i guess fine-tuning is fun!🤖
Delhi, India
Joined September 2021
- Tweets1.7K
- Following806
- Followers98
- Likes112K
Imhimanshu retweeted
India needs a ministry of compute.
Like ministry of oil, we need laser-focus from the top to build compute necessary for participating in the future economy.
Otherwise, we will repeat in a decade what we’re doing now: spending expensive forex for importing what we don’t have.
Imhimanshu retweeted
india keeps rewarding mugging up PCM and then wonders where the hackers went.
JEE is not computer science. GATE is not cybersecurity. rank is not competence.
our examination fetish is killing technical talent:
ni5arga.github.io/notes/indi…
Imhimanshu retweeted
Alignment is not as hard to solve as many claim, but it is in the end an algorithmic problem which a lot of ML community mostly stopped working on.
Formulation of alignment stated by three (or four) laws of robotics can take us very far, so we roughly know the objective.
The tricky part is, how do we take gradient with respect to alignment? We have two algorithms right now at our disposal: pretraining and RL.
Pretraining takes gradient with respect to next token prediction - that's def not the alignment objective.
We could create RL environments that embody the alignment objective, but:
- those are expensive to create, so often cheaper, hackable proxies are used in practice
- RL as an objective needs successful and unsuccessful rollouts to happen to take the gradient step. We DO NOT want to harm any humans in the process of aligning our modes - this is a pretty big problem.
Therefore there are two solutions forward for the alignment problem:
- either we RL models in a simulated environments with simulated alignments and decreasing the likelihood of harming simulated humans, which will never be perfect
- or we create a new algorithm that can teach our models to not harm humans without harming any humans in the process
Science is the process how we solve the hardest problems ahead and that is one of them
Imhimanshu retweeted
Computer science is gradually returning to the domain of physicists, mathematicians, and electrical engineers as large language models automate much of what we currently call software engineering.
The field’s center of gravity is shifting away from manual code writing and toward deeper theoretical thinking, mathematical insight, and systems-level reasoning.
Imhimanshu retweeted
Six faculty members from IIT Madras have been selected for the Anusandhan National Research Foundation (ANRF) J.C. Bose Grant, one of India’s highest honours for senior scientists and engineers. The grant provides ₹25 lakh annually for five years to support frontier research.
Their work spans prosthetics, clean water, next-generation optical networks, sustainable energy, biomedical innovation and fundamental chemistry, demonstrating how cutting-edge research at IIT Madras is translating into technologies that improve millions of lives.
The six professors and their impactful research are:
Prof. Sujatha Srinivasan develops affordable prosthetics and assistive devices tailored for Indian users.
Prof. Thalappil Pradeep creates low-cost water purification technologies that have brought safe drinking water to millions and founded the International Centre for Clean Water.
Prof. Deepa Venkitesh leads India’s first multicore optical fibre testbed and advances high-capacity optical communication systems.
Prof. Jitendra S. Sangwai works on carbon capture, gas hydrates and sustainable energy solutions for reducing emissions.
Prof. Ramakrishnan Swaminathan designs biomedical devices and instrumentation to improve diagnosis of brain and muscle-related conditions.
Prof. Sundargopal Ghosh advances transition-metal boron chemistry for better catalysis and small-molecule activation.
We need to celebrate these real-life heroes.
News Source: Times of India
Imhimanshu retweeted
It's so inspiring to see Gen-Z take a stand for their future in India, and stand united across different parts of the country.
After all, with their entire lives ahead of them, they're the ones with most to lose.
Everyone knows India's education system is a big mess. We're not adequately prepared for the future. We're not investing in basic R&D.
We're the backoffice of the world, but we shouldn't settle for that, shoudl we?
Statistically, the largest population in the world ought to have the largest number of geniuses. I think they're still there, but our education system has failed to harness their potential.
India's largest pool of youth was once hailed as an asset, but is it now?
I feel there's still hope our country can deploy a billion brains to craft a better future for entire humanity. But that won't happen unless the youth feels empowered, and unless the country invests in highest quality education at all levels - primary, secondary and higher.
So more power to the young generation!
Imhimanshu retweeted
Why are Indian companies not developing frontier models?
The answer is in economics, not talent. The biggest hurdle for an Indian company trying to develop frontier model is that right from the get go, it has to compete with Anthropic and OpenAI free plans + all Chinese open weight models.
Chinese models mostly flourished due to the demand from their domestic market which OpenAI/Anthropic couldn’t meet as they’re unavailable in China. This unmet demand created space for innovation.
