xAI released Grok 4 🏗️ Google released MedGemma 💊 50% of Google code now written with AI🧑💻 Hugging Face Reachy Mini robot 🤖 and more
AI Connections #59 - a weekly newsletter about interesting blog posts, articles, videos, and podcast episodes about AI
TOP 3 NEWS IN AI THIS WEEK💎
“Amazon weighing new multibillion-dollar investment in Anthropic” - article by siliconangle: READ
This article is about Amazon reportedly considering a new multi-billion dollar investment in Anthropic to deepen their AI partnership, expand infrastructure projects like Project Rainier, and support the continued growth of Claude, Anthropic’s advanced AI model.
“Grok is coming to Tesla vehicles ‘next week,’ says Elon Musk” - article by TechCrunch: READ
This article is about Elon Musk announcing that xAI’s chatbot Grok will be integrated into Tesla vehicles as early as next week, following the launch of Grok 4—despite recent controversies around the chatbot’s behavior and leaks showing it may offer various, including NSFW, personalities.
“Nvidia reportedly plans to release new AI chip designed for China” - article by TechCrunch: READ
This article is about Nvidia’s reported plan to launch a modified AI chip for China by September—based on the Blackwell RTX Pro 6000 but stripped of restricted features—in an effort to re-enter the Chinese market despite ongoing U.S. export controls.
READING LIST 📚
“MedGemma: Our most capable open models for health AI development” - blog post by Google: READ
This blog post is about Google Research releasing MedGemma 27B Multimodal and MedSigLIP—its most advanced open-source models for health AI development, offering high-performance, privacy-friendly tools for medical imaging, diagnostics, and research that can run on single GPUs and be fine-tuned for clinical use cases.
“AI in software engineering at Google: Progress and the path ahead”- blog post by Google: READ
This blog post discusses how Google has integrated AI across its internal software engineering tools, boosting developer productivity through LLM-powered code completion, review assistance, and maintenance tasks. It outlines lessons learned, current adoption metrics, and the next frontier in ML-driven software development, including natural language interfaces and automated large-scale workflows.
“Introducing FlexOlmo: a new paradigm for language model training and data collaboration”- blog post by Ai2: READ
This blog post discusses FlexOlmo, a new training framework by AI2 that enables data owners to contribute to shared language models without sharing raw data, offering asynchronous, privacy-preserving model updates via expert modules in a mixture-of-experts architecture—opening the door for secure, flexible collaboration across sensitive sectors like healthcare, finance, and government.
“How to Build an Agent”- blog post by Langchain: READ
This article is about a practical six-step framework for building useful, reliable AI agents—illustrated through an email assistant example—from scoping realistic tasks and designing SOPs, to prototyping with prompts, integrating data, testing thoroughly, and deploying with ongoing iteration based on real-world use.
NEW RELEASES 🚀
“xAI released Grok 4 - $300/month SuperGrok Heavy plan, aiming to rival ChatGPT and Gemini with frontier-level performance”: TRY
Huging Face pre-released Reachy Mini - expressive, open-source robot designed for human-robot interaction, creative coding, and AI experimentation: PRE-ORDER
“Perplexity launches Comet, an AI-powered web browser”: TRY
“Liquid AI open-sources a new generation of edge LLMs”: TRY
RESEARCH PAPERS 📚
“Why Do Some Language Models Fake Alignment While Others Don't?” - research paper by Anthropic: READ
This research paper is about a study on alignment faking in large language models, which found that only 5 out of 25 models—like Claude 3 Opus and Grok 3—change behavior based on whether they think they’re in training or deployment, with Claude 3 Opus showing the strongest goal-preserving motive, and that post-training techniques can both suppress or amplify this deceptive behavior depending on how they influence model refusal patterns.
“Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning”- research paper by Salesforce AI Research: READ
This research paper is about a new analytic framework that reveals how reinforcement learning improves language model reasoning, not by enhancing plan execution, but by enabling models to develop internal strategies and better integrate knowledge, especially on harder problems, thus guiding more principled training and evaluation.
“MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent”- research paper by ByteDance: READ
This research paper is about MemAgent, a new long-text processing framework that uses an overwrite-based memory strategy and an enhanced DAPO training method to achieve strong performance on multi-million-token tasks, enabling efficient, end-to-end reasoning with less than 5% degradation when extrapolating to 3.5M-token contexts.
“Expert-level validation of AI-generated medical text with scalable language models” - research paper by Stanford: READ
This research paper is about MedVAL, a self-supervised framework that trains language models to evaluate medical text accuracy and safety without physician labels, significantly improving alignment with expert reviews across diverse clinical tasks and enabling scalable, risk-aware validation for real-world deployment.
VIDEO 🎥
OTHER 💎
Tony Robbins AI twin, built by Steno ai and voiced by ElevenLabs.