Sushant Kumar

Director of AI, SproutsAI
Prev: Head of AI, SquareYards, Co-founder Azuro, Bank of America, IIT Bombay 2014

[Sushant Kumar @sushantkumar_23]

Mind-blowing AI-generated tweets

With @OpenAI's GPT-3 model (thanks to @gdb), I built an app that generates its own tweet given any word.

Also, these AI tweets are indistinguishable from human tweets.

Try for yourself (replace 'word' below with yours):
https://t.co/LZN7RkTbE1

[Sushant Kumar @sushantkumar_23]

Feature request for Claude Code: would love a user-scoped memory layer, similar to Codex.

Project/worktree memory is useful, but a durable cross-repo personal memory for preferences, coding style, tooling habits, and workflow defaults would be a huge unlock. @trq212 @bcherny

[Sushant Kumar @sushantkumar_23]

I'm a simple person; I see a Rich Sutton lecture, I drop everything and watch!

[Sushant Kumar @sushantkumar_23]

Exactly what I was searching the internet for today! Thanks @rasbt for always being on point!

[Sebastian Raschka @rasbt]

Couldn't resist.
Here's a pure PyTorch from-scratch re-implementation of Gemma 3 270M in a Jupyter Notebook (uses about 1.49 GB RAM): https://t.co/M2f8EB0KBE

[Sushant Kumar @sushantkumar_23]

@thetronjohnson Great idea and execution!

Is this built in Swift or Electron?

[Sushant Kumar @sushantkumar_23]

For me, GPT-OSS was the much-awaited OpenAI architecture reveal since their detailed GPT-2 architecture!

Some key observations:

✅ MoE GPT-2-style Transformer (36 / 24 layers)
✅ 128 / 32 experts with top-4 routing ⇒ only 5.1 B / 3.6 B active params
✅ RMSNorm
✅ Grouped Query Attention + RoPE attn (biases still present in QKV matrices)
✅ 131 K context via YaRN (Sliding window)
✅ 4-bit MXFP4 packs 120 B on 80 GB & 20 B on 16 GB
✅ SwiGLU activation

[Sushant Kumar @sushantkumar_23]

Looking forward to this … 🚀

[Lee Robinson @leerob]

I'm joining Cursor to teach the future of coding!

There are millions of developers learning how to use AI and they need pragmatic advice:

1. We need to teach new developers strong foundations, so they know what to learn, and how to solve issues when debugging.

2. We need to teach experienced developers how AI can automate the tedious parts of coding, or save them time reading docs and fixing bugs.

3. We need to help developers become even more competent. AI may end up writing most of your code, but you have to review, understand, and maintain that software.

This is why some experienced devs are having a great time with AI. They can ask for a pattern like "add an exponential backoff" instead of “make it more robust to errors” which may or may not work.

I want to help developers become an order of magnitude more productive, and help more people contribute to building software.

This is going to take a *lot* of education and retraining. So expect more videos soon, and if you have ideas for what I should teach, let me know!

[Sushant Kumar @sushantkumar_23]

This lecture is a masterpiece from @RichardSSutton that's moving the conversation forward on what the future with AI could look like.

You can literally feel the clarity that comes when you have spent half-a-century thinking deeply about intelligence.

https://t.co/BT1v8LAZmm

[Richard Sutton @RichardSSutton]

https://t.co/dW1mBPdS2f

Machines that learn from experience were explored by Alan Turing almost eighty years ago, which makes it particularly gratifying and humbling to receive an award in his name for reviving this essential but still nascent idea.

[Sushant Kumar @sushantkumar_23]

@shl Taskaid AI

Link: https://t.co/ImN63Sa8xn

[Sushant Kumar @sushantkumar_23]

Next week will be 🔥

[DeepSeek @deepseek_ai]

🚀 Day 0: Warming up for #OpenSourceWeek!

We're a tiny team @deepseek_ai exploring AGI. Starting next week, we'll be open-sourcing 5 repos, sharing our small but sincere progress with full transparency.

These humble building blocks in our online service have been documented, deployed and battle-tested in production.

As part of the open-source community, we believe that every line shared becomes collective momentum that accelerates the journey.

Daily unlocks are coming soon. No ivory towers - just pure garage-energy and community-driven innovation.

[Sushant Kumar @sushantkumar_23]

@himanshustwts This is really good direction especially for those training small models for experimentation.

[Sushant Kumar @sushantkumar_23]

@_apoorvnandan Thanks for the post along with code.

Two issues that I faced and had to resolve:
1. env.reset() - https://t.co/1j9sA4JFUS

Fix:
- observation, info = env.reset()

2. env.step() - https://t.co/3Za26VAhhw

Fix:
- observation, reward, done, _, info = env.step(action)

[Sushant Kumar @sushantkumar_23]

One of best experiments to get hands dirty for building foundational models. ~30M model so cheap enough to pre-train.

Added RoPE to GPT2 architecture. Raised PR! Pre-trained to a validation loss 1.224 in less than 1 hour H100 (40GB)!

Great work, @alve_om! Toy MoE next?

[Om Alve @alve_om]

So I pre-trained an LLM from scratch this week, yes, pretrained not finetuned. This is how it went - A thread 🧵

[Sushant Kumar @sushantkumar_23]

@karpathy Spend millions and decades perfecting the art and science of 3D rendering, only to have general Transformers waltz in and say, ‘Neat. I’ll just watch a bunch of videos and wing it.’

"The Bitter Lesson" all over again!

[Sushant Kumar @sushantkumar_23]

Pre-training SLMs seems to be accessible way into AI research.

1 day of H100-80GB pretrains ~1B SLM on roughly about 15B gpt2 tokens if properly optimised (was able to get upto ~200k tok/sec throughput)

anyone know of SLMs (sub 1B) that has 50 MMLU (or in other words useful)?

[Sushant Kumar @sushantkumar_23]

Major AI labs are hitting the wall with larger models. True AI needs Continual Learning, Agency, and Adaptive behavior—not just bigger models and more data.

My thoughts that Path to Prize (SuperIntelligence) lies beyond the comfortable confines of LLMs

https://t.co/WkFdzpXf1t

[Sushant Kumar @sushantkumar_23]

@anoushk77 @vercel https://t.co/Y1GRn6kd3c

[Sushant Kumar @sushantkumar_23]

@SuperteamDAO @devfolio Thanks @SuperteamDAO for this amazing opportunity! Looking forward to the next 2 months!

[Sushant Kumar @sushantkumar_23]

GM,
After a month-long selection process, I’m super pumped to become a fellow at @SuperteamDAO and @Devfolio’s *Solana India Fellowship*.

#Solana is one of the fastest growing blockchains with faster transactions and higher throughputs for building on cutting-edge domains.

[Sushant Kumar @sushantkumar_23]

@Austin_Federa @SuperteamDAO @devfolio Thanks Austin! Overwhelmed by the love and support from the Solana community!

[Sushant Kumar @sushantkumar_23]

Over next 2 months of fellowship, I’m looking forward to sharing my experiences about exploring Rust in greater depths and building on #Solana for some exciting use-cases such as NFTs, creator-economy, DeFi, Gaming.

#WAGMI