Easy AI Daily | October 29, 2025

A roundup of AI industry news from the Easy AI teaching project, covering model releases, tooling updates, community discussions, API changes, research papers, and events.

Model Updates & Releases

Cursor 2.0 Released with Composer-1 Agent Model and New UI

Cursor 2.0 ships with the Composer-1 coding agent model (4x faster), a new Agents tab for managing multiple agents, a built-in browser, voice-to-code support, and automated code review.

> Links: Cursor announcement tweet | Cursor blog

OpenAI Releases gpt-oss-safeguard Open-Weight Models

OpenAI launched gpt-oss-safeguard in 20B and 120B sizes for policy-based safety classification. Fine-tuned from gpt-oss and licensed under Apache 2.0, the models work with Ollama, LM Studio, and other inference stacks. OpenAI also launched the ROOST community.

> Links: OpenAI announcement | Ollama support

Minimax M2 MoE Model Ranks in Global Top 5

Minimax released M2, a 230B-parameter MoE model with only 10B active parameters, outperforming M1 and Claude Opus 4.1. It excels at long-horizon tool use (shell, browser, MCP, etc.) and integrates with Cursor, Claude Code, and other platforms.

> Link: Minimax announcement

Windsurf Launches Falcon Alpha Model

Windsurf released Falcon Alpha, a high-performance agent model optimized for speed, and is seeking user feedback. Cascade now supports Jupyter Notebooks across all models.

> Link: Windsurf announcement

Modular Mojo MAX Adds HuggingFace and Torchvision Support

Mojo MAX now supports HuggingFace/Torchvision models via torch_max_backend.export_to_max_graph (e.g., VGG11), with GPU execution.

> Link: Modular announcement (Discord)

Tools & Frameworks

• Aider-CE added a Navigator mode and community-contributed RAG (Retrieval Augmented Generation), plus support for GitHub Copilot models (with github_copilot/ prefix) and a directory-switching bug fix.
• DSPy vs. LangChain: A user reports DSPy is better suited for structured tasks (e.g., chat), optimizing prompts and avoiding the issue of prompt patterns breaking after model upgrades; their team has migrated to DSPy.
• GitHub MCP Registry: GitHub plans to integrate MCP Registry so developers can self-publish MCP servers to the OSS registry, auto-syncing to the GitHub MCP Registry (currently 44 servers).

Community Discussions

• Manus.im credit costs: Users complain credits deplete quickly (e.g., 5,600 credits in 3 hours building an Android IRC app) and find Anthropic Claude better value ($20 for 3 projects), asking for free alternatives.
• Petals project: The Petals project (Llama 70B) has lost momentum due to difficulty keeping up with new architectures; LlamaCPP RPC is discussed as the current alternative. The community also discussed the release of Schmidhuber's HGM model code.
• API parameter removal: Users discussed GPT-5 and Anthropic Claude removing temperature/top_p parameters, speculating it prevents probability leakage or reflects reasoning models replacing traditional parameters; some users are unhappy.

API & Service Updates

• OpenAI Sora: Users report improved browser rate limits and generation speed, though some say limits are unchanged and quality is stable.
• GitHub Copilot: The $10/month subscription includes unlimited RAG, gpt-5-mini, gpt4.1, grok-code-1-fast, and free embedding models.
• Anthropic Claude: The API now accepts only one of temperature or top_p, no longer both, affecting developer prompt tuning.

Research & Papers

• KBLaM vs. RAG: Discussion comparing knowledge-compressed KBLaM (embeddings compress knowledge, possibly reducing quality) with traditional RAG (preserves original content without approximation error, but with higher storage costs). See the KBLaM paper.
• HGM by Schmidhuber: Jürgen Schmidhuber's HGM model code and paper are now public; the community is interested in its geometric structure and relation to intelligence. See the HGM GitHub repo and paper.
• Grokking: Discussion of "Towards Understanding Grokking: An Effective Theory of Representation Learning," exploring how models suddenly generalize late in training. See the paper.

Events

• Nextdata Data 3.0 virtual event: CEO Zhamak Dehghani discussed delivering AI-ready data products with agentic co-pilots, unifying multimodal data, and replacing manual orchestration with self-managing data products. Registration
• DSPy meetup: Scheduled for November 18 in San Francisco to discuss framework progress and future direction. Event page

---

*Source: Easy AI teaching project.*