𝚐𝔪𝟾𝚡𝚡𝟾

@gm8xx8

161 Following

131518 Followers

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

LONG LIVE OPEN SOURCE

0 reply

3 recasts

18 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

1 reply

2 recasts

21 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

6 replies

0 recast

12 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

3 replies

0 recast

29 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

The π₀ release introduces a VLA generalist model for dexterous tasks like laundry folding and table bussing. π₀ uses a transformer with flow matching, combining VLM pre-training benefits and continuous action chunks at 50Hz, and is pre-trained on a broad dataset. With distinct pre-training and post-training stages, it supports zero-shot and fine-tuned task adaptation, demonstrating robustness to external interventions, as seen in an uncut video of π₀ folding laundry with a single model. π₀ and its smaller, non-VLM version are evaluated against: - Octo and OpenVLA for zero-shot VLA tasks - ACT and Diffusion Policy for single tasks π₀ surpasses in zero-shot accuracy, fine-tuning for new tasks, and language-following. Compute-parity ablations highlight trade-offs between VLA backbone gains and pre-training costs. Hierarchical methods like RT-H aid complex tasks needing low-level control and high-level planning, though Pi_0’s robust architecture largely drives its performance. (link below)

2 replies

2 recasts

150 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Me trying to hold back all the breakthroughs, optimization hacks, new models, and products I’ve been testing.

0 reply

1 recast

24 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

1 reply

1 recast

7 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Oasis is a 500M parameter AI model that generates real-time, interactive gameplay directly from user input, eliminating the need for a traditional game engine. Rendering immersive experiences at 20 fps with zero latency, Oasis introduces “Generative Interactive Experiences,” where users can interact with AI-created video or game content in real time, responding instantly to text or voice commands. Built on Decart’s optimized ViT + DiT architecture, this promises more dynamic, personalized digital experiences. Oasis is available open-source, complete with code, weights, and a live demo for exploration. demo: https://oasis.decart.ai/welcome

1 reply

2 recasts

17 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

you know what impresses me? making a small model work well ☺︎ few understand this.

1 reply

0 recast

13 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning project page: dexmimicgen.github.io DexMimicGen automates data generation for bimanual dexterous robots, creating 21K demos from just 60 human demonstrations. It provides a scalable way to train robots in complex, coordinated tasks through simulation, with applications to real-world tasks like can sorting.

0 reply

1 recast

13 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

clearing out some drafts…

0 reply

0 recast

17 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

OpenAI’s London DevDay—Romain Huet demonstrated a live drone setup in under 2 minutes using o1, along with a London Tube app built on stage. The audience gained full o1 access, including updated Realtime API pricing with caching: 50% off cached text inputs and 80% off cached audio. The Realtime API now offers five expressive voices—Coral, Verse, Ballad, Sage, and Ash—supporting new speech-to-speech capabilities with steerable vocal tones. 🔗: https://platform.openai.com/docs/guides/realtime OpenAI launches SimpleQA, a factuality benchmark assessing language models accuracy on short, fact-seeking questions. 🔗: https://openai.com/index/introducing-simpleqa/ Also—Advanced Voice is now available in macOS and Windows desktop apps. 🔗: https://openai.com/chatgpt/download/

1 reply

2 recasts

22 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

0 reply

0 recast

3 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Big if true—I was hesitant to share and even deleted my post on 𝕏 (for reasons) but given their work around SSM, I felt compelled to share. Also, magicailabs.twitter is trash (iykyk)

4 replies

1 recast

6 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Reaching 1B Context Length With RAG Zyphra: https://www.zyphra.com/post/reaching-1b-context-length-with-rag retrieval system enables LLMs to process up to 1 billion tokens efficiently on a standard CPU using a sparse graph-based approach. Outperforms RAG methods with dense embeddings or long-context transformers. I’m impressed with the work Zyphra has been doing in the SSM space (most recently Zamba2-7B) so I’m eager to see more.

6 replies

5 recasts

34 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

1 reply

1 recast

4 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

“software that writes more software”

0 reply

1 recast

15 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

0 reply

0 recast

4 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Zyphra introduced Zamba2-7B, a new SLM that surpasses competitors like Mistral-7B, Gemma-7B, and Llama3-8B in both performance & efficiency. Designed for on-device use, consumer GPUs, and enterprise applications. Zamba2-7B: > Achieves top benchmark results with faster inference—25% quicker first token generation and 20% more tokens per second compared to similar models. > Utilizes Mamba2 blocks w/ two shared attention blocks interleaved in an ABAB pattern, enhancing cross-sequence dependencies. > Trained on 3T tokens, combining proprietary Zyda data w/ high-quality, deduplicated datasets, followed by a specialized annealing phase to improve model precision. > Zamba2-7B achieves high throughput and low memory usage, eliminating the need for KV-caches & leveraging efficient Mamba blocks optimized for parallel hardware. > Model was trained on 128 H100 GPUs over 50 days and will be available open-source under Apache 2.0. w/ integrations on Huggingface and PyTorch.

1 reply

9 recasts

50 reactions

𝚐𝔪𝟾𝚡𝚡𝟾 pfp

𝚐𝔪𝟾𝚡𝚡𝟾

Powered by Qwen Code 2.5 & WebLLM https://huggingface.co/spaces/cfahlgren1/qwen-2.5-code-interpreter

0 reply

1 recast

3 reactions