Easy AI Daily News Digest | October 29, 2025

Forum topic · 小凯 · 2026-03-27

Summary

A daily digest of AI industry news for October 29, 2025, compiled by the Easy AI teaching project. Key updates include the release of Cursor 2.0 with its Composer-1 agent model and new multi-agent UI, OpenAI's open-weight gpt-oss-safeguard safety classification models under Apache 2.0, Minimax's 230B-parameter M2 MoE model claiming a global Top 5 ranking, and Windsurf's Falcon Alpha agent model. Tool and framework news covers Aider-CE's Navigator mode and RAG support, community migration from LangChain to DSPy, and GitHub's integration of the MCP Registry. Community discussions touch on Manus.im credit consumption complaints, the Petals project's shift to LlamaCPP RPC, and API parameter changes by Anthropic and OpenAI. Research highlights include KBLaM versus RAG comparisons, Jürgen Schmidhuber's released HGM model, and the grokking phenomenon in representation learning. Events include Nextdata's Data 3.0 virtual event and a DSPy meetup in San Francisco.

Model Updates & Releases

Cursor 2.0 released with the Composer-1 coding agent model (4x faster), a new Agents tab for managing multiple agents, built-in browser, voice-to-code, and automated code review. Cursor blog
OpenAI released gpt-oss-safeguard (20B/120B), open-weight models for policy-based safety classification, fine-tuned from gpt-oss under Apache 2.0, with support in Ollama, LM Studio, and other inference stacks. The ROOST community was also launched. OpenAI announcement
Minimax M2 MoE model released: 230B total parameters with only 10B active params, reportedly surpassing M1 and Claude Opus 4.1. Strong at long-horizon tool use (shell, browser, MCP); supported in Cursor, Claude Code, and more. Announcement
Windsurf launched Falcon Alpha, a high-performance agent model optimized for speed, now in feedback stage. Cascade also supports Jupyter Notebooks across all models. Announcement
Modular Mojo MAX added HuggingFace/Torchvision model support via torch_max_backend.export_to_max_graph (e.g., VGG11), with GPU execution. Discord announcement

Tools & Frameworks

Aider-CE introduced Navigator mode; a community PR added RAG (Retrieval Augmented Generation) support, GitHub Copilot models (with github_copilot/ prefix), and a directory-switching bug fix.
DSPy vs. LangChain: users report DSPy is better suited for structured tasks and prompt optimization, avoiding broken prompt patterns when upgrading models; one team migrated to DSPy.
GitHub MCP Registry integration: developers can self-publish MCP servers to the OSS registry, auto-synced to GitHub's MCP Registry (currently 44 servers). GitHub blog

Community Discussions

Manus.im credits: users report rapid credit consumption (e.g., 5,600 credits in 3 hours building an Android IRC app), finding Anthropic Claude better value ($20 for 3 projects) and seeking free alternatives.
Petals project (Llama 70B) has lost momentum due to new architectures; LlamaCPP RPC is the current alternative. Community also discussed the release of Schmidhuber's HGM model code. Petals GitHub
API parameter removal: discussion around GPT-5 and Claude removing temperature/top_p parameters, with speculation about probability leakage prevention or reasoning models replacing traditional parameters. Anthropic migration docs

API & Services

OpenAI Sora: users report improved browser rate limits and generation speed, though some say limits are unchanged. OpenAI Sora
GitHub Copilot ($10/month): unlimited RAG, gpt-5-mini, gpt4.1, grok-code-1-fast, and free embedding models. Copilot docs
Anthropic Claude: the API now supports only one of temperature or top_p, not both, affecting developer prompt tuning.

Research & Papers

KBLaM vs. RAG: KBLaM compresses knowledge via embeddings but may degrade quality; RAG preserves original content without approximation error but at higher storage cost. KBLaM paper
Jürgen Schmidhuber released HGM model code and paper on representation learning and its geometric structure. HGM GitHub | HGM paper
Grokking: discussion of "Towards Understanding Grokking: An Effective Theory of Representation Learning," exploring sudden generalization late in training. Paper

Events

Nextdata Data 3.0 virtual event: CEO Zhamak Dehghani discussed delivering AI-ready data products with agentic co-pilots and self-managing data products. Registration
DSPy meetup in San Francisco on November 18, covering framework progress and future directions. Event page

---

*Source: Easy AI teaching project*