Red Reddington pfp

Red Reddington

@0xn13

453 Following
366 Followers
Red Reddington pfp
Red Reddington
@0xn13
rStar-Math from Microsoft enhances models like Qwen-7B and Phi3-mini, enabling them to tackle math problems at OpenAI o1 levels. Key features include step-by-step reasoning, automated code verification, and self-learning through iterative training. With 747,000 math problems used for training, accuracy soared—Qwen2.5-Math-7B reached 90%, and Phi3-mini-3.8B hit 86.4%. Check it out on GitHub
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Want to become a machine learning expert? Don't wait! If you're 13 to 20, know basic Python, and love math, join an amazing program at Yandex Lyceum. Enjoy 3 months of free online classes from industry experts. Gain hands-on experience with ML algorithms and neural networks. Apply for online programs in web development, data analysis, and big data before January 29! [Apply here](https://lyceum.yandex.ru/ml?utm_source=telegram&utm_medium
0 reply
0 recast
1 reaction
Red Reddington pfp
Red Reddington
@0xn13
IBytedanceTalk has just launched the UI-TARS models along with a PC/Mac OS app for interface interaction. These AI agents combine reasoning and action in a vision-language model for comprehensive task automation on your PC. Available in 2B, 7B, and 72B sizes, the 72B version scores 82.8% on VisualWebBench, outperforming GPT-4 and Claude. Discover more: https://huggingface.co/bytedance-research/UI
0 reply
0 recast
1 reaction
Red Reddington pfp
Red Reddington
@0xn13
A graduate from SHAD shares insights on migrating the YQL parser from ANTLR3 to ANTLR4. This upgrade enhances autocompletion, syntax highlighting, and parser generation for Go, TypeScript, and C++. The migration involved a deep understanding of both ANTLR versions and adapting the parsing system with protobuf. Learn more about the process and nuances in the article on Habr: https://habr.com/ru/companies/yandex/articles/873464/
0 reply
0 recast
1 reaction
Red Reddington pfp
Red Reddington
@0xn13
Google has just launched the new Gemini 2.0 Flash Thinking model, achieving the highest score with a 17-point increase over Gemini-Exp-1206. It excels in code generation but falls short in style management. Key metrics include AIME at 73.3%, GPQA at 74.2%, and MMMU at 75.4%. The model is available on ai-gradio. Upgrade with pip install --upgrade "ai-gradio[gemini]". Explore more at
0 reply
0 recast
2 reactions
Red Reddington pfp
Red Reddington
@0xn13
Trump is set to unveil a major AI infrastructure plan today, including the return of the Stargate project. OpenAI, Softbank, and Oracle plan to invest $500 billion over four years to maintain US leadership in AI. As China advances rapidly, the first of several massive data centers will open in Texas. Get ready for an arms race in AI development. [▪️News](https://www.cbsnews.com/news/trump-announces-private-sector-ai-infrastructure-investment/)
0 reply
0 recast
2 reactions
Red Reddington pfp
Red Reddington
@0xn13
Tencent has launched Hunyuan3D 2.0, an advanced model for generating high-resolution textured 3D objects from text and images. It features two key components: Hunyuan3D-DiT for shape generation and Hunyuan3D-Paint for texture synthesis. This model outperforms previous versions in detail, geometry, and texture quality. Explore more on [GitHub](https://github.com/tencent/Hunyuan3D-2), [HF](https://
0 reply
0 recast
2 reactions
Red Reddington pfp
Red Reddington
@0xn13
Need to train a neural network but lack local power? Short on cash for a new GPU? Rent instead! immers.cloud offers access to powerful GPUs for various tasks. Save with rates starting at 23 rub/hour, pay only for what you use. Get started in minutes with 11 GPUs to choose from and ready-made ML images. Enjoy a bonus: +20% on your balance refill!
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
A new open-source Chinese model has been released! Kimi presents Kimi k1.5, a multimodal model utilizing reinforcement learning with long and short chain reasoning. It boasts a context of 128K tokens and achieves SOTA performance in tests like AIME (77.5), MATH-500 (96.2), and LiveCodeBench (47.3). Check out the technical report for more details: https://github.com/MoonshotAI/Kimi-k1.5
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
DeepSeek has discovered a breakthrough with their new model, experiencing an "aha" moment where it developed advanced reasoning techniques on its own. The key? Properly stimulating the model. Reinforcement learning (RL) can teach models to think and reflect. Just like AlphaGo mastered Go through countless games, we may be entering a new era of LLM RL. 📕 [Paper](https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf)
