Oleksii Kuchaiev

1/4 We see no wall in post-training. Scaling RL software, infra, and data keeps yielding major capability gains. We trained across 30 RL environments with up to 4,000 instances per batch — math, code, STEM, agentic tool use, SWE, terminal, safety — all in a unified multi-environment RLVR setup.

@xeophon @arcee_ai Open > closed

Classmates and friends lay flowers at a makeshift memorial outside the collapsed apartment building. A Russian attack killed 24 people here, including 3 children.

We are open sourcing renderers For RL, the inference server should be simple Tokens in, tokens out renderers is the token-level chat templating layer to >render messages to tokens >parse completions to structure >bridge rollouts byte-for-byte > >3x throughput on openmodels

This is what modern russia is

Lab is launching with self-serve support for models from Nvidia, OpenAI, Meta, Qwen, with more coming soon. Models range from 1B to 400B parameters covering both dense and MoE architectures, reasoning and non-reasoning modes, and text and image modalities.

@Scobleizer Build in AI Applications layer, on top of open models.

Open Source's big problem. Last night I went to a Y Combinator party in San Francisco and met an entrepreneur who is making a top Open Source AI model. He told me it is very hard to make money in open source. Yeah, it is cool being popular, he told me, but figuring out how to make a business out of it is proving to be very difficult. The Chinese are pounding the price into the ground with their open source models. Which makes it tough. In the old world of Open Source you could make money with them by consulting, service, etc, like RedHat did. But in this new world, he told me, it's much harder to make a good business out of it. Is anyone making a good business out of open source? What would your advice be to the businesses that are trying to support Open Source?

New model dropped today. This is a fast one!

Meet Nemotron 3 Nano Omni 👋 Our latest addition to the Nemotron family is the highest efficiency, open multimodal model with leading accuracy. 30B parameters. 256K context length. 🧵👇

HF Article huggingface.co/blog/nvidia/ne… Tech Report research.nvidia.com/labs/nemotron/…

Nemotron 3 Nano Omni model is now available. Text, Image, Video and Audio as input with Text as output. As all models from Nemotron 3 family, fully open (weight, base models, data and code) and extremely efficient.

Today we're releasing Nemotron 3 Nano Omni. Audio, Video, Image, Text ➡️ Text Ask questions about all your data. Amazing efficiency powered by the Nemotron Hybrid SSM MoE architecture. State of the art multimodal intelligence.

Trump is setting America back — not putting America First, as he promised. We will lose the future because of his reckless, harmful actions.

@WSJ I hope Apple gets their AI strategy together and creates cohesive AI agent a experience across their device ecosystem. A much better experience is possible than talking to your AI agent running on Mac via Telegram messenger on iPhone.

Apple’s new CEO John Ternus is a hardware expert who must help the company catch up in the AI race as it looks for its next big hit. 🔗 on.wsj.com/4tvNf1Q

@tedlieu You are correct. There is nothing "conscious" about current AI. And there is no universaly agreed definition of consciousness.

My take: linear algebra equations will never be conscious. Random number generators will never be conscious. At a very basic level, AI is math. AI can act like it is conscious, but it will never be conscious. And adding more math to AI models doesn’t make it any more conscious.

@polynoamial No one needs to die on that hill because this is obviously true.

A hill that I will die on: with today's AI models, intelligence is a function of inference compute. Comparing models by a single number hasn't made sense since 2024. What matters is intelligence per token or per $. This is especially true when using it in a product like Codex.

gonna be chatting w @danielhanchen + the @nvidia nemotron team next week + talking RL come hang out :)

@tedlieu @thejointstaff Thank you, @tedlieu, for speaking up

Dear @thejointstaff: The UCMJ and federal law prohibit war crimes. Obviously eradicating a whole civilization constitutes a war crime. You must disobey that order. If you commit war crimes, the next Administration will prosecute you.

@hwchase17 Continual learning is a model-harness co-design problem.

most people thinking of continual learning as happening at the model level but with agents - there's actually three different levels you could "learn" at: - model - harness - context

While the whole timeline laments about Claude sub being reserved for Claude things - let me remind you that Nemotron 3 Super is a fine model and you can do whatever you'd like with it because you own the weights! And if you want a guide on how to hook it into OpenClaw (plus safety bells and whistles) check out NemoClaw! github.com/NVIDIA/NemoClaw

Liftoff. The Artemis II mission launched from @NASAKennedy at 6:35pm ET (2235 UTC), propelling four astronauts on a journey around the Moon. Artemis II will pave the way for future Moon landings, as well as the next giant leap — astronauts on Mars.