Chowdhury

Today, I’m introducing PlusB plusb.in The most efficient way to work with AI. PlusB is an Agentic AI Suite, powered by 7 experts, each equipped with the best SOTA models and a suite of curated tools. All experts in PlusB share the same brain, so each performs tasks personalized to you. Every expert knows you, your likes, and dislikes. You can share context with any expert anytime. All your messages remain fully encrypted and yours. You control everything. Here’s a quick look at a shared workflow.

Everyone's talking about AI. Very few are actually building it from scratch. I tried. Not with funding. Not with a team. Not with grants. With my own money. I cut corners where I had to. Reduced data. Simplified architecture. Optimized everything just to survive the cost, and still managed to build a working model. Not a wrapper. Not a UI. Not a "ChatGPT but with a different color." A real model. But one thing that is keeping me away is - funding. I applied. Again and again. Silence. No feedback. No direction. Just silence. Meanwhile: • Fear-based AI content gets millions of views • People with surface-level knowledge lead the conversation And the ones trying to build something real… are just trying to stay afloat. We have talent. We have engineers. We have ambition. But somewhere, we stopped backing the people who actually want to build. I don't want India to just use AI. I want us to create it. To compete. To lead. To build models that millions rely on every single day. Because I know we can. I've seen it. I've felt it. I'm living it. I'm still going to build. With or without support. But imagine how many builders like me gave up quietly… because the fight outside was harder than the one inside. If you feel the same fire - to research, to build, to create something real… Or if you believe in backing something from zero You’re welcome.

Today My product PlusB is launching in @producthunt after @ycombinator rejection :) Wish me luck! producthunt.com/products/plusb…

I am currently working on building the Treevik model family. As a result, I have built a 2.5B model: Treevik 2 From Scratch (not finetuned from other models- pretrained with a custom architecture, then SFT). It still does not perform well on factuality, but it serves the goal of enabling models to run on low-end hardware, and that is the research direction. treevik.plusb.in

I am currently working on building the Treevik model family. As a result, I have built a 2.5B model: Treevik 2 From Scratch (not finetuned from other models- pretrained with a custom architecture, then SFT). It still does not perform well on factuality, but it serves the goal of enabling models to run on low-end hardware, and that is the research direction. treevik.plusb.in

1/ Ever wondered how LLMs "understand language" ? It's because of embeddings. A thread 🧵👇

6/ Generation looks impressive. Embeddings do the real work. I have written a detailed article on embeddings. Read here medium.com/towards-artifi…

5/ Embeddings don't "understand" language. They encode statistical structure of meaning. But that structure is powerful enough to feel like understanding.

Thread 🧵 1/ Hi guys. I think nobody remembers me. That’s okay. Life happens.

2/ Some time back, I was building something close to my heart aiexchange.tech Dreams were big. Nights were longer.

GPT 5 >> Any other Model in the world.

There is no good Indian AI Product. Why?

I Failed in building a LLM from scratch! It cost me $400, 15 days of my life, and happiness. - I started collecting high quality data and ended up with a dataset with 200M tokens (approx.) - My model Parameters was around 2.6B - 85 GPU hours on a 80GB H100. - The early signs were good. But, after training, after 20-25 tokens it is hallucinating like hell. - Finding out where it went wrong. Maybe I'll start again with a SLM.

What is the Use case of @perplexity_ai ? It's useless for me.