Andrea Silvi

AI and Computer Science Engineer | Algorithmic Finance Developer

[Andrea Silvi @ansi_code]

@paoloardoino @keet_io 2/2 however I think we need to wait for a proper WebGPU adoption on modern browsers to achieve state of art models inference.

Here is another module I hacked months ago that could bring NumPy like features on Js leveraging on WebAsembly. (Still PoC)

https://t.co/GNpEpR5FQW

[Andrea Silvi @ansi_code]

@paoloardoino @keet_io Great, I’m trying to raise awareness on offline/privacy preserving and censorship resistant neural network inference too.
Worked on this months ago (still PoC) for Browser or Node.js and support by design also IPFS pinned models.

https://t.co/69UKqWKkMU

[Andrea Silvi @ansi_code]

I agree with you that ObservableHQ is a good platform to do a variety of things. I also agree that Tensorflow.js is without doubt a masterpiece made by Google for Machine Learning.

However JavaScript continue to lacks of something that made Python suitable for Data Science. I think that one of those missing pieces of the puzzle is a proper Lightweight free&open-source library for N-dimensional array (SIMD optimised?) operations the same as NumPy.

Exploring lot of repos you can see production live libraries that use horrific self-made functions to do basic stuffs on TypedArray(s). I‘ve seen a lot of horrible codes in my hacking days, but when recently I’ve seen a 2k+ stars repo running a TopK (for LLM) extraction by entirely sorting the array (50k items) and then picking the top k=5 elements… I’ve spotted this market need.

Btw writing my other weekend project about LLM+IPFS I personally needed something like NumTs.
https://t.co/kgRo6YVkeF

[Andrea Silvi @ansi_code]

@palashkaria @tesseralis @LaptevMartin I think now Js (in 2023 supporting WASM or WebGPU and many more) could be one of the best candidate for the next generation of data scientists. It “only” lacks of a proper Math libraries. If you think @tesseralis Jupyter like notebooks editor should be built in JavaScript instead

[Andrea Silvi @ansi_code]

@GBallarani I experienced it few years ago in Kyoto, Japan. It was not a test but a real alert (with the siren ringtone at maximum volume) in the middle of the night advising to evacuate houses due to a big typhoon. Everything was written in Japanese.😭

[Andrea Silvi @ansi_code]

@GBallarani @jakidale @jakidale @GBallarani quando volete creiamo un contenuto a tema “sfatare i falsi miti dell’informatica”
Ultimamente, soprattutto sull’AI, ne ho sentite davvero di belle ⚰️

[Andrea Silvi @ansi_code]

@GBallarani @jakidale e perché quel computer è proprio di Jeff Bezos? 😉

[Andrea Silvi @ansi_code]

@cipherproxyllc @willdepue Currently rewriting the full inference by scratch using my other tensor library project (WebAssembly SIMD)

https://t.co/GNpEpR581o

[Andrea Silvi @ansi_code]

@hardmaru WebGPU will be the best candidate for a web based dense matrices backend. But before we should ask if we really need it. I think quantised and sparse matrices ops computation on CPU (SIMD) is faster and good enough.

[Andrea Silvi @ansi_code]

@hardmaru I absolutely agree. We are limited in using dense matrices. Personally I’m exploring things to implement a custom JavaScript/WebAssemby backend for sparse matrices. It would enable LLM to run in browser. I will post it when ready.

https://t.co/69UKqWKkMU

[Andrea Silvi @ansi_code]

@skirano I love the humankind endless effort to "Humanize" Machine Learning models. But, as it is now, for me they are just (lot of) matrices, graphs and statistics.😄

[Andrea Silvi @ansi_code]

@itsafiz @midjourney @OpenAI @elevenlabs @D_ID_ @christianortner This post is 100% built using Artifical Intelligence without human interactions (after setup) and without using external third party services (excluding GPT 3.5).
It automatically publish on all social networks. @HumanTimesMedia

[Andrea Silvi @ansi_code]

I think @AssemblyScript powered math/tensor library can enable all Javascript developers use it for machine-learning related tasks.
In Ts/Js world we are like in Python but without NumPy.
This is going to change with the new library called NumTs (PoC on NPM).
What do you think?
https://t.co/5qMy4MvUYp

[Andrea Silvi @ansi_code]

@nonmayorpete I think another big leap will happen once a large language model can fit and run inside a web page. That’s why I developed this PoC here (it supports also model pin/pull on IPFS) https://t.co/69UKqWKkMU

[Andrea Silvi @ansi_code]

@xenovacom @huggingface In numts I wrote a cross-compatible backend (AssemblyScript or TypeScript) with concrete methods. Then the build pipeline compiles and pack the encoded binary in base64 (to avoid static external files).

At runtime the usage is straightforward in selecting wasm or js backend.

[Andrea Silvi @ansi_code]

@xenovacom @huggingface I also think that JavaScript ecosystem needs a proper algebra/tensor library.

Here is another module I hacked weeks ago to achieve it leveraging on WebAsembly. (Still incomplete but with great potential)

https://t.co/GNpEpR581o

[Andrea Silvi @ansi_code]

@xenovacom @huggingface Cool, I worked on something similiar weeks ago (still PoC). It can run on Browser or Node.js and support by design IPFS pinned models.
Appreciating a star🔥

I think there is space to merge efforts.

https://t.co/69UKqWKkMU

[Andrea Silvi @ansi_code]

@mishig25 @xenovacom I also think that JavaScript ecosystem needs a proper algebra/tensor library.

Here is another module I hacked weeks ago to achieve it leveraging on WebAsembly. (Still incomplete but with great potential)

https://t.co/GNpEpR581o

[Andrea Silvi @ansi_code]

@mishig25 @xenovacom Sure, I worked on something similiar weeks ago (still PoC). It can run on Browser or Node.js and support by design IPFS pinned models.
Appreciating a star or shoutout🔥

https://t.co/69UKqWKkMU