Easy AI Daily Digest | December 5, 2025
A roundup of AI industry news for December 5, 2025, curated by the Easy AI teaching project.
Model Releases and Updates
Google releases Gemini 3 Deep Think mode
Available to Google AI Ultra subscribers, Deep Think improves complex reasoning using parallel thinking. It scores 45.1% on ARC-AGI-2, far above GPT-5.1's 17.6%, and supports math and science tasks.- Google AI announcement
- Google DeepMind details
- OpenAI Devs announcement
- Cursor integration
- Hugging Face model page
- Nous Research blog
- Mistral AI announcement
- Google Titans memory architecture: Combines RNN efficiency with Transformer performance, supporting contexts of 2M+ tokens; early results shown at NeurIPS. Google Research
- TorchAO MoE quantization: New MoEQuantConfig enables quantization of mixture-of-experts models for faster inference. PyTorch PR
- VATTENTION paper: First sparse attention mechanism with (ϵ, δ) guarantees, improving long-text processing. arXiv
- STRAW: Sample-tuned rank-augmented weights that mimic neuromodulation, dynamically adjusting model weights for better task adaptability. Substack
- Fast ODE solver for diffusion models: Generates 4K images in 8 steps, matching the quality of 30-step DPM++2M SDE. Hugging Face Space | arXiv
- Anthropic acquires Bun as Claude Code reaches a $1 billion annual revenue milestone. Anthropic news
- Perplexity receives investment from Cristiano Ronaldo, framed as "inspiring curiosity worldwide." Announcement
- Harvey raises $160M Series F at an $8 billion valuation, serving 700+ law firms with legal AI. Source
- Antithesis raises $105M led by Jane Street, focused on deterministic simulation testing of AI-generated code. Source
- GPT-5.1 vs Gemini 3 for bug finding: OpenAI Discord users report GPT-5.1 finds code bugs that Gemini 3 misses.
- Z-Image content filtering: Reddit users report the model still filters gore/nudity despite claims of being uncensored, showing a "maybe not safe" flag. Reddit
- AI and tech jobs: Redditors debate whether AI will reduce software roles, with consensus leaning toward role transformation rather than replacement. Reddit
- VibeVoice-Realtime feedback: LocalLlama users test the model's English/Chinese support; some note Mandarin accent issues. Reddit
- Gemini 3 Deep Think benchmarks: Extensive discussion of its 45.1% ARC-AGI-2 score vs GPT-5.1's 17.6%. Reddit
- OpenRouter State of AI report: Analyzed 100 trillion tokens; open-source models are ~50% roleplay, paid models ~50% coding, with Claude accounting for 60% of coding workloads. Report
- Windsurf integrates GPT-5.1-Codex Max: Free trial for paid users with Low/Medium/High reasoning levels. Announcement
- mcp-apps-sdk open-sourced by General Intelligence Labs, enabling ChatGPT apps to be embedded in other platforms. GitHub
- tinygrad fix: PR fixes train_step not utilizing input tensors, improving training efficiency. GitHub PR
- DSPy + Claude Code proposal: A community suggestion for native Claude Code support in DSPy using its Read/Write/Terminal tools.
- Gemini 3 Deep Think: 45.1% on ARC-AGI-2, roughly 2.5x GPT-5.1's 17.6%, excelling on complex reasoning tasks.
- Mistral Large 3: #1 on lmarena's coding leaderboard among open-source models.
- DeepSeek V3.2: Baseten serving metrics show TTFT of 0.22s and 191 tokens/sec; strong lmarena rankings in math, law, and science. Baseten
- GPT-5.1-Codex Max: Users report improved code quality and efficiency after Cursor integration.
OpenAI launches GPT-5.1-Codex Max
Available through the Responses API and integrated into the Codex agent harness, with support for IDEs such as VS Code and Cursor.Microsoft releases VibeVoice-Realtime-0.5B
A lightweight, open-source realtime text-to-speech model supporting English and Chinese, hosted on Hugging Face.Nous Research releases Hermes 4.3
Built on ByteDance Seed 36B with performance approaching Hermes 4 70B, trained via the Psyche network with MoE support.Mistral Large 3 tops open-source coding leaderboard
Ranked #1 on lmarena and available via Ollama Cloud; the community has confirmed its coding strength.Technical Progress and Research
Industry News and Funding
Community Discussions
Tools and Platform Updates
Benchmarks and Performance
*Source: Easy AI teaching project.*