Ashvanth.S

@ashvanths.bsky.social

Deep Learning Practitioner | Language Lead for Tamil @ HuggingFace | Interested in Continual Learning and Generative Models | Website : https://ash-01xor.github.io/ X : https://twitter.com/ashvanth_s1

"Code appears. It looks right. The diff gets approved. But no one held it in their head. No one walked through the house; they just glanced at a photograph and called it home." Wrote about coding, Sisyphus, and building where the line keeps moving. link: asherr.bearblog.dev/in-pursuit-of/

In pursuit of

In the span of ten seconds, a fleeting moment, lives can be transformed forever. 100 Meters (Hyakuemu), produced by Rock 'n' Roll Mountain, is a film based u...

asherr.bearblog.dev

Have you ever felt like you lost your focus while reading a book and wandered into deep internet rabbit holes? Introducing sollu : AI-powered dictionary. Uses the Gemini model under the hood. It is open-sourced as well :).

BildBild

Quite a humbling experience every day while coding. You start with an issue and a vision about how to solve the problem and then pretty much the road traveled often to reach the solution isn't straightforward. Humbled each and every day to understand and accept and that it is how it is.

Over a period of time , getting to realize that im having my flow states during certain periods of time and getting to schedule tasks around it. Guess the goal is to build systems that can make sure we enter such states like on and off button.

Not able to point of the difference particularly , but gpt-4o-mini seems to work way too fast over the last day. From taking around 4 to 5 mins to process a 65-page PDF for extraction, it takes around 3 mins. Do you guys want me to run benchmark tests and probably write a blog post about it ?

Building SmolGPT myself , have plans to extend it. but before that struggling with managing python versions !!! Had to use pyenv and then pip. like now i get why experienced devs are frustrated with python package management

Interesting to see the hype of agents and using them , but almost everyone who uses the term throws it away just like that. All I get to see is a clearly well-defined workflow in a constrained environments most of the time and yet they are being called 'agents'.

Since being used to python development from the start i dont think i never had an issue using pyenv , venv , conda etc. Like it never felt like a chore. But then hearing about devs from other communities really does make me question why .

Somehow deep down i always get to think about how optimization of any process leads to boredom over a period of time. The excitement and the risks once taken might decreases due to the numbers the clouds our judgement. Like while recruiting , where folks are given standard questions to solve or ..

Its been great few weeks reading about agents after going through the course conducted by Berkeley. While there was lots of insightful talks , one that was particularly insightful was on week 4 , where Burak Gokturk got to talk about enterprise trends about agents. The most important trend...

Introducing Maya – A New Multimodal Multilingual Vision-Language Model. Maya is a completely open source, open weight, and open dataset, designed to handle 8 languages, cultural diversity, and nuanced real-world contexts in vision-language models. Paper: arxiv.org/abs/2412.07112

Maya: An Instruction Finetuned Multilingual Multimodal Model

The rapid development of large Vision-Language Models (VLMs) has led to impressive results on academic benchmarks, primarily in widely spoken languages. However, significant gaps remain in the ability...

arxiv.org

Weekend was quite cool , got to just keep my head down and work on retrieval engines over a custom database. Implemented thing right from TF-IDF to RAG and saw quite some interesting things. Anyone who says TF-IDF is not the best must need to implement it and check it first...