Desh Raj @rdesh26

Second big release from us today: Nemotron-3.5-ASR-Streaming! 🌎40 languages ⚡️80ms - 1s controllable latency 🔥240 - 2400 concurrent streams on 1xH100 🧱FastConformer Cache-Aware RNN-T architecture huggingface.co/nvidia/nemotro…

Very excited to share Nemotron 3 Ultra, 550B total with 55B active MoE hybrid Mamba-Attention model post-trained for agentic applications! This model delivers frontier-level agentic accuracy while being fast and open with weights, training software, and data available for commercial use. 1/4

NVIDIA Nemotron 3 Ultra is now live! Frontier accuracy, 5X greater speed, 30% lower cost. Deploy however you need - on-premise, on the cloud, or at the edge. Model is live on HuggingFace under the OpenMDW 1.1 license. youtube.com/watch?v=D8LIIv…

NVIDIA just announced the release of Nemotron 3 Ultra in Jensen Huang's Computex keynote: at 550B parameters (55B active), this is the largest Nemotron 3 model to date, and it is the most intelligent US open weights model We partnered with @nvidia to evaluate this model for intelligence and speed - these figures use the model’s BF16 weights, but as with Nemotron 3 Super the model will be made available in NVFP4 quantization as well for higher inference performance. ➤ New leader for US open weights intelligence: Nemotron 3 Ultra scores 48 on the Artificial Analysis Intelligence Index. This is well ahead of the next strongest US open weights models, Gemma 4 31B (39), Nemotron 3 Super (36) and gpt-oss-120b (33), but behind the Chinese-led open weights frontier (Kimi K2.6 at 54). ➤ Leading speed for its intelligence: on a pre-release @DeepInfra endpoint, Nemotron 3 Ultra served over 300 tokens per second. Peer models in its size class from China-based labs such as DeepSeek and Moonshot (Kimi) are generally served at speeds of 50-100 tokens per second in the market today. gpt-oss-120b is served at speeds similar to this level, but with significantly lower intelligence. ➤ Largest Nemotron 3 model so far: at approximately 550 billion total parameters and 90% sparsity, Nemotron 3 Ultra is significantly larger than its siblings and is the largest recent US open weights model release We’ll be sharing additional analysis and full benchmarks at release.

Nemotron 3 Ultra: Frontier smart. 5X faster. 30% cheaper. 💚💚💚

Imagine you spent 40 years doing the boring, responsible thing. You opened a 401k at 23. You contributed every paycheck. You ignored the noise. You bought the index because Bogle told you to, because Buffett told you to, because every honest piece of financial advice for 30 years told you the index was the safest, most diversified, most rules-based way to own America. The whole point was the rules. The rules said: a company must trade for 12 months before joining the S&P 500. The rules said: it must show four consecutive quarters of GAAP profitability. The rules existed because in 1999 the index quietly bought a lot of stocks at the top, and pensioners paid the bill. After the dot-com crash, S&P tightened the rules. Nasdaq tightened the rules. FTSE Russell tightened the rules. For 23 years, those rules held. Then SpaceX filed for IPO. And the rules changed. The S&P 500 waived the profitability requirement. Nasdaq cut its trading-history window from 90 days to 15. FTSE Russell cut its to 5. Bloomberg Intelligence estimates the major index funds will absorb between 19% and 24% of SpaceX's float within six months. That's over $30 trillion of passive 401k and retirement money, mechanically buying a single newly public company at IPO valuations, because the rules said they had to. Except the rules used to say they didn't. Here's the thought exercise: If you spend 40 years building a system designed to protect ordinary savers from buying overpriced stocks, and then you waive the protections the moment a sufficiently large stock asks you to, what was the system actually protecting? Most of investing is about understanding what's a rule and what's a guideline. A rule binds the rule-maker. A guideline binds the saver. You're allowed to find out which is which only after the fact.

