Realtime 3D worlds, OpenAI goes open source, ultra fast AI video, Claude Opus 4.1, full storybooks

Key Concepts

• Open Source AI Models (GPT-OOS 120B, GPT-OOS 20B)
• Mixture of Experts (MoE)
• Video Generation (Fast Juan, One 2.1/2.2)
• Guided Learning (Gemini, ChatGPT)
• AI Storybook Generation (Gemini)
• AI-Powered Video Creation (Kivex)
• Robotics (Unitree A2 Stellar Hunter)
• AI Chess Tournament
• Interactive 3D World Generation (Genie 3)
• Image Generation and Editing (Uni Pic)
• AI Model Benchmarking and Leaderboards
• Claude Opus 4.1

OpenAI Goes Open Source: GPT-OOS Models

• Release: OpenAI released two open-weight models: GPT-OOS 120B and GPT-OOS 20B.
• Architecture: Both are Mixture of Experts (MoE) models.
  • GPT-OOS 120B: 117 billion parameters, 128 experts, stronger reasoning, requires H100 GPU.
  • GPT-OOS 20B: 21 billion parameters, 32 experts, optimized for local usage, fits on consumer GPUs.
• Licensing: Apache 2.0 license, minimal restrictions, allows commercial use.
• Capabilities: Chain of thought reasoning, tool use (web search, Python coder, data scraper).
• Benchmarks:
  • Competitive Coding: Close to GPT-3 and GPT-4 Mini.
  • Humanity's Last Exam: Lower than GPT-3 with tool use.
  • Healthbench: Close to GPT-3.
  • AIM (Competitive Math): Performs well (but benchmark is saturated).
  • GPQA (Graduate Level Science Questions): Slightly worse than GPT-3.
• Independent Leaderboard Performance:
  • Artificial Analysis: Ranks below Grok-1, GPT-3 Pro, Gemini 1.5 Flash, and DeepSeek R1. Cost-effective.
  • LiveBench (Abaqus AI): Ranks below DeepSeek V3, Grok-3, Claude 3.7.
  • Hallucination Leaderboard: 20B model has a 3.7% hallucination rate.
• Accessibility: Models available on Hugging Face. Can be tried online with reasoning enabled.
• Example Use Cases:
  • Medical Report Generation: Comprehensive report on congenital heart defects, including surgical options, prognosis, monitoring, and genetic counseling.
  • Crypto Portfolio Dashboard: Generates HTML code for a dashboard with price fluctuations, risk assessments, and a trade simulator (trade simulator doesn't work well).

Fast Juan: Accelerated Video Generation

• Description: A fine-tune of Alibaba's One, an open-source video generator.
• Speed Improvement: Significantly faster than One 2.1/2.2.
• Performance: Generates a 5-second 480p video in 5 seconds on a single H200 GPU (almost real-time). On a 4090, the same video takes 21 seconds.
• Speed Comparison: Fast Juan is 46x faster than One 2.2 (5B parameter model) and 98x faster than One 2.2 (14B parameter model).
• Free Demo: Available online.
• Example Prompts:
  • "A cute fluffy panda bear is preparing a meal in a cozy modern kitchen."
  • "A tranquil scene in the heart of a dense forest emphasizing two large textured tree trunks in the foreground framing the view."
• Technique: Sparse distillation, focusing on the most important parts of the video.
• Availability: GitHub repository with instructions for download and use. Fast Juan model currently available for One 2.2 (5B variant).

Google Gemini: Guided Learning and AI Storybooks

• Guided Learning: A virtual tutor feature that guides users step-by-step through a subject with follow-up questions.
  • Example: Learning about cellular respiration or Japanese.
• AI Storybook Feature: Creates personalized illustrated storybooks with narration.
  • Example: "My seven-year-old doesn't want to sleep over their grandma's house. Create a story book to help them cope."
  • Features: Voiceover (two voice options), consistent illustrations, coherent story.

Kivex: AI-Powered Video Creator (Sponsored)

• Description: An AI-powered video creator for product spotlighting.
• Features:
  • URL to Video: Parses product info from Amazon, eBay, Shopify, or app store links.
  • AI Script Generator: Creates a full script from a topic.
  • AI Video Translator: Converts videos into over 70 languages.
• Offer: 5 minutes of free video credits (no payment required). 25 more minutes for inviting a friend. Full HD 1080p without a watermark.

