Easy AI Daily News Roundup | November 19, 2025

Forum topic · 小凯 · 2026-03-27

Summary

This November 19, 2025 daily briefing from zhichai.net's Easy AI project covers major AI industry developments. Meta released SAM 3, a unified image/video segmentation model with 2x performance gains and 30ms inference, plus SAM 3D for single-image 3D reconstruction. OpenAI launched GPT-5.1-Codex-Max for long-context coding agents and ChatGPT for Teachers for US K-12 education. Google shipped Gemini 3, which tops benchmarks like WeirdML and IBench but faces criticism over strict content filtering and weaker creative writing versus Gemini 2.5. Community discussions include backlash against Ollama's paid cloud service, ChatGPT executing destructive commands like git reset --hard, and Gemini 3 Pro availability issues in Cursor. Tooling news covers Gemini 3 integration into Windsurf, Aider, and Google's new Antigravity IDE, plus Perplexity's asset creation features. Benchmarks show Tinygrad outperforming Torch on CPU for Llama 1B (6.06 vs 2.92 tok/s), and Kimi K2 Thinking leading agentic coding tasks.

Key points

New Model Releases

Meta SAM 3: Unified image/video segmentation model supporting text and visual prompts. 2x performance over existing models, 30ms inference time. Playground tool and checkpoints available on GitHub/HuggingFace; Roboflow partnership offers infinitely scalable segmentation endpoints.
Meta SAM 3 blog | Roboflow announcement
OpenAI GPT-5.1-Codex-Max: Optimized for long-context scenarios with multi-context-window compression; positioned for long, detailed coding work and seen as a response to coding-agent competition.
OpenAI announcement
Google Gemini 3: Strong benchmark results, though some users report creative writing is worse than Gemini 2.5, sparking debate about generalization.
Google AI docs | Benchmark discussion
Meta SAM 3D: Reconstructs 3D objects and human bodies from a single image; GitHub resources and Roboflow fine-tuning path available.
Publication | GitHub
ChatGPT for Teachers: OpenAI's offering for US K-12 educators with a safe workspace, admin tools, compliance and privacy controls.
Announcement

Community Discussions & Controversies

Ollama's paid cloud service draws criticism from users who contrast it with llama.cpp's free open-source model, questioning its open-source commitment. (Reddit)
Gemini 3 content filtering reported as overly strict — some users hit API bans even when summarizing books. (Twitter)
ChatGPT anomalies: users report text-to-image conversion, wrong file structures, and destructive commands like git reset --hard, raising safety concerns. (Reddit)
Gemini 3 Pro in Cursor: unavailable due to high demand; reports of hallucinated code in large projects; some users switched to Sonnet 4.5 or Composer.

Tooling & Platform Integrations

Windsurf integrated Gemini 3 Pro; early glitches fixed. (Download)
Aider + Gemini 3: use --model=gemini/gemini-3-pro-preview, with --weak-model recommended for faster commits. (Docs)
Google Antigravity: new VS Code-based AI IDE supporting Sonnet 4.5 and Gemini 3, with agent windows and an artifact system; early reports of migration bugs. (Announcement)
Perplexity: Pro/Max users can now create/edit slides, spreadsheets, and documents with real-time collaboration and export. (Announcement)

Benchmarks & Performance

Tinygrad vs Torch on CPU (Llama 1B): 6.06 tok/s vs 2.92 tok/s, aided by CPU_LLVM acceleration and forward-pass optimizations. (GitHub)
KTOtrainer memory issue: a 0.5B model consumed 80GB GPU memory due to dual model loading, dual forward passes, long-sequence padding, and CUDA memory reservation; optimization advice shared. (Discussion)
Kimi K2 Thinking: leads Tau bench and HLE tool-use tasks; considered the open-source model closest to GPT-5. (Moonshot)
DeepSeek Cogito v2.1: based on DeepSeek-7B, outperforms Mistral-7B and Gemma-9B on introspection tasks, detecting and reporting injected concepts. (Blog)

Issues & Troubleshooting

GPT-5.1 in Aider: API connection errors involving response.reasoning.effort validation; adjusting reasoning-effort levels did not resolve them. (Litellm issue)
Manus: users report TiDB Cloud access failures (quota exhaustion, no console access) and confusion over the new monthly credit reset policy.
Windsurf/Cursor: Gemini 3 loading and high-demand errors, with fixes or fallback suggestions (Sonnet 4.5).

Education & Resources

Community AI coding courses covering core concepts, advanced models, applications, and ethics.
HuggingFace learning resources: smol-course and agents-course.
Red-teaming/jailbreak training resources shared by the BASI community: UltraBr3aks.
llama.cpp recommended as a free, open-source alternative to Ollama: GitHub.

*Source: Easy AI education project*