Sourav Suman

Blockchain Developer | Solidity | ZK

[Ishita @Ishita_30]

A lil glimpse of what we’ve been building in private credit.

> PRIVATE_CREDIT
> status: ONCHAIN
> deployed: DEFI
> liquidity: EXIT.... wait (・・;)ゞ

Tried to include as many as possible after a lot of research. Missed yours? Drop it below.

[Pranshu Rastogi is at @NS @pranshurastogii]

Somewhere between forgotten promises, missed birthdays, unpaid bills, screenshots you'll never open again, and "I'll do it later"...

we slowly started carrying too much in our heads.

We built TuckBack because your mind deserves to think, dream, create, and live, not act as unlimited storage.

Just tuck it away.
We'll bring it back when it matters.
Private by design.
Built for real life.

Sign up to get early access. 💚
https://t.co/Q2dTpmCEod

[Sourav Suman @smn_srv]

Fun fact: GPT-3's tokenizer has ~50,257 tokens (using byte-pair encoding). The famous "Attention Is All You Need" Transformer paper originally used ~37,000 tokens. The choice matters—fewer tokens mean faster processing; more tokens, better precision! #AI #Transformers #llm

[Sourav Suman @smn_srv]

We often hear Transformers called “next-token predictors.” But what exactly is a "token"? Are tokens just words, or is there more to it? Let's explore this 🧵👇 #Transformers #NLP #llm #genai

[Sourav Suman @smn_srv]

Are tokens and words the same? Not always!
Example: "My name is sourav" has 4 words, but GPT-3 tokenizes it into: ['My', 'name', 'is', 's', 'our', 'av'] — that's 6 tokens! 🚀
More tokens ≠ more meaning, but helps models handle language flexibly. #GPT3 #Tokens

[Sourav Suman @smn_srv]

How is the number of layers decided?
It's typically a balance between:
• Performance: More layers can handle more complex patterns.
• Computational cost: Too many layers can lead to diminishing returns and slower training/inference.
Example:
- GPT-2: 12 layers(decoder layers)

[Sourav Suman @smn_srv]

Did you know the transformer architecture diagram you often see isn't exactly the full picture? 🤔
Let's explore a key detail you might've missed! 🧵👇

[Sourav Suman @smn_srv]

But why multiple layers? 🤷‍♂️
Each layer refines the representation, capturing increasingly complex relationships. Stacking layers allows Transformers to better understand context, dependencies, and nuances in language.

[Sourav Suman @smn_srv]

Interpretation: DL models are like black boxes and we don't exactly know what's happening so we cannot say why the model gave this result. But in ML we exactly know what model gave this result.

[Sourav Suman @smn_srv]

Day 2 of learning Deep learning

Deep learning vs Machine Learning
1. Data Dependency: Deep learning(DL) models are data hungry and require more data for training , whereas the performance of ML model stagnates after providing data beyond a limit.

#understanding_deeplearning

[Sourav Suman @smn_srv]

Training time : DL models training time is high as compared to ML models

[Sourav Suman @smn_srv]

Supervised learning models define a mapping from input data to output predictions.

[Sourav Suman @smn_srv]

Day 1 of learning Neural Networks

Artificial Intelligence is basically building a system which simulates intelligent behavior.

Machine Learning is a subset of AI which learns to make decisions by applying a mathematical model on observed data.

#understanding_deeplearning

[Rabble @0xRabble]

Last Saturday's MAHa-thon was a blast!

The @okto_web3 office hosted code monkeys hacking away, including a vacationing Burmese dude who won. Corporate peeps also got their hands dirty, snagging cash from the $1000 prize pool.

Kudos to @vijaytechmm @oirohit @PritamK94689631 @smn_srv @0xyashgarg for winning Saturday.

More Mini-App hackathon madness coming soon, so bone up on @telegram Mini-Apps unless you enjoy looking clueless.

What's a Telegram Mini-App? Check 🧵.

[Sourav Suman @smn_srv]

https://t.co/3gysGCbaTp

[Sourav Suman @smn_srv]

An open-source Attestation Protocol that allows contributors to create templates and organizations to use those templates for document attestation. This protocol utilizes zero-knowledge proofs on top of @MinaProtocol . Check out the concept we submitted to the #zkIgnite cohort 2.