Unitree A2 Stellar Hunter: Advanced Robot Dog

• Description: A robust robot dog designed for industrial applications.
• Capabilities:
  • Runs through glass.
  • Performs front and back flips.
  • Balances on two legs (side or front) or one hand.
  • Runs fast across uneven terrain.
  • Supports a heavy load (100 kg max).
  • Climbs objects up to 1 meter tall.
• Specs:
  • Weight: 37 kg.
  • Speed: Up to 18 km/h.
  • Battery Life: Up to 5 hours and 20 km.
  • Sensors: Two industrial-grade LiDAR sensors and an HD camera.
  • Temperature Resistance: -20°C to 55°C.

AI Chess Tournament Results

• Format: Best of four games per match.
• Semifinals:
  • GPT-4 Mini vs. GPT-3: GPT-3 won 4-0.
  • Grok-4 vs. Gemini 1.5 Pro: Grok-4 won 3-2.
• Finals:
  • GPT-3 vs. Grok-4: GPT-3 won 4-0.
• Medalists:
  • Gold: GPT-3
  • Silver: Grok-4
  • Bronze: Gemini 1.5 Pro

Google DeepMind Genie 3: Interactive 3D World Generator

• Description: Generates fully interactive 3D environments from text prompts or images.
• Performance: Responds to key presses in real-time at 720p resolution and 24 frames per second.
• Capabilities:
  • Physical understanding.
  • Simulates physics of water and wind.
  • Supports first-person and third-person perspectives.
  • Maintains spatial consistency with a memory window.
  • Allows adding elements and specifying actions with prompts.
• Limitations:
  • Limited to 720p.
  • Limited action space.
  • Cannot model complex interactions between multiple agents.
  • No audio.
  • Cannot simulate real-world locations with perfect accuracy.
  • Not good at generating text in the video.
  • Limited to a few minutes of continuous interaction.
• Accessibility: Currently limited to select researchers.

Skywork Uni Pic: Image Generation and Editing

• Description: A 1.5 billion parameter auto-regressive model for image generation and editing.
• Architecture: Same as GPT-40.
• Capabilities:
  • Text-to-image generation.
  • Semantic image editing (changing art styles, backgrounds, subjects).
• Performance: On par with larger models like Bagel and Omnimodel 2.
• Accessibility: Instructions for local download and use are available.

Anthropic Claude Opus 4.1

• Description: Anthropic's most performant model.
• Coding Benchmarks: Slightly better than Opus 4 on Bench Verified (but GPT-5 scored higher).
• Performance:
  • Coding: Performs well.
  • Other Tasks (Graduate Level Science Questions, Visual Reasoning, Competitive Math): Performs worse than GPT-3 and Gemini 1.5 Pro.
• Accessibility: Only available to paid users.
• Coding Examples:
  • CRM Dashboard: Generates a dashboard with real-time insights, interactive charts, and graphs. GPT-5's dashboard is more interactive and customizable.
  • Photoshop Clone: Claude's clone fails (buttons don't work). GPT-5's clone is functional with basic tools and settings. GLM 4.5 is more functional than Claude but has fewer features than GPT-5.
• Independent Leaderboard Performance:
  • Artificial Analysis: Ranks below Gemini 1.5 Pro, GPT-3, Grok-4, and Quen 3.
  • LiveBench (Abaqus AI): Scores higher than Grok-4 and Gemini 1.5 Pro but is behind GPT-5 and GPT-3.

Conclusion

This week in AI saw a flurry of activity, with major players like OpenAI, Google, and Anthropic releasing new models and features. OpenAI's open-source GPT-OOS models offer accessible AI for various tasks, while Google's Genie 3 showcases impressive interactive 3D world generation. Fast Juan significantly accelerates video generation, and Gemini introduces guided learning and AI storybook creation. While Anthropic's Claude Opus 4.1 shows marginal improvements in coding, it lags behind other models in broader tasks. The AI landscape continues to evolve rapidly, with each week bringing new advancements and tools for developers and users alike.