Easy AI Daily News Digest | November 14, 2025
A curated roundup of AI industry news from the Easy AI teaching project community.
Model Releases & Updates
OpenAI releases GPT-5.1 with adaptive reasoning and tool capabilities
OpenAI launched GPT-5.1, featuring adaptive reasoning (allocating more tokens to complex tasks), newapply_patch (code editing) and shell tools, and prompt caching extended to 24 hours. Pricing is unchanged from GPT-5.0, and the model is already integrated into GitHub Copilot, VS Code, and other tools.> Links: OpenAI blog | GitHub Copilot announcement
DeepMind releases SIMA-2, agents that learn autonomously in virtual 3D worlds
DeepMind introduced SIMA-2, a Gemini-based agent that can play games and reason in virtual 3D worlds, self-improving through trial and error. It supports worlds generated by Genie 3 without requiring additional human data.> Link: DeepMind blog
Google quietly rolls out Gemini 3 to Pro subscribers
Google is rolling out Gemini 3 to Pro users via Gemini Drops, with a focus on developer tools and enterprise applications. Free users will receive the update in the coming weeks.> Link: Google blog
Zhipu AI releases GLM-4.6, hosted by Together AI
Zhipu released GLM-4.6, and Together AI has put it into production, claiming performance close to Claude Sonnet 4 with 15% fewer tokens used.> Link: Together AI announcement
API & Tooling Ecosystem
- OpenRouter deprecates Polaris Alpha, ships GPT-5.1 variants: OpenRouter deprecated the early GPT-5.1 version Polaris Alpha and launched GPT-5.1 Chat, Codex, and Codex-Mini. Privacy settings affect model availability, and some users hit API rate limits. (announcement | status page)
- Aider-ce integrates DeepSeek API: The integration improves agent-mode performance and addresses slow GPT-5-high responses. Users praise its features but worry about maintainer communication. (GitHub | DeepSeek)
- Propercode v0.1 released: A multi-agent coding CLI built on Pydantic AI, aiming to improve coding reliability, with autonomous and learning-guide modes. (GitHub)
- DeepSeek OCR API tool open-sourced: Deployable custom OCR using Unsloth inference, supporting image extraction from URLs or base64. (GitHub)
- Eleuther cleans Lakh MIDI dataset: Produced 44k structured JSON entries; planned upload to Hugging Face with community collaboration invited.
- Cleaned French Wikipedia dataset uploaded: 2.7 million files with templates, tables, infoboxes, and links preserved; an English version is in progress. (Hugging Face)
- RLinf Basic repo incoming: Qwen3-VL VLA-adapter training has started, with repo cleanup and LIBERO dataset evaluation planned. (GitHub)
- Anthropic plans $50 billion US data center investment: Facilities in Texas and New York will create jobs, but debate continues over scale, staffing, and environmental impact. (announcement)
- Parag Agrawal's new company raises $100M Series A: Parallel Web Systems will use the funding to build an AI-native web. (announcement)
- YC-backed Kimi For Coding criticized over quota limits: Its 2048 weekly quota was called "daylight robbery" by users questioning its value. (tweet)
- LMArena Discord: Mixed reactions to GPT-5.1—some report improved reasoning, others find it lacking creativity. Gemini 3's release may slip to next week.
- Perplexity AI Pro: Users report image generation limits, inaccurate lab token resets, and bans tied to the referral program.
- Unsloth AI Discord: Members advise LoRA alpha should be half the rank to avoid gradient explosions; discussion of RTX 5000/6000 price-performance and a community GPU pooling proposal. (docs)
- HipKittens paper flags Mojo MHA performance issues: Mojo's MHA kernel reaches only ~50% of peak performance on MI355X due to bank conflicts; optimized memory layouts are recommended. (arXiv:2511.08083)
- Memorization entanglement disentanglement paper: Proposes a method to separate memorized information within model weights, aiding interpretability of model decisions. (arXiv:2510.24256)
- NVIDIA NVFP4 GEMV optimization hackathon: Co-hosted with GPU MODE; CuTeDSL recommended, with B200 GPU access provided. (details)
- Modular Mojo plans WebGPU and FFI C support: Users discuss C FFI integration possibilities; currently relies on Python interop conventions. (Mojo docs)
- tinygrad updates: Added OpenCL device checks and VGG16 model loading; pmap/vmap features coming soon. (GitHub)
- Cursor Auto Mode slows after GPT-5.1 integration: Users report slowdowns and inability to edit files after GPT-5.1-Codex integration, suspecting server overload. (forum)
- Manus.im checkpoint issues: Checkpoint system fails to find git commits; users must contact support for manual sync. (feedback)
- HuggingChat paywall disappointment: Users complain free features are now restricted compared to the previous unlimited free version. (HuggingChat)
Open-Source Projects & Datasets
Industry News & Funding
Community Discussions
Research & Papers
Hardware & Infrastructure
User Experience & Issues
*Source: Easy AI teaching project*