Pinned Tweet
Had a great time at the @nvidia AI Summit.
Got to meet the computing pioneer of the 21st century and my CEO, the one and only Jensen Huang!
Yash retweeted
We spent $2m so you don't have to.
And so you can learn how to train MoEs.
Mixture-of-Experts LLMs are what powers the AIs you use everyday, and few people in the world know how to train them at scale (a handful at each top AI lab).
This handbook distills everything we learned in meticulously crafted chapters and figures.
Disclaimer: it may take days to read.
We hope to see 1000x more people able to train their own frontier models all over the world! ๐ค
Congrats @Nouamanetazi @lvwerra @Thom_Wolf
Yash retweeted
Ramayana: The Legend Of Prince Rama (1995) 4K movie is now available on Internet Archive (Without watermark)
Link:
archive.org/details/@ramayanโฆ
Meet NVIDIA Nemotron 3.5 Transcribe ๐๏ธ
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyoneโs talking at once.
1/ ๐งต
Yash retweeted
Wouldnโt it be nice if there was at least one actually pure website? One without the usual gotchas.
No asking for your email. No parts paywalled. No ads. No data collection.
I think so, which is why I created teddyrookbook.com
Homeschool families get every Kโ12 course free (with teaching guides and practice that checks answers), plus a lesson planner that builds a week-by-week plan. Science labs. Languages. Music. Etiquette. Logic. Meal planning. Trades and farming. And much more!
Iโm building this for my family to use and sharing with others because Iโm open to feedback on how it can be better.
Yash retweeted
We just launched OpenUSD support on the @huggingface Hub ๐
Yesterday, @nvidiaomniverse put 2,500+ SimReady assets on the Hub. Forklifts, robot arms, engines, hospital equipment, all OpenUSD. You can also search the whole library in plain English
huggingface.co/spaces/nvidiaโฆ
huggingface.co/datasets/nvidโฆ
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-bโฆ
hsay.in
๐
Modi Govt deals another lethal blow against drugs with Operation Trident.
Our agencies launched a barrage of simultaneous crackdowns across 38 locations in Assam, Mizoram and Manipur on 23 October, smashing an international narco-cartel. The operation revealed โน100 crore worth of hawala transactions, froze 1,424 bank accounts, attached properties worth โน17.82 crore, seized 500 detonators and โน12 lakh in cash, and led to the arrest of 18 people.
The network trafficked narcotics across the borders of India, Myanmar and Bangladesh through the Death Triangle Corridor. The offensive attests to Indiaโs Himalayan resolve to realise the mission of eradicating every drug network by 31 December 2029. India will hammer until the last nail is driven into the coffin of drug cartels.
Heartiest congratulations to the anti-narcotics team comprising central agencies, and police forces of the three states.
From autonomous vehicles to smarter factories, companies are putting physical AI to work.
See what's new with NVIDIA Cosmos: nvda.ws/4dHB5x0
Explore the latest industrial AI announcements: nvda.ws/3VLNABB bit.ly/4su4ReH
Explore Microduckโs path from simulation to the real world with NVIDIA Isaac Lab in this Hugging Face blog. bit.ly/4iagHG3
Yash retweeted
โ๐๐๐ ๐ข๐ฆ๐ ๐๐ก๐๐ง๐ ๐,โ ๐ญ๐ก๐๐ฒ ๐ฌ๐๐ฒ.
Their mask is off. The sinister agenda is now out in the open. Unable to win the trust of 140 crore Indians during elections, the INDI Alliance and its radical cabal are openly challenging the very foundations of the Indian Republic.
Their own words expose their intent.
Rahul Gandhi brazenly declares, โOur fight is against the Indian State itself.โ Not against a political party. Not against a government policy. But against the Indian State itself and its constitutional machinery.
Prashant Bhushan openly calls for โregime changeโ and has spoken of revolution involving bloodshed, guns and the Army. This is not the language of democratic dissent or a peaceful transfer of power through elections. It raises a fundamental question: when they speak of regime change, are they talking about defeating a government at the ballot box or bringing the State to its knees through street violence?
Then comes Neha Bora, glorifying foreign turmoil with the slogan, โLal Salaam to Bangladesh for showing the way.โ This is an open romanticisation of the upheaval, violence and mob lynchings associated with Bangladeshโs political crisis, while holding up events beyond the constitutional process as a model to follow.
