Applied ML ▪ Senior SWE (AI) - R&D @SiemensHealth ▪ Tech x Finance x Chess x Books x Writing
Joined May 2019
- Tweets 1,627
- Following 449
- Followers 213
- Likes 2,829
205 Photos and videos
Pinned Tweet
9 Nov 2025
A powerful line I've read in a while:
Your ambition doesn't (and shouldn't) fit into someone else's structure.
133
"Why speculate having trillion dollars when you can have it for real?"
6
Day 69 of learning why liquidity is more important. Sit on cash!
25
No matter what numbers you achieve, there’s always a higher number.
Fortunately, happiness and contentment is still immeasurable.
"I Can't Sleep": Gurgaon Man With BMW, Rs 40 Lakh Salary Still Thinks He's 'Poor', Doctor Explains ndtv.com/offbeat/i-cant-slee…
126
Uh what speed up now? 👀
52
Mrutyunjay Biswal retweeted
May 25
Today, we announce @NeocambrianAI
The future of AI will be physical. But Physical AI has no internet scale dataset to learn from.
We’re building the data foundation of Physical AI - a high fidelity, pre training scale database of Human Action.
From India. For the world.
61
53
579
78,635
Alright, imma shoot a shot!
Hi @reach_vb @gabrielchua
Me and my homies got frustrated with stock research. Hundreds of deep research threads across ChatGPT, Grok, Gemini, Claude > thousand discussions to build a thesis, a case around a stock or an industry > following sub-reddits, X accounts for near-live news, banging heads onto walls, etc. etc.
Finally, a month back, we built ArthKosh. A financial data intelligence platform with multi-agent systems and harness to offer hyper-personalised, proof-backed research chat interface for retailer investors.
Many don't understand whats' a spread, why PEG of a stock matters and where, how event A affects company Z, and the list goes on.
ArthKosh integrates heterogeneous data sources under one-umbrella and unleashes a multi-agent system on 'em to provide relevant information in an ocean of overwhelming always-updating event scope.
Now, we built it at the Codex Bengaluru hackathon, roughly a month back. Scope was too big, couldn't get it to a desirable state. Entered @opencode hackathon the same weekend, finished on top-5 under multi-agent theme. Been token-maxxing ever since on the generous ChatGPT Pro subscription, until now.
So, would you kindly extend the earlier generosity with continued Codex Pro access and some API credits for me to finish the MVP with improved harness and core functionalities? I am happy to share the progress so far on dm :)
127
Anthropic on generational-talent-maxxing arc!
May 19
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
65
“work on things that matter and not think of rewards.”
May 16
We did research when pay was low.
We did research when pay was uncertain.
We did research even when we were lucky enough to be paid well.
One way is to figure out what to work on is to work on things that matter and not think of rewards.
We are still quite early into what makes a frontier model all the way from optimization, architecture and objectives. Big token wants to convince you otherwise.
71
Just letting @sama know that @ChatGPTapp latently been approving me of being “directionally” right.
23
Mrutyunjay Biswal retweeted
May 12
🧵1/ A single neuron is sufficient to bypass safety alignment in LLMs.
Across 7 models, 2 families, and scales from 1.7B to 70B, suppressing one MLP neuron bypasses refusal behavior — with no fine-tuning and no prompt engineering.
We call them refusal neurons.
We also study concept neurons: neurons that encode harmful knowledge itself. As a proof of concept, we identify suicide-related neurons. Our analysis reveals several interesting results⬇️
Joint work with @AtoosaChegini (equal contribution) , Maria Safi
15
80
461
30,742
Mrutyunjay Biswal retweeted
May 12
We just hit #1 trending on @huggingface Spaces 🎉
“The Ultimate Guide to RL Environments” dives into building & scaling RL environments for LLMs.
If you're exploring RL agents, this might be useful
May 5
Excited to release the Ultimate guide to RL environments!
Definitions of RL environments differ wildly in the LLM era, so we spent the last month building several RL environments across 6 different frameworks, domains and complexities to map out which are easiest to build with and which can be scaled to 1000s.
8
16
202
18,548
Mrutyunjay Biswal retweeted
The Real Singularity is the Friends We Made Along the Way geohot.github.io//blog/jekyl…
7
12
189
8,665
Every once in a while, life presents you with split roads. More often to test your conviction. You’ve set out for one goal, life would present an equally rewarding opportunity or opportunities. And you’re forced to take a decision: your conviction vs near term gratification.
This is a vicious cycle, and there’s no better test of character than this.
44
67