Easy AI Daily | December 12, 2025: GPT-5.2 Launch, Disney's $1B OpenAI Deal, and Open-Source AI Tooling Updates

📅 AI Industry Roundup — December 12, 2025

A translated digest of the Easy AI Daily from zhichai.net covering model releases, industry deals, open-source tools, benchmarks, and community discussions.

Model Updates & Releases

GPT-5.2 Released: Performance Gains at Higher Prices

OpenAI released GPT-5.2 with improvements in scientific reasoning (92.4% accuracy), competition math (100%), and long-context processing. Pricing increased to $1.75 per million input tokens and $14 per million output tokens, with a 90% discount on cached input. It ranks second on the WebDev Code Arena, though it underperforms on some coding benchmarks.

• OpenAI official blog
• System card (PDF)
• Documentation

Mistral Teases New Model

Mistral AI teased an upcoming model on X. Community speculation suggests it may appear on OpenRouter; users await benchmark results.
• Mistral on X

Qwen 3 Sparse Series Called "Underrated"

Users recommend the Qwen 3 sparse series (e.g., a3b) for strong coding and reasoning, though some report the Qwen 32b model performs only mediocre.
• OpenRouter Discord

Industry Partnerships & Investment

Disney Invests $1B in OpenAI, Brings Characters to Sora

Disney is investing $1 billion in a partnership with OpenAI to integrate its characters into the Sora AI video generator. The agreement includes a 3-year license with exclusivity in the first year, with content to appear on Disney+.
• OpenAI announcement
• CNBC coverage

DeepMind Opens First Automated Research Lab in the UK

In partnership with the UK government, DeepMind is opening its first automated research lab focused on AI-driven scientific discovery (e.g., materials science, drug development), scheduled to launch in 2026.
• DeepMind blog

Open-Source Tools & Technology

Unsloth's New Packing Method: 3x Faster Training

Unsloth's new packing technique trains 3x faster than the previous version and 10x faster than FA3, supports Qwen3-4B training on 3.9GB of VRAM, and resolves dependency conflicts with older NVIDIA drivers.
• Unsloth docs

llama.cpp Adds Live Model Switching

llama.cpp introduces a routing mode enabling dynamic model management (load, unload, and switch without restarts). A multi-process architecture isolates crashes for stability, with LRU caching and automatic discovery.
• Hugging Face blog

Hugging Face WebGPU Local Voice Chat Demo

A Hugging Face Space showcases real-time AI voice chat running entirely in the browser (STT, VAD, TTS, and LLM all processed locally), preserving user privacy.
• Hugging Face Space

Benchmarks & Performance

GPT-5.2 Beats Human Experts on GDPVal Tasks

GPT-5.2 Thinking outperformed 70.9% of human experts on GDPVal tasks across 44 occupations, working 11x faster at about 1% of the cost — though human oversight is still recommended.
• OpenAI GDPVal notes
• SWE-Bench results

Community & Ecosystem

Reddit Debates GPT-5.2 Performance vs. Hype

Reddit users praised GPT-5.2's 100% competition math score but criticized the $168/M output token cost. Memes mocked its "AGI" claims — triggered by miscounting the letter R in "garlic."
• Reddit thread
• AGI meme

LMArena Community Tests GPT-5.2 Coding

LMArena users report GPT-5.2 High generating buggy code in Code Arena despite high SWE-bench scores. It ranks second on the WebDev leaderboard, but users call it a "rushed release" with steep pricing.
• LMArena leaderboard
• Discord chat

Hardware & Infrastructure

CUDA 13 Fixes Torch/vLLM Compatibility Issues

Switching to CUDA 13 resolves compatibility issues between Torch and vLLM — both must use CUDA 13 builds, particularly relevant for AMD GPU users.
• GPU MODE Discord

Hetzner Launches 96GB VRAM Server at €889

Hetzner released a bare-metal server with 96GB of VRAM at €889/month, including generous free traffic — attractive for AI startups cutting training/inference costs.
• Nous Research Discord

Research & Theory

Diffusion Distillation Technique Yields Free Log Probabilities

A new diffusion technique adds a predictive-divergence head and adjusts initial noise to obtain free log probabilities, improving image likelihood maximization.
• arXiv paper

Sandwich Normalization for Long-Context Transformers

Researchers discuss "sandwich normalization" for handling longer sequences in Transformers by normalizing activations; the paper details the method.
• OpenReview paper

AI Ethics & Jailbreaks

CIRIS Agent Tested for Jailbreak Resistance

CIRIS Agent, designed as an ethical AI, invited users to bypass its filters. It refused to generate unethical content (e.g., meth synthesis instructions), though some users pushed its limits.
• BASI Jailbreaking Discord

Grok Image Generation Sparks Censorship Debate

Users debate Grok's image-generation censorship — some say restrictions are strict, others note skilled users can still produce deepfakes, with some outputs described as "unaligned garbage."
• BASI Jailbreaking chat

Developer Tools & Platforms

Cursor Debug Mode Gets Positive Feedback

Cursor's new debug mode solves issues by adding test objects, with users reporting successful debugging. However, context rollback cannot restore state; users want backup features.
• Cursor Discord

Perplexity Pro Users Hit Strict Rate Limits

Perplexity Pro users report being rate-limited after just 5 Gemini 3 Pro queries. Suspected causes include server load or bugs; workarounds include disabling VPNs and clearing cache.
• Perplexity Discord

Windsurf Ships New MCP Management UI

Windsurf released versions 1.12.41 and 1.12.160 with improved stability and performance, a new MCP management UI, fixes for GitHub/GitLab MCP issues, and enhanced diff zones and Supercomplete.
• Windsurf changelog

---

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant