GPT-5 Cracks Math Conjectures & Deception Tests as Alibaba Eyes 10-Trillion-Parameter AI

AI Models & Research

GPT-5 Advancements: Solving Math Conjectures and Social Intelligence

• GPT-5 demonstrates progress in solving minor open math problems (typically requiring PhD-level effort) and shows "flashes of originality" in reasoning, though it struggles with cross-paper synthesis. A study titled "Gödel Test: Can Large Language Models Solve Easy Conjectures?" evaluates its performance on five conjectures.
  • GPT-5 may represent the beginning of progress toward models capable of passing the Gödel Test
• In a separate study, GPT-5 outperformed other AI models in Among Us, showcasing advanced deception, persuasion, and theory-of-mind skills.
  • Researchers made AIs play Among Us to test their skills at deception, persuasion, and theory of mind. GPT-5 won.
    • Study: AmongAIs

OpenAI’s New "Alpha Models" for Pro Users

• OpenAI introduced exclusive "Alpha Models" for Pro subscribers, offering enhanced performance and advanced features. Details on capabilities remain limited.
  • New "Alpha Models" for Pro users

GPT-5-Codex: Enhanced Coding Reasoning

• Users report GPT-5-Codex excels in coding tasks, with improved reasoning and error recovery, making it a standout tool for developers.
  • GPT-5-Codex seems to be on fire!

Alibaba’s Ambitious Qwen Roadmap

• Alibaba unveiled plans to scale Qwen models aggressively, including:
  • Context length: 1M → 100M tokens
  • Parameters: Trillion → 10 trillion
  • Focus areas: Synthetic data generation, agent capabilities, and test-time compute scaling.
  • Alibaba just unveiled their Qwen roadmap. The ambition is staggering!

AI Hardware & Infrastructure

China’s Fenghua No.3 GPU: CUDA-Compatible with 112GB+ HBM

• China’s latest GPU, Fenghua No.3, claims CUDA compatibility, ray tracing (RT) support, and 112GB+ HBM memory, positioning it as a competitor in AI hardware.
  • China's latest GPU arrives with claims of CUDA compatibility and RT support
    • Tom’s Hardware: Fenghua No.3 Details

AI Agents & Local Deployment

Fully Local AI Agent for Raspberry Pi 5

• A developer built a lightweight, fully local AI agent for Raspberry Pi 5 using small models (Qwen3:1.7B, Gemma3:1B). The agent handles wake-word detection, transcription, and LLM inference on-device.
  • I built a tiny fully local AI agent for a Raspberry Pi
    • Blog: Agent Desktoy
    • GitHub: MaxHeadbox

AI in Visual Reasoning

Google’s Veo 3: Chain-of-Frames for Visual Tasks

• Google’s Veo 3 introduces Chain-of-Frames, a visual analog to Chain-of-Thought, enabling diffusion models to tackle complex visual reasoning tasks (e.g., Arc AGI, Clockbench) without relying on multimodal LLMs.
  • Google's Veo 3 Demonstrates Chain-of-Frames behavior
    • Video Zero-Shot Project

Enterprise & Sovereign AI

Microsoft, OpenAI, and SAP Partner for Germany’s Public Sector AI

• A collaboration between Microsoft, OpenAI, and SAP will deploy sovereign AI in Germany’s public sector by 2026, using Microsoft Azure and SAP’s Delos Cloud to ensure data privacy and compliance.
  • Microsoft, OpenAI and SAP are bringing sovereign AI to Germany’s public sector
    • OpenAI: Germany Initiative

Anthropic’s Claude Integrated into Microsoft 365 Copilot

• Claude (Anthropic) is now available within Microsoft 365 Copilot, enhancing the platform’s AI capabilities.
  • Anthropic and Microsoft deal was successful