China knows this story well, but here in India, we like unfettered, globalized free market, and that requires an upstart to compete with established companies from across the world.
If India is serious about the AI race, it needs to create a massive guaranteed market for domestic AI or else it’ll be economically irrational for any funder to underwrite a product in a highly competitive, already established market.
Imhimanshu retweeted
Getting used to being liked likely means you are overfit to RLHF.
The problem with overfitting is that the pain overwhelms the limbic system once you try to sample trajectories outside the known distribution.
As more people like you, your sampling regime becomes smaller and smaller to avoid negative feedback. Eventually you get stuck and become a slave to your own feelings.
That’s why I have never seen a model student happy once they become a “model”. Their weights are frozen and cannot be updated anymore. They cannot risk being better than their own SOTA.
Imhimanshu retweeted
its high time that india starts to embrace the need of "working together" and start building towards the sovereign ecosystem. there will be a shortage of gpu in future, and you will be deprieved of large model's usage in that case, what do you do then?
the only feasible solution is to build a self-sustaining ecosystem for ai, where different start-ups work in co-operation for mutual growth (this is the need of the hour). for instance, i would like to cite an example of how @PrimeIntellect, @arcee_ai and @datologyai works in sync while achieving their own set of goals. datology builds and works on coming up with excellent synthetic data pipelines and flows, prime invests heavily around the ai-infra/post-training research and development, arcee works on frontier models from scratch. they're improving steadily, they're not wanting to go more than others in valuation but accelerating research and making it accessible.
we need to have the same ecosystem in india as well. you cannot expect @sarvamai to carry all the heavy-weight and deliver a model which out-performs opus/gpt, its an extremely impracticle demand. instead, i ask the vcs to invest more in companies which can accelarate other domains of research (small models, ai infra, hardware accelaration, data curation (india is a data mine note that), post-training etc). in this way, we're investing in the ecosystem not the immediate results and investing in a process will eventually result in long term benefits, its bound to happen.
so instead of being dependent on 1-2 big labs and work according to their operation strategies, why don't we have something for ourselves which would help us even in times of adversities? research has to be democratized not concentrated in the hands of few.
think.
This is big: all access to Mythos and Fable AI models disabled for everyone outside America.
First thoughts:
1. Technology is the ultimate weapon. National sovereignty, national security, all of it is now about technology.
2. Globalization is dead and Bharat must find her own way ahead.
We must keep these two ideas in mind.
What can our government do right now? Ensure that orgs in India embrace smaller models, both Indian and Chinese open source ones. With a bit of effort, we can make them work. Anyway, why pay money to people who don't even want to sell to you?
We must deepen our R&D. Sarvam has been on it and we have been on it but remember that the latest models cost not only huge GPU budgets to train, the GPUs themselves are restricted. So we can't afford the scale of money (of the order of $100+ billion to even get in the game!) and even if we could come up with the money, we can't get all the GPUs. I would not like to ask the government to fund tens of billions of dollars on this anyway - the money has far better uses.
Zoho has been pursuing alternative R&D approaches that are far, far less expensive but by its nature cutting edge R&D takes time and we are patient. I am confident we will get there.
Any remaining people in India who have delusions about globalization should wake up now.
Imhimanshu retweeted
The Redpoint InfraRed 100 is now live.
These are the companies building the infrastructure that powers everything happening in AI right now, from world models and agent runtimes to the sandboxes, databases, and security tools agents depend on.
Congratulations to this year's honorees!
Read the full 2026 InfraRed Report: our state of the union on AI and cloud infrastructure 👉 redpoint.com/reports/the-inf…
Imhimanshu retweeted
The bitter lesson in 26 words:
Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.
Imhimanshu retweeted
This is the best slides I've seen so far about claude code pro tips (esp. for AI researchers): xingjianz.com/assets/talks/a…
By @_Jimmy_Zhang_
Imhimanshu retweeted
Excited to launch the accompanying free RLHF Course for my book. To kick it off, I've released:
- Welcome video
- Lecture 1: Overview of RLHF & Post-training
- Lecture 2: IFT, Reward Models, Rejection Sampling
- Lecture 3: RL Math
- Lecture 4: RL Implementation
I'm going to add question & answer videos throughout the lecture to go deeper on topics that need it, and potentially cover some topics that are too recent and in flux to go in print. I expect 10-15 videos in total over the next few months.
At the same time, development around the code for the book is picking up. It's a great time to build the foundation for post-training methods.
YT playlist and course landing page below.
