This tweet is about how I have studied ML and made it my profession. I'll share the resources I've used and the sequence of my study. Straight to point ML pre-requisites(maths) : Linear Algebra, Probability Theory, Calculus, Optimization Theory(optional), Information theory(optional) Linear Algebra: Lecture course by Gilbert Strang Probability theory: MIT 6.041 (it contains parts of Bayesian inference as well) Calculus: your high school and college classes are enough Once basic maths is done then we move to ML. Classical ML : CS229. Either by Andrew NG or someone else. Follow their lecture notes and solve their problem sets. Reference books for classical ML that I followed: PRML by Christopher bishop, Pattern Classification by Duda, Hart and Stork After getting comfortable with classical ML we move to Deep Learning and everything else. Deep Learning and Computer Vision: CS231n. Very good lecture and assignments Reference book: Deep Learning by Ian Goodfellow. This is the best book on deep learning. I’ve read some chapters of it many many times. Beautiful maths and intuitions MLOps: dvc, WandB, MLFlow NLP: I just read hugging face blogs. I haven’t spent much time with classical NLP though. Alignment/AI safety/AI explainability: Anthropic Blogs(I’m a noob in this, just started learning couple months ago) Additionally: Blogs: Lilian Weng(OpenAI)’s blogs, colah’s blogs Additionally: arxiv. I read many papers from arxiv Karas and Tensorflow blogs: for introductory code about modern deep learning frameworks Competitions: Kaggle Cloud compute. GCP/collab/Kaggle notebooks PS: this is not a roadmap. Just what I followed till now and I find it quite structured. Even after 5 years I still find myself learning new stuff everyday.

If you're building LLMs on @Apple silicon, this is a gold mine

We are deadass sitting on a gold mine. Qwen 3.6 comes with MTP, so no external draft model is needed for multi token rpediction unlike normal speculative decoding.

Building a general purpose harness is impossible given every model being trained on different rewards What you can aim for is a continuously evolving meta system that adjusts your harness as you do tasks while observing where the model and tools lack

Last year I got a health checkup in Japan. To my surprise, the medicine I got from the Japanese pharmacy was made in India. Even after my health insurance, the medicine cost me 4 times as much as it would have cost me back in India if I just bought it from a pharmacy And the Japanese doctor told me how good and effective the medicines made in India are. India is a pharmacy giant.

The Crunchyroll translation of "Smoking Behind the Supermarket with You" is done by the famous Indian translator Sriram Gurunathan

Today we drop Trinity-Large-Thinking. SOTA on Tau2-Airline, frontier-class on Tau2-Telecom, and the #2 model on PinchBench, right behind Opus. On BCFLv4, we're in the mix with the best. 26 people with under $50M raised and a ruthless pursuit of greatness. What this team just pulled off is nothing short of incredible. One hell of an accomplishment and I couldn't be more proud of Arcee. And we've got more to prove.

Scratch the surface, and the free-market enthusiast often turns out to be a socialist who wants tax payers to fund his pet idea.

There are many ways in which India can win in AI Talent density. We have lots of talented people, we just need less disturbance from external people and more support for the focused work, productive environment and good pay Data. India generates huge amount of data. And US companies are taking full advantage of it. Whether it’s collecting cheap egocentric view for robotics or simple data labeling. There is definitely a cost arbitrage and we should take benefit of that as well. Data collection and labeling is a huge challenge and we can do it reliably and cheaply Compute. We still lack compute but recent efforts by sarvam and other IIT labs have shown that with good funding and support we can reach the level of compute we need to run big experiments Optimism. This is where we lack. From everyday opinions from people saying “it’s too late” or “it’s just post training nothing special like deepseek” to leaders saying we don’t need sovereign LLM(spoiler we do). We have to be optimistic on the work our people are doing. Policies should support that as well

Y’all not gonna believe how deeply Hinduism is rooted in one piece . In the latest chapters, Imu literally used “Om” en with Gyan Mudra, and the mural line, “God dances and leads the world to its end and brings a new dawn,” directly parallels Shiva ( Nataraja )

Post-training an open weight model is the most efficient way to build sovereign AI. We’ve seen it already with Cursor Composer. If your country is trying to build LLM from scratch, that’s wasting your tax money. LLM is all about dataset, and none of country has enough data as US/China.

80.2 on SWE bench verified??? I’ll give it a shot today

Indians are blaming Infy/Murthy/Nandan for a lack of home-grown LLM. But that blame is misplaced. Infy/Murthy/Nandan could never start an AI company. They had money and they could have fund something. But that again, is their money and why should they risk it? The real culprit here is Amitabh Kant and other IAS like him. These are the people, why I left India and started two companies in the US. These are the people, why much of Indian talent left to work for US based companies. And these folks did well. Let's say the government gives me $10B and ask me to set up an AI lab in India. Am I qualified to do it? YES. Will I do it? HELL NO And you would ask why? Some would say that I have a cozy life in the US. Some would say, I have deep connection in the US, including family. All of that is correct but does not pin point the reason why I wont start a company in India. The real reason is Babu. Unlike, many In India who think competing for 1000 seats using some bullshit essay writing contest makes Babu some wizard, I have not come across one, I will hire as an analyst. Under no circumstance, I am gonna report to a babu (Happy report to Dharmendra Pradhan or Smriti Irani though). Also, under no circumstances, I will accept a position where I am unable to fire and put an IAS in jail if they reported to me and indulged in some corruption. Till this babu problem is fixed, no NRI would come to India. If I were the CIA or CCP, trying to ensure that India does not gain AI independence, I would make every effort to protect Babu fiefdom.

Wait what? Rio 3.5 Open 397B, developed by IT company of Rio de Janeiro's city government is now SOTA open source and even outperforming Qwen 3.7? What is happening today. Never heard of them before.

SITUATION DETECTED: The city of Rio de Janerio has post-trained a model. Based on Qwen 7/2, Rio 3.5 Open 397B adds SwiReasoning on top of the base Qwen model — a framework that dynamically switches between standard chain-of-thought and latent-space reasoning, guided by entropy-based confidence signals, so the model only "thinks out loud" when it needs to and otherwise reasons silently in hidden space for better token efficiency.

Last year, I was presenting to an internal team we had put together to develop a SOTA geospatial foundation model and this was a slide from that ppt. We have the talent and ambition, but the only entities that have helped us throughout this journey is govt IndiaAI mission and NVIDIA, who gave us priority compute after seeing our initial results. This is the kind of support that should have come from Indian IT behemoths. A country that can develop nuclear weapons and land on the moon is very well capable of developing SOTA models. But for that, we must first dream to be more than AI coolies.

Govt has been incredibly supportive of our AI efforts through the IndiaAI Mission and other schemes. Zero support from any Indian industry. Thanks to NVIDIA, we trained India's first Geospatial Foundation Model (think of it like ChatGPT for satellite data) SLM using ISRO data. Dropping soon!

Don't settle for mediocrity. If you're going to do something, go all in. Focus intensely on a topic, field, or skill for a good stretch of time—whether that's six months or a year. Commit fully and don't stop until you're genuinely satisfied with what you've achieved. Don’t stop and don’t distract yourself with anything else. Push yourself to the point where you can confidently say, "This is the best I could have done." Then, look at the results. You'll be amazed to discover that what you lacked wasn't intelligence or resources, but persistence. This isn’t about DSA or CP or ML or development, arts, or sports. It applies to everything. At least once in your life, experience this phase. Once you do, you'll realize you can accomplish anything.

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs. Kimi 2.7 is amazingly good. More below.

yess!!