Exploring ML/AI and local AI research. Author of Under the Hood (Building LLMs) - love தமிழ் and building தமிழ் models!
Austin, TX
Joined November 2020
- Tweets1K
- Following1.1K
- Followers3.5K
- Likes785
Pinned Tweet
This one is personal. Of everything on the launch list, getting this book into India at an Indian price is the one I cared about most, and for a while I genuinely wasn't sure it would be possible. Notion Press prints both editions locally, and that's the only reason it's ₹1,599 for the paperback and ₹1,899 for the hardback instead of a $49.99 / $79.99 import. And I am sorry to all those who kept asking, and I had no answer for.
Both links below cost you exactly the same - the direct one just helps me the most. And to be honest,the best part of this launch hasn't been the rankings. It has been the ability to finally have my family order one for themselves. So happy this is finally in reach for folks back home. Thank you again for all the love!♥️
Thanks to @sriiniivas for the lovely unwrapping video!
🇮🇳 Paperback + Hardback, direct: direct.notionpress.com/in/re…
🇮🇳 On Amazon.in: amazon.in/dp/B0HF122KXS?&tag…
🇮🇳 India Kindle: amazon.in/dp/B0H9FYVK3W
US Kindle: amazon.com/dp/B0H9FYVK3W
US Paperback: amazon.com/dp/B0H9FWPCDV
Leanpub: leanpub.com/under-the-hood
He is absolutely right. Also this has been around for a while now, with people like @jobyaviation doing it the right way. I mean, if I you want to do circus tricks in a controlled space, sure, wait another 10 years for it be rolled out. But 2028 is a lie.
Ramchand Kumaresan retweeted
Today I'm launching Chitra Studio 🌹
Your AI Media Workstation for the Mac.
Images, Video, Voices, Music and Sound Effects from 60+ models, in one app.
Know the cost before you generate.
Every prompt, cost and edit documented.
Your library stays on your Mac.
$49/year, 50 credits included.
14-day Money Back Guarantee.
chitra.studio/
Quick update: I found an issue with the HF Spaces where the preview is hosted. This would have prevented 127 of the 167 tested prompts from giving bad answers. If you can - please do test again. Links are in the OP here - xitter.cf/Mechramc/status/211037…
Agaram 2.5B, @MuraiLabsHQ's first Tamil–English model trained from scratch, is open for a 4-day public preview 🎉
Chat in Tamil, Tanglish or English, try the tools, and rate replies 👍/👎.
🔗 huggingface.co/spaces/mechra…
📄 arxiv.org/abs/2610.18420
Weights come after the preview - waiting on licensing to be cleared.
So far 405 prompts tested by real users across 169 sessions. Please use the 👍or 👎folks so I know which ones worked for you and which ones did not. So far I have 8 👍and 17👎- this shows the model needs a bit more finetuning and thats valuable info. More testers please! Thank you!
xitter.cf/Mechramc/status/211037…
Agaram 2.5B, @MuraiLabsHQ's first Tamil–English model trained from scratch, is open for a 4-day public preview 🎉
Chat in Tamil, Tanglish or English, try the tools, and rate replies 👍/👎.
🔗 huggingface.co/spaces/mechra…
📄 arxiv.org/abs/2610.18420
Weights come after the preview - waiting on licensing to be cleared.
Agaram 2.5B, @MuraiLabsHQ's first Tamil–English model trained from scratch, is open for a 4-day public preview 🎉
Chat in Tamil, Tanglish or English, try the tools, and rate replies 👍/👎.
🔗 huggingface.co/spaces/mechra…
📄 arxiv.org/abs/2610.18420
Weights come after the preview - waiting on licensing to be cleared.
What helps us most: ask it real things you'd ask a friend, in Tamil, Tanglish or English. Try the 🔧 tools: weather, currency, maths, Wikipedia. Then tap 👍 or 👎. Every rating shows us where it's weak and we can refine it better based on what is not working.
It's a research preview: it gets facts wrong, so check before you rely on it. With your consent, we keep the conversations to evaluate and train the next version. Please don't type personal details. Open until Sunday 18 Oct, 19:30 IST. The paper covers how we built it: a tokenizer made for Tamil, 50B tokens of pretraining, fine-tuning, and what works and what doesn't. Model weights and code come after this preview. அகரம் — the first letter.
Yeah @elonmusk we would like to know the answer to this. Also, it is also odd that you haven’t had an opinion on the h1B topic in a while. Can you share your view on it?
