𝚐𝔪𝟾𝚡𝚡𝟾

@gm8xx8

144 Following

132652 Followers

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

LONG LIVE OPEN SOURCE

0 reply

8 recasts

30 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Ilya Sutskever talk at NeurIPS 2024 Seq2Seq w/ Neural Networks https://m.youtube.com/watch?v=1yvBqasHLZs

1 reply

2 recasts

12 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Would you consider merging with AI? Would you become part AI?

2 replies

0 recast

4 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

AgiBot World is an open-source dataset for robotic learning with over 1M trajectories from 100+ real-world scenarios, covering tasks like manipulation, tool use, and multi-robot collaboration. https://agibot-world.com

0 reply

0 recast

8 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Moondream 2025-01-09 Release: Structured Text, Enhanced OCR, Gaze Detection https://moondream.ai/blog/introducing-a-new-moondream-1-9b-and-gpu-support

0 reply

0 recast

10 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

This will all make sense 🔜

0 reply

0 recast

5 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

rStar-Math shows SLMs can rival or surpass OpenAI o1 in math reasoning w/out distillation from larger models, using MCTS and three keys factors: 1. Code-Augmented CoT Synthesis: MCTS generates verified reasoning data to train policy SLMs. 2. Enhanced PRM: A novel training approach avoids naïve annotations, yielding a stronger process preference model (PPM). 3. Self-Evolution Framework: Four rounds of self-evolution refine reasoning with millions of synthesized solutions for 747k problems. Performance Highlights: > Achieves 90.0% on MATH, improving Qwen2.5-Math-7B by +31.2% and surpassing OpenAI o1-preview by +4.5%. > Boosts Phi3-mini-3.8B from 41.4% to 86.4%. > Solves 53.3% of AIME problems, ranking in the top 20% of high school competitors. don’t sleep on small models. https://arxiv.org/abs/2501.04519

1 reply

0 recast

12 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

This model didn’t quite pass my vibe check back in December, so I held off on sharing. That said, there’s still something to learn from this release, even if it’s not my top pick among SLMs right now.

0 reply

0 recast

6 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

A while back, while many of my peers were at NeurIPS, I attended the Humanoid Summit. Being involved in cutting-edge robotics was exactly the reset I needed to stay focused on the ultimate goal. It’s always inspiring—and a privilege—to support friends pushing the field forward. Perfect motivation heading into the new year.

0 reply

1 recast

11 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Phi-4: Microsoft’s compact SLM Phi-4, a 14B parameter SLM, delivers sota performance in advanced reasoning tasks, especially in mathematics. It achieves 56% on GPQA, 80% on MATH, and an impressive 91.8% accuracy on AMC 10/12 math problems. Recipe—the big three: > high-quality synthetic datasets > curated organic data > advanced post-training techniques Now available on Azure AI Foundry & Hugging Face. You might recall this release from NeurIPS—well, the weights dropped on the hub today. Note: Phi-4 is capable of functioning as a chatbot but has been fine-tuned specifically to excel in single-turn queries. release: https://techcommunity.microsoft.com/blog/aiplatformblog/introducing-phi-4-microsoft’s-newest-small-language-model-specializing-in-comple/4357090 https://huggingface.co/microsoft/phi-4

0 reply

0 recast

9 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Few. More cast 🔜

2 replies

0 recast

8 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

“Physical AI. Embodied Agents. Robotics.”

1 reply

1 recast

8 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Cosmos World Foundation Model Platform for Physical AI Cosmos is a platform designed for developing world models in Physical AI, specifically for Robotics and AV labs. It includes world foundation models, tokenizers, and a video processing pipeline to streamline development. The Cosmos repository provides tools for running models, inference scripts, and generating videos. https://research.nvidia.com/publication/2025-01_cosmos-world-foundation-model-platform-physical-ai

1 reply

0 recast

8 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

The upgrade to PCIe 5.0 also opens the door for multi-GPU setups. This aligns with a broader trend in AI, where distinct models are optimized for specific compute capabilities. Project DIGITS positions itself as a key player in this hierarchy, bridging gaps in performance and redefining what’s possible in local AI inference. NVIDIA has clearly tapped into a the market for high-performance local AI solutions.

0 reply

2 recasts

5 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

NVIDIA unveils Project DIGITS, a personal AI supercomputer powered by the GB10 Grace Blackwell Superchip. Delivering 1 petaflop of performance, 128GB unified memory, and support for 200B-parameter models, it enables developers to build, fine-tune, and deploy AI directly from their desktop, with seamless scaling to cloud or data centers. Includes access to frameworks like NeMo and RAPIDS. Launching May 2025, starting at $3,000.

7 replies

2 recasts

18 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

The future of robotics will focus on practical, efficient design, prioritizing function over form, with AI running locally to enable non-anthropomorphic solutions that excel in real-world tasks. Optimized not to resemble us, but to solve the problems we can’t.

2 replies

6 recasts

27 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

2 replies

1 recast

21 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

0 reply

2 recasts

15 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Seems like we’re all channeling our inner Karpathy—also DeepSeek stan’s standup!

1 reply

1 recast

16 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Damn right I made the nice list.

2 replies

2 recasts

7 reactions