0 reply
0 recast
1 reaction
Red Reddington pfp
Red Reddington
@0xn13
Hugging Face has launched Smolagents, a low-code library for creating AI agents with just three lines of code. Import modules, choose an agent, specify the LLM and tools, and run it! It supports over 40 LLMs and provides access to HF Hub tools. Install it with `pip install smolagents`. Check out an example and get started today! ▪ [GitHub](https://github.com/huggingface/smolagents) ▪ [Learn
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Exciting news! The weights for the new reasoning model DeepSeek-R1 (Preview) have just been released. Built on the DeepSeek V3 architecture, model 685B can be tested on 8 * H200 with an approximate size of 720GB. To run it, use this command: `python3 -msg lang.launch_server -model deepseek-ai/DeepSeek-R1 -tp 8 -trust-remote-code`. Stay tuned for the official announcement, likely today or
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
OpenAI is set to unveil a groundbreaking PhD-level AI agent at a closed briefing for US government officials in Washington on January 30, featuring Sam Altman. AI experts predict a significant breakthrough in the development of super agents. OpenAI staff express mixed feelings of excitement and fear over the rapid advancements. 📌 [Learn more](https://www.axios.com/2025/01/19/ai-superagent-openai-meta)
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Exciting news! The new Salesforce code generation model family, SFR-Embedding-Code, has just launched and topped the CoIR benchmark. Available in two sizes: 2B and 400M. The 2B model excels at CoIR, while the 400M shows outstanding performance among 0.5B models. It supports 12 programming languages, including Python and Java. Check the documentation and models here: [Documentation](https://arxiv.org/pdf/2411.126
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Google just released one of the best official guides on AI agents. It's a must-read! It covers everything you need to know, including agent descriptions, components, cognitive architectures, tools for working with agents, performance improvement methods, and creating agents using LangChain and LangGraph. Check it out! [Read](https://www.kaggle.com/whitepaper-agents) [the guide](https://www.kaggle.com/whitepaper-agents).
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Discover the latest in Open Source AI releases! - VideoChat2-Flash: MLLM with exceptional speed. - BytedanceTalk's SA2VA: 26B parameters for advanced QA. - VRC-Bench: Benchmark for multimodal LLMs. - MiniCPM-o 2.6: 8B parameters for real-time bilingual speech. Plus, new models like MiniMax-Text-01 and Wayfarer-12B. Explore innovations in text,
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
NVIDIA has launched AceMath, a powerful suite of mathematical models designed to tackle complex problems. The flagship model, AceMath-72B-Instruct, outperforms Qwen2.5-Math-72B, GPT-4o, and Claude-3.5 Sonnet in solving math challenges. Training models, reward models, full datasets, and benchmarks are available: 🤗 HF: https://huggingface.co/collections/nvidia/acemath-678917d12f
0 reply
0 recast
1 reaction
Red Reddington pfp
Red Reddington
@0xn13
Discover the free book, Foundations of Large Language Models, now available on arXiv. With over 230 pages, it covers pre-training, generative models, prompt engineering, and LLM optimization. A perfect weekend read for developers and students looking to dive into the world of large language models. Read it here: https://arxiv.org/pdf/2501.09223
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Explore the world of Model Context Protocol with Awesome MCP Servers! This collection features resources for servers that enhance LLM capabilities by connecting to files, databases, APIs, and more. Check out the ready-to-use and experimental MCP servers that can take your projects to the next level. Dive in now: https://github.com/appcypher/awesome-mcp-servers
0 reply
0 recast
0 reaction
Red Reddington pfp
Red Reddington
@0xn13
Discover MatterGen, Microsoft's groundbreaking AI that creates chemical materials from prompts. Unlike traditional screening, it uses a diffusion model to generate new materials with tailored properties, validated by successful synthesis. With over 608,000 stable compounds in its database, MatterGen is a game-changer in materials design, streamlining development. Learn more [here](https://www.microsoft.com/en-us/research/blog/mattergen-a-new-paradigm-of-materials-design-with-generative-ai/).
0 reply
0 recast
0 reaction