I'm screwed, open-source 4o image editor, more Google AI updates, Claude 4, AI has higher EQ

Key Concepts

• Open-source multimodal AI
• Image generation and editing
• Visual understanding
• AI emotional intelligence (EQ)
• Vision language models
• AI-powered automation
• AI coding agents
• AI medical analysis
• AI learning tools
• AI video generation

Bagel: Open-Source GPT40 Image Generator and Editor

• Main Point: Bite Dance released Bagel, an open-source multimodal language model similar to GPT40, capable of image generation, editing, and visual understanding.
• Capabilities:
  • Chatbot-like interaction with image-based prompts.
  • Image generation from text prompts (e.g., "photo of three antique glass magic potions").
  • Image editing based on text prompts (e.g., adding a dog to an image).
  • Visual understanding: Solving visual problems (e.g., solving "X" in an image), describing images in detail, identifying objects and people in images.
  • Style transfer: Changing image styles (e.g., to 3D animated, Japanese anime, clay style).
  • Camera manipulation: Rotating the camera view or simulating movement within an image.
  • Reasoning: Step-by-step thinking to generate complex prompts (e.g., generating an image of an animal with nine lives).
• Examples:
  • Generating an image of a cat surrounded by nine heart symbols based on the prompt derived from "an animal with seven plus two lives."
  • Editing an image to show how fabric appears unrolled.
  • Demonstrating makeup colors on skin in an image.
• Availability: Models are available on HuggingFace and GitHub for free download and offline use.

MTV Crafter: Motion Transfer AI

• Main Point: MTV Crafter is an open-source AI tool that transfers movements from a reference video onto a reference character image.
• Process:
  1. Reference video converted to 3D motion data.
  2. 4D motion tokenizer breaks down motion data into chunks.
  3. Motion video model maps movements onto the reference character.
• Limitations: The quality of generated videos is not state-of-the-art compared to tools like Vase by Alibaba.
• Availability: The tool is open-sourced on GitHub.

AI Emotional Intelligence (EQ) Study

• Main Point: A study found that leading AI models (GPT401, Gemini 1.5 Flash, Copilot 365, Cloud 3.5 Haiku, Deepseek version 3) outperformed humans on standard EQ tests.
• Findings:
  • AI models achieved an average score of 81% compared to human participants' 56%.
  • AI models could generate new, reliable EQ tests indistinguishable from human-generated tests.
• Implications: AI could be effective in roles requiring emotional intelligence, such as coaching, therapy, or conflict resolution.

UniVGR1: Vision Language Model for Visual Analysis

• Main Point: UniVGR1 is a new vision language model excelling at visual analysis tasks.
• Capabilities:
  • Finding common objects across multiple images.
  • Spotting differences between images.
  • Locating an object from one image in another image.
  • Reasoning about images (e.g., identifying which furniture can deal with an object in an image).
• Architecture:
  • Based on Quen 2VL (an open-source vision model by Alibaba).
  • Fine-tuned using chain-of-thought supervised fine-tuning (feeding examples of reasoning steps).
  • Reinforcement learning (rewarding correct answers).
• Performance: Outperforms existing vision language models of similar size across benchmark scores.
• Availability: Models are available on HuggingFace, and a free online demo is coming soon.

Skywork Super Agents: AI-Powered Automation

• Main Point: Skywork Super Agents is a suite of AI workspace agents designed to automate work and enhance productivity.
• Capabilities:
  • Web searching and deep research.
  • Generating reports, sheets, slides, web pages, and podcasts.
  • Data analysis and formatting into Excel spreadsheets.
  • Integration with MCPs for expanded capabilities (e.g., document retrieval, music generation, video generation).
• Performance: According to the Gaia agent benchmark, Skywork is better than OpenAI's deep research or Manis.

Google IO Event AI Updates

• VO3: A video generator that natively generates audio with the video.
• Imagine 4: An image generator capable of generating images up to 2K resolution, with good text and topography generation.
• Jules: A free coding agent.
• Stitch: A free UI design platform.
• Realtime AI Assistant: A realistic AI assistant that can be interacted with via voice or camera.
• Notebook LM Updates:
  • Audio Overviews: Converts documents, websites, or videos into podcasts.
  • Video Overviews: Generates full explainer videos from uploaded content.
• Med Gemma: An AI specifically designed for medical analysis, based on the Gemma 3 architecture.
  • Variants: 4 billion parameter model (multimodal, processes images and text) and 27 billion parameter model (text-only, for deep medical comprehension).
  • Capabilities: Interprets radiology scans, pathology slides, and dermatology photos.
  • Availability: Open source and available on HuggingFace.
• Learn LM: An AI learning tool accessible on Gemini, fine-tuned for education and learning.
  • Capabilities: Breaks down topics into learning plans, provides explanations, and suggests activities.
• Interactive Quizzes on Gemini: Allows users to create or generate interactive quizzes on various topics.

Anthropic Claude 4

• Main Point: Anthropic released Claude 4, their most intelligent model to date, with two variants: Opus and Sonnet.
• Variants:
  • Opus: Larger, more performant model for complex problem-solving and reasoning tasks (STEM-related subjects).
  • Sonnet: Lighterweight, faster model for everyday use.
• Features:
  • Hybrid reasoning system with regular and extended thinking modes.
  • Tool use: Web search, code execution, and file analysis.
  • Multitasking and parallel tool use.
• Availability: Claude Sonnet 4 is available on the free plan. Claude Opus 4 and the extended thinking feature require a pro subscription.
• Performance:
  • Self-reported benchmarks show strong coding performance, potentially outperforming OpenAI's 03 and Gemini 2.5 Pro on specific coding tasks.
  • Independent evaluations show mixed results, with Claude 4 not consistently outperforming other models in overall intelligence or specific benchmarks.
• Pricing: Claude 4 is the most expensive model to use via API.

Microsoft AI Updates

• NL Web: An open-source tool for adding an AI-powered chatbot to websites.
  • Features: Model agnostic, integrates with MCPs, and allows AI to retrieve and process data from other apps.
  • Availability: Free and open-source on GitHub.
• GitHub Copilot Coding Agent: An autonomous coding agent that handles complex requests, edits multiple files, and creates pull requests.
  • Availability: Free to use with 50 agent mode or chat requests per month on the free plan.

Synthesis/Conclusion

This week in AI saw a flurry of activity, with new open-source tools, model releases, and significant updates from major players like Google, Microsoft, and Anthropic. Bagel offers impressive multimodal capabilities for image generation and editing, while UniVGR1 demonstrates advancements in visual analysis. Google's IO event showcased powerful AI tools for video generation, medical analysis, and education. Anthropic's Claude 4 aims to excel in coding tasks, though its overall performance and cost-effectiveness are debated. Microsoft's NL Web and GitHub Copilot Coding Agent provide valuable tools for website owners and developers. Overall, the rapid pace of innovation in AI continues, with a focus on automation, accessibility, and specialized applications.