When you say that you will make pricing affordable for indians does that entail increasing pricing on American Starlink users to fund the price reduction?
Surely if it was possible to offer Starlink at these reduced prices you would prioritize American Starlink users first right? Why should the poor family in rural America pay more for Starlink than the rural poor family in india?
I say all of this as a major supporter of Starlink and one of the earlier adopters.
That ML engineer at xAi searching chat transcripts for Elon’s card info…
To buy stuff, you can just take a picture of your credit card, drop it in the chat and Grok @Bot will scour the Internet for the best deal on the item you want and order it
Readers added context they thought people might want to know
xAI explicitly warns users not to include personal information in prompts, yet credit card photos may contain such information.
Grok Bot's terms state that user data and credentials may be stored in its cloud environment and require human review for financial transactions.
x.ai/legal/privacy-…
x.ai/legal/grok-bot…
Hahaha 😂
Key disadvantages of allowing Starlink in India include:
National security risks: Foreign-controlled network limits India's ability to shut off service in sensitive border areas or ensure availability during crises. Past reports of terminals with militants or smugglers raise misuse concerns.
Data sovereignty issues: Laser links could route traffic outside Indian gateways, complicating monitoring and lawful interception.
High costs: Hardware and monthly fees likely far exceed affordable terrestrial options like Jio or Airtel, limiting rural access.
Market impact: Could disrupt local telecoms and foster over-reliance on a single US provider instead of domestic alternatives.
Agaram 2.5B update 🧵 Yes, it went from 1.5B to 2.5B - part of the reason I haven’t made updates was because when 1.5b version was done, it was not good enough. It could not stop and was rambling, and was looping responses. After that, I went into a full on rework without any idea if it will work. Studied a lot more papers than I want to admit, and revised the corpus and added twice the amount of synthetic data.
At one point, I was close to stopping it and releasing a model with a write-up of where I failed. However, I guess I am just too stubborn- so blew another big chunk of money and time and decided to do it all again the right way. So after another two months - it is almost done.
A first-of-its-size, Tamil-first language model trained from scratch at Murai Labs: ~50B tokens, its own tokenizer, Tamil + Tanglish chat, tool calling.
Pretraining ✅ Chat fine-tuning ✅
Training a final round now on identity, tool use and safety.
Next week: results, and open weights if they hold up. அகரம் வருகிறது 🌱
The read is clear - investors world wide know what this means for IT jobs and how the companies will adapt.
They tried to do a hit job on India's stock markets yesterday.
They spread fake news and panic via X by very prominent US handles, saying the "Indian market crashed" showing a heatmap of all Indian stocks in red.
The Indian stock markets were not even open then. Today the Indian indices showed a big Indian finger to these small losers. Ruh roh.
The future is here!
After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.
I am speechless - i need to test this out but if this works, and if we can extend this on larger scales- it could mean a new paradigm shift. Congrats!
Can an LM, starting from random init (!!), learn to generate all of its pretraining data?
Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities.
A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.
20% almost, of the known unsolved math problems - solving one could have been a Ph.d dissertation or master’s thesis. They solved 90. And people who play identity politics and mislead folks into fighting each other, don’t have any clue the earth is going to be wearing a new avatar in another six months. AI’s message seems to be - catch me if you can, or wither into uselessness.
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Ramchand Kumaresan retweeted
Introducing Genex
The desktop app to build games with AI
→ Open source, MIT
→ Local models, or Claude Code / ChatGPT sub
→ Self-improving harness for game dev
→ Three.js + Blender, Meshy & other tools
→ Unity & Unreal plugins soon
genex.games/desktop
Ramchand Kumaresan retweeted
Kudos to the team for the great work. The rest should also be enabled modulo NV custom bits.
BOOM!!!! 🔥🔥🔥
$AMD has cleared a serious software milestone!!!!
Their GPUs now have >90% parity on vLLM’s critical “gating” tests, the automated checks that must pass before code enters vLLM’s official codebase.
That makes AMD support less likely to break with updates, improving production confidence. It’s software reliability progress, not a performance benchmark, but it reduces a real part of NVIDIA’s CUDA ecosystem advantage.
@AnushElangovan and his team need to get proper kudos!!!!!! 🫡🫡🫡
#AMD #AI
Ramchand Kumaresan retweeted
BTW @AMDRyzen AI Halo with 128GB is still available for $3999: a.co/d/0ir0BJ7p