๐๐ก๐ ๐ฉ๐๐ญ๐ญ๐๐ซ๐ง ๐ข๐ฌ ๐ฎ๐ง๐ฆ๐ข๐ฌ๐ญ๐๐ค๐๐๐ฅ๐: ๐๐ก๐๐ฅ๐ฅ๐๐ง๐ ๐ ๐ญ๐ก๐ ๐๐ญ๐๐ญ๐, ๐ฉ๐ซ๐จ๐ฏ๐จ๐ค๐ ๐๐จ๐ง๐๐ซ๐จ๐ง๐ญ๐๐ญ๐ข๐จ๐ง ๐ฐ๐ข๐ญ๐ก ๐ฅ๐๐ฐ ๐๐ง๐๐จ๐ซ๐๐๐ฆ๐๐ง๐ญ, ๐๐ง๐ ๐ญ๐ก๐๐ง ๐๐ซ๐ฒ โ๐จ๐ฉ๐ฉ๐ซ๐๐ฌ๐ฌ๐ข๐จ๐งโ ๐ฐ๐ก๐๐ง ๐ญ๐ก๐ ๐ฉ๐จ๐ฅ๐ข๐๐ ๐ฉ๐๐ซ๐๐จ๐ซ๐ฆ ๐ญ๐ก๐๐ข๐ซ ๐๐จ๐ง๐ฌ๐ญ๐ข๐ญ๐ฎ๐ญ๐ข๐จ๐ง๐๐ฅ ๐๐ฎ๐ญ๐ฒ ๐ญ๐จ ๐ฆ๐๐ข๐ง๐ญ๐๐ข๐ง ๐ฅ๐๐ฐ ๐๐ง๐ ๐จ๐ซ๐๐๐ซ. ๐
๐ซ๐จ๐ฆ ๐ฌ๐ญ๐ซ๐๐๐ญ ๐ฏ๐๐ญ๐จ๐๐ฌ ๐ญ๐จ ๐ซ๐ก๐๐ญ๐จ๐ซ๐ข๐ ๐๐๐จ๐ฎ๐ญ ๐๐ฅ๐จ๐จ๐๐ฌ๐ก๐๐, ๐ญ๐ก๐ข๐ฌ ๐ฉ๐จ๐ฅ๐ข๐ญ๐ข๐๐ฌ ๐ซ๐ข๐ฌ๐ค๐ฌ ๐ฉ๐ฎ๐ฌ๐ก๐ข๐ง๐ ๐ ๐ซ๐ข๐ฌ๐ข๐ง๐ , ๐๐จ๐ง๐๐ข๐๐๐ง๐ญ ๐๐ก๐๐ซ๐๐ญ ๐ญ๐จ๐ฐ๐๐ซ๐๐ฌ ๐ข๐ง๐ฌ๐ญ๐๐๐ข๐ฅ๐ข๐ญ๐ฒ ๐๐ง๐ ๐๐ก๐๐จ๐ฌ.
But India is not Bangladesh. It is not Nepal.
India is a strong, resilient Republic, anchored in its Constitution, democratic institutions and the will of its people. And beyond law enforcement stands the greatest safeguard of all: crores of law-abiding Indians who will not allow the sanctity of our Republic, our Tricolour or our sovereignty to be undermined.
No political ambition is greater than the nation. No street mobilisation can substitute for the peopleโs mandate. And no call for upheaval can override the Constitution. Every attempt to undermine democratic institutions must be confronted through the rule of law, constitutional safeguards and the unwavering resolve of the Indian people.
๐๐จ๐ง๐ ๐๐ข๐ฏ๐ ๐ญ๐ก๐ ๐๐ง๐๐ข๐๐ง ๐๐๐ฉ๐ฎ๐๐ฅ๐ข๐. ๐ฎ๐ณ
Jai Hind!
Congrats to StepFun on Step 5 Preview ๐ Their largest open weights release.
The 600B MoE model is built to handle agent tasks lasting up to 24 hours.
The model runs at nearly 8K tokens/sec/GPU on NVIDIA GB300 using NVFP4.
NVP4 checkpoint: huggingface.co/stepfun-ai/Stโฆ
Deploy with @SGL_project - docs.sglang.io/cookbook/autoโฆ ย or @vLLM_project - recipes.vllm.ai/stepfun-ai/Sโฆ
Post-train with NeMo Automodel: github.com/NVIDIA-NeMo/Automโฆ
Run it with NVIDIA NIM using a single command: catalog.ngc.nvidia.com/orgs/โฆ
Step 5 Preview is now open weights.
BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog.
โข Sparse MoE: 600B total, 27B active per token
โข 1M-token context window
โข Native text and vision input
โข Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance
Frontier intelligence, open weights. Build what's next.
Hugging Face: huggingface.co/stepfun-ai/Stโฆ
ModelScope: modelscope.ai/models/stepfunโฆ
GitHub: github.com/stepfun-ai/Step-5โฆ
Yash retweeted
๐ Nemotron-3.5-Super-VL is here! ๐
โ
Same base architecture as Super-3.0: 120B total, 12B active plus Vision encoder
โ
Ahead of Ultra-3.0 on most benchmarks! And now with visual reasoning!
โ
SWEBench-Pro 51.79%, AA index of 26 vs 23 for Ultra-3.0
Post-training pipeline is a greatly simplified and improved version of what we had for Ultra-3.0.
We reduced number of teachers, SFT set size while significantly improving data and environments quality and made it multimodal! 1/5
๐ค Made with AI

































