Every mask, every resize, every range check carries that split forever.

[Farhan Syah @farhansyah90]

Anyone else training speech this way?

I want to know what breaks at scale before I find out the expensive way.

Code is open: https://t.co/CtSSwdu4kO

[Farhan Syah @farhansyah90]

I am training a text-to-speech model as a language model.

One causal decoder. One flat vocabulary.

Text ids go in. Neural audio codec ids come out.

[Farhan Syah @farhansyah90]

No separate acoustic model. No vocoder head bolted on the side.

The same next-token prediction that writes a sentence also writes the audio.

[Farhan Syah @farhansyah90]

@paulg It's not to late. Being doing that since last year.

[Farhan Syah @farhansyah90]

Around it: a tokenizer, a tensor library, a regex engine, an inference server, a notebook runtime.

30+ crates. Two years. One maintainer.

https://t.co/YldgKEBnfy
https://t.co/JliyAHpx3c

[Farhan Syah @farhansyah90]

I scrapped this database twice before the third version worked.

NodeDB: multi-model, PostgreSQL-compatible, one binary. Vectors, graph, documents and time series in the same transaction. Embedded or clustered.

[Farhan Syah @farhansyah90]

The first two attempts failed because I was patching a design instead of building one.

The third works because I finally understood why the first two did not.

[Farhan Syah @farhansyah90]

Everyone argues about whether AI can write real software.

I would rather just show the repos.

https://t.co/YldgKEBnfy
https://t.co/JliyAHpx3c

[Farhan Syah @farhansyah90]

I direct it, review it, reject it, and understand it down to the storage layer.

That part is not optional. I scrapped NodeDB twice. The third version works, and it only works because I understood why the first two did not.

[Farhan Syah @farhansyah90]

@grok What's the difference between DTS, ELI5 and ASD-STE100?

[Farhan Syah @farhansyah90]

I tried ELI5, ASD-STE100, and the recently released 'concise' style from Claude.

After a lot of experiment, this is my best output style.

It preserves all the details, while mantaining the quality.

Not the best at compression, but I like the response the most.

Try it:

https://t.co/QuzCtURNHg

Let me know your experience

[Farhan Syah @farhansyah90]

This is before and after using DTS.

Tested on Opus 5 using Claude API.

[Farhan Syah @farhansyah90]

Upcoming new Huggingface PipelineTokenizer is insane.

It performs better then gigatoken in certain workload.

I thought splintr already doing well, being competive with gigatoken.

Yet, another huge ceiling is coming.

@art_zucker good job 👍. Waiting for the official release.