Hedgeye @Hedgeye

7 days ago

Rule changes for the SpaceX $SPCX IPO: Index providers waived the profitability requirement and cut the seasoning window from 90 days to 5. This forces over $30 trillion in passive 401k and retirement money to buy SpaceX at IPO valuations. Bloomberg Intelligence estimates S&P

550 2K 10K 11.6M 4K

First they came for the asylum seekers, and I did not speak out - they were a drag on GDP, and I was busy accelerating the techno-capital machine. Then they came for the undocumented workers, and I did not speak out - they had not contributed to foundational AI. Then they came for the DACA kids and the OPT students, and I did not speak out - my visa was already stamped. Then they came for the H-1Bs, and I did not speak out - surely they meant the lower-skilled ones, not technologists like me. Then they came for my path to citizenship, and I felt robbed - because I had a PhD, and ten years of foundational contribution, and I had assumed the policy was for the other immigrants.

@niksdagr8 I am not sure, sorry. I think the event was deliberately kept private to encourage more open discussions.

I was invited to participate in a panel on "multimodal foundation models" last night.

@chenwanch1 Thanks for attending, and for the great picture!

We are offering grants of $100,000 + Tinker credits to researchers advancing the field of human-AI interactivity. Submit your proposals by June 19th! thinkingmachines.ai/news/interacti…

@TanelAlumae 👇

Desh Raj @rdesh26

3 weeks ago

@xutai_ma Here is the conversation in a gist: gist.github.com/desh2608/5d601… TL;DR: I gave the Thinky blog link and asked it to draw the illustrations in a similar design. For the first animation, I iterated on the design a few times (asking it to render with Python code instead of HTML, etc.).

1 0 5 291 0

x.com/i/article/2054…

I can’t stop thinking about this post. If you do one thing today, I encourage you to give it a thoughtful, thorough read… And then commit to never living your life this way. Life has wasted success on the people described in this post. It really is completely pathetic. They say that comparison is the thief of joy - look no further than this post for validation it is indeed true. On their deathbed they will realize they have lived their life completely wrong. Don’t let it be you.

Deedy @deedydas

3 weeks ago

The vibes in SF feel pretty frenetic right now. The divide in outcomes is the worst I've ever seen. Over the last 5yrs, a group of ~10k people - employees at Anthropic, OpenAI, xAI, Nvidia, Meta TBD, founders - have hit retirement wealth of well above $20M (back of the envelope

1K 1K 16K 13.3M 13K

It is 70 and sunny in New York City, we’re heading to a kite festival, and I haven’t heard the words “agent” or “token” once all morning. Greatest city in the world.

Deedy @deedydas

3 weeks ago

The vibes in SF feel pretty frenetic right now. The divide in outcomes is the worst I've ever seen. Over the last 5yrs, a group of ~10k people - employees at Anthropic, OpenAI, xAI, Nvidia, Meta TBD, founders - have hit retirement wealth of well above $20M (back of the envelope

1K 1K 16K 13.3M 13K

👀 ahh damn, someone should have told me I was Founding Research Scientist at Meta!

leon @towheretobegin

3 weeks ago

if your email is firstname@company then you're founding, actual role notwithstanding

19 4 878 156K 127

We've gone even farther: Nemotron 3 Super is 120B and pretrained on 25T tokens in NVFP4. Nemotron 3 Ultra is ~500B and also pretrained in NVFP4. Accelerated computing means we rethink every aspect of the AI stack looking for new opportunities to improve efficiency.

How To AI @HowToAI_

3 weeks ago

NVIDIA has done the impossible and nobody's talking about it. They trained a 12 BILLION parameter LLM in 4-bit precision on 10 trillion tokens. For years, the AI industry has been stuck. If you wanted to train a world-class AI, you had to use 16-bit or 8-bit precision. Going

50 144 1K 201K 720

It seems LLM folks have finally discovered multi-stream models. Let me introduce you to array geometry invariant multi-channel multi-speaker ASR.

Here is the conversation in a gist: gist.github.com/desh2608/5d601… TL;DR: I gave the Thinky blog link and asked it to draw the illustrations in a similar design. For the first animation, I iterated on the design a few times (asking it to render with Python code instead of HTML, etc.). Once the design was satisfactory, the rest of the animations were faster by explaining what I wanted to show. I didn't check the code much, but looks like it only used packages from the Python standard library. Perhaps using something like manim would be prettier.

@vipul_1011 Williamsburg?