Easy AI Daily | November 19, 2025

A roundup of AI industry news, community discussions, benchmarks, and troubleshooting from the Easy AI teaching project.

New Model Releases and Updates

Meta Releases SAM 3 Unified Image/Video Segmentation Model

Meta launched SAM 3, supporting text and visual prompts for unified image/video segmentation. Claims 2x performance over existing models with 30ms inference time. A Playground testing tool and checkpoints on GitHub/HuggingFace are available, with a Roboflow partnership offering infinitely scalable segmentation endpoints.

Links: Meta SAM 3 official blog | Roboflow partnership announcement | SAM 3 GitHub resources

OpenAI Releases GPT-5.1-Codex-Max

Optimized for long-context scenarios with multi-context-window compression, positioned for long, detailed coding work — seen as a response to recent coding agent competition.

Link: OpenAI announcement

Google Releases Gemini 3 Multimodal Model

Gemini 3 performs strongly on benchmarks, but some users report creative writing is weaker than Gemini 2.5, sparking generalization debates.

Links: Google Gemini 3 overview | Benchmark discussion

Meta Releases SAM 3D

Reconstructs objects and human bodies in 3D from a single image, with GitHub resources and a Roboflow fine-tuning path; early demos show strong text-prompt tracking.

Links: SAM 3D announcement | GitHub

OpenAI Launches ChatGPT for Teachers

For US K-12 education: a safe workspace with admin tools, compliance, and privacy controls to help teachers integrate AI into classrooms.

Link: Announcement

Community Discussions and Controversies

• Ollama pricing controversy: Ollama's new paid cloud service drew criticism from users who see it as abandoning its open-source ethos, favoring llama.cpp's free model. Reddit thread
• Gemini 3 content filtering: Users complain filtering is overly strict — even summarizing books triggered API bans; some say it exceeds ChatGPT's moderation. Twitter
• ChatGPT anomalous behavior: Reports of text-to-image responses, wrong file structures, and destructive commands in code generation. Reddit
• Gemini 3 Pro in Cursor: Unavailable due to high demand, hallucinating code and ignoring prompts in large projects; some users switched to Sonnet 4.5 or Composer.
• git reset --hard safety debate: ChatGPT reportedly executed git reset --hard, losing local code; community calls for tools like Cursor to block dangerous commands. Reddit safety discussion

Tools and Platform Integrations

• Gemini 3 in Windsurf: Integration announced; early glitches fixed — download the latest version. Announcement | Download
• SAM 3 × Roboflow: Infinitely scalable segmentation endpoints via text prompts (e.g., "green umbrella"), comparable with Claude and YOLO World.
• Gemini 3 with Aider: Requires --model=gemini/gemini-3-pro-preview; pairing with --weak-model speeds up commits. Aider docs
• Google Antigravity IDE: A VS Code-based AI IDE supporting Sonnet 4.5 and Gemini 3, with agent windows and an artifact system; early migration bugs and prompt limits reported. Announcement
• Perplexity asset creation: Pro/Max users can create/edit slides, sheets, and documents in-platform with real-time collaboration and export. Announcement

Performance Tests and Benchmarks

• Tinygrad beats Torch on CPU for Llama 1B: 6.06 tok/s vs Torch's 2.92 tok/s, aided by CPU_LLVM acceleration and forward-pass optimizations. Tinygrad GitHub
• KTOtrainer high memory usage: 80GB GPU memory for a 0.5B model — attributed to dual model loading, dual forward passes, long-sequence padding, and CUDA memory reservation; optimization advice provided. HuggingFace discussion
• Gemini 3 benchmarks mixed: SOTA on WeirdML and IBench, but criticized on creative writing versus Gemini 2.5. Reddit
• Kimi K2 Thinking leads agentic coding: Ahead on Tau bench and HLE tool tasks; considered the open-source model closest to GPT-5. Moonshot announcement
• DeepSeek Cogito v2.1: Based on DeepSeek-7B; outperforms Mistral-7B and Gemma-9B on introspection tasks, detecting and reporting injected concepts. Blog

Issues and Troubleshooting

• Manus/TiDB Cloud access: Users can't access Manus-provisioned TiDB Cloud accounts (quota exhausted, no console access); ticloud CLI lacks API keys. TiDB Cloud docs
• GPT-5.1 connection errors in Aider: Errors involving response.reasoning.effort validation persist despite adjusting reasoning-effort (low/medium/high); possibly related to OpenAI API changes. LiteLLM issue
• Manus credit system changes: Shift to monthly resets confused users with older never-expiring plans. Manus site
• Windsurf fixed Gemini 3 loading glitch: Users should update to the latest version.
• Cursor "high demand" error for Gemini 3 Pro: Team suggests Sonnet 4.5 as an alternative.

Education and Resources

• A community member offers an AI coding education course covering core concepts, advanced models, practical applications, and ethics (contact via DM).
• Users asked about certification for HuggingFace's smol-course and agents-course. HuggingFace courses
• Tinygrad benchmark tests comparing CPU performance with Torch are being added to test/external.
• AI security red-teaming resources, including jailbreak prompts and AzureAI widget testing methods. BASI red-team resources
• Ollama open-source alternative: llama.cpp recommended for free local inference. llama.cpp GitHub

---

*Source: Easy AI teaching project.*