Frontier AI Update: xAI releases Grok 4, Apple to integrate Alibaba AI into iPhones in China, Alibaba Qwen launches desktop version, and more major tools and opinion updates
NVIDIA's Huang said: "Artificial intelligence is a great equalizer."
Don't underestimate the young people using it. They are putting pressure on those who spent decades mastering skills without AI.
When everyone has a personal tutor, expertise is no longer a moat.
......
Latest advancements in large models:
OpenAI delays open-source model release: Sam Altman announces that the open-source weight model scheduled to be released next week will be postponed due to the need for additional security testing.
xAI releases Grok 4: The official blog post for Grok 4 has been released. Gary Marcus comments that it shows good progress on public benchmarks but still lacks in areas like visual understanding, with performance improvements mainly coming from the integration of symbolic tools.`
https://x.ai/news/grok-4
Moonshot (Moonshot) releases K2 model: Released the open-source K2 model with trillion parameters, claimed to perform strongly on coding and agent tasks, with performance comparable to Claude Opus 4 and GPT-4.1. The model was pre-trained using 15.5T tokens.
https://www.kimi.com/
Mistral AI launches Devstral series: Launched the Devstral Small and Medium 2507 versions, optimized for coding agent and software engineering tasks.
Hugging Face releases SmolLM3: Released the 3B parameter SmolLM3 model, which outperforms models like Llama-3.2-3B in multiple benchmarks, and supports 6 languages and Chain-of-Thought (CoT) reasoning. It can run directly in the browser via WebGPU, and the model is open-source.
https://huggingface.co/spaces/HuggingFaceTB/SmolLM3-3B-WebGPU?v1=
DeepSeek releases DeepSeek-R1-0528: The new version improves benchmark performance, reduces hallucinations, and supports JSON output and function calls.
https://chat.deepseek.com/
Runway launches Gen-4: Introduced the new generation AI video generation model Gen-4, significantly enhancing fidelity, dynamic motion, and controllability.
https://runwayml.com/research/introducing-runway-gen-4
Luma Labs launches Ray2: Introduced the new video generation model Ray2, creating realistic videos with natural and coherent motion through a 10x computational power expansion.
Liquid AI open-sources edge LLM LFM2: Open-sourced the new generation edge computing large language model LFM2, combining reasoning speed and quality, launching three versions: 350M, 700M, and 1.2B.
https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models
Exploring New Architectures: EBTs and Atlas: Researchers have proposed Energy-Based Transformers (EBTs) and the Atlas architecture, aimed at unlocking broader reasoning capabilities and addressing long-term memory issues.
......
Latest AI tools and platform advancements:
Google Gemini app feature update: Sundar Pichai announced that users have created over 40 million videos using Veo 3, and a new "Photo to Video" feature has been added to the Gemini app.
GitHub Copilot feature enhancements: The GitHub mobile app now supports Copilot code review features, while official documentation provides guidance on optimizing prompt engineering.
Cursor integrates Grok 4: AI code editor Cursor announces integration with Grok 4 model.
Anthropic launches Claude Campus Program: The program is open for global students to apply, offering free API credits, event support, and compensation.
Perplexity AI launches Comet browser: Announces the release of Comet, a web browser designed for the modern internet.
Alibaba Qwen launches desktop version and official website: Launched the Qwen Chat desktop version supporting the MCP protocol, and launched the qwen.ai official website integrating chat, research, and API.
qwen.ai/download
OpenAI increases structured output limits: Significantly increased the limits for structured outputs (Structured Outputs) in the API to support larger data schemas.
Apple to integrate Alibaba AI into iPhone in China: According to reports, Apple will integrate Alibaba's AI technology into iPhones in China.
https://www.cnbc.com/2025/02/13/apple-will-integrate-alibabas-ai-into-iphones-in-china-chairman-joe-tsai-says.html
......
Latest AI technology tutorials / Open-source projects:
Nguyen Duy Anh releases new courses: Released two new courses for beginners, "AI Python Programming" and "Post-training for LLMs," both available for free for a limited time.
https://www.deeplearning.ai/short-courses/ai-python-for-beginners/
https://www.deeplearning.ai/short-courses/post-training-of-llms/
GenAI Processors open-source library: Google DeepMind has released a new open-source Python library, GenAI Processors, for building asynchronous, composable AI pipelines suitable for real-time agents and generative AI.
AI video creation tutorial: Demonstrates how to coordinate video production using Claude Code and Hugging Face tools, as well as a project using Gemini and Veo to transform fashion sketches into runway videos, Sketch2Runway.
Best Practices for Rapid Engineering with GitHub Copilot
https://docs.github.com/en/copilot/concepts/prompt-engineering-for-copilot-chat
......
Latest AI in medical advancements:
Microsoft Launches BioEmu to Simulate Protein Motion: Microsoft CEO Satya Nadella and Microsoft Research announced the release of the AI system BioEmu, which can simulate the conformational ensembles of protein structures,有望 in hours provide insights that previously required years of simulation, thus accelerating drug discovery.
ESM3 Model for Protein Research: ESM3, a generative language model capable of reasoning about the three fundamental properties of proteins—sequence, structure, and function—was showcased in the journal Science, and is being made freely available to researchers.
MedSigLIP for Medical Image Embedding: Introducing MedSigLIP, a model for creating embeddings of medical images (such as X-rays, CT scans, MRI) and text, which can be used for tasks like classification, semantic image retrieval, etc.