SM1: Mamba1's d_state=1 Closed-Form Solution Runs on Blackwell

An interesting variant of the Mamba1 architecture, "SM1 (Scalar Mamba1)", was announced on Reddit's MachineLearning community. The idea: by specializing for the d_state=1 condition, the entire selective scan can be replaced with just two native PyTorch operations.

Specifically, a closed-form solution combining `torch.cumprod` and `torch.cumsum` reportedly produces results completely identical to the sequential computation in floating-point precision. This is not an approximation but an exact solution under the boundary condition of d_state=1. For d_state=2 and above, the closed form no longer holds.

The practical benefits are substantial as well. Mamba1's scan intermediate representation is 4-dimensional (B, T, F, S), but SM1 eliminates the S dimension entirely. Compared to Mamba1 with d_state=16, scan memory is reduced to 1/16. A 130M-parameter model's inference state fits in about 14,080 floats (56KB), achieving O(1) per-token cost with no KV cache required.

Training a 130M-parameter model is feasible on an RTX 5060 Ti (16GB), and the author is training on 163K MIDI files (roughly 2.5 billion tokens). The claim that a scalar d_state is sufficient as long as structure is encoded through tokenization is intriguing.

Note, however, that this is a community-driven experimental implementation, not an official Google recommendation.

Spice: An OSS Layer Governing AI Agents' Decision-Making

An open-source project called "Spice" aims to build a decision layer that sits on top of AI agents' execution layer.

Execution agents like Claude Code, Codex, and Hermes excel at "doing things", but the judgment of "what to do" and "when to do it" currently depends on the user's prompts. Spice is an attempt to bring structure to this decision-making piece.

The core loop is "perception → state model → simulation → decision → execution → reflection": understand the context, simulate multiple futures, make structured decisions, and delegate tasks to the appropriate agent. Auditability and traceability are also touted.

Rather than replacing existing agents, it is designed as the "brain" that sits above them. It's an interesting idea, but the project is still in its early stages, and its real-world effectiveness remains to be verified.

Claude Code's /workflows Feature Removed the Day After Appearing in the Changelog

An interesting verification article was published on Qiita. The changelog for Claude Code v2.1.147 (released May 20, 2026) included the addition of a new feature called "Workflow tool". It was a feature for controlling sequential chains of multiple agents with deterministic code rather than LLM judgment, but reportedly, the very next day a GitHub Actions commit removed the entry.

As a case of divergence between official documentation and the actual implementation, it is sparking discussion among developers. A feature once listed in a changelog being removed without any announcement is a concern for developers who value the stability of APIs and tools.

This incident can be seen as one example of the "documentation reliability" problem brought about by the rapid iteration of AI development tools.

A Practical Benchmark Visualizing RAG Chunking Quality

On Reddit, a comparative experiment on RAG chunking and embeddings using content from three production sites (Intercom, HubSpot, and KPMG) was shared.

The content from each site was stratified (HIGH/MEDIUM/LOW/REJECTED) and retrieval quality was evaluated, showing that RAG effectiveness varies dramatically depending on content density. 96% of Intercom's HIGH chunks were help center articles, HubSpot's HIGH chunks were concrete case studies, and KPMG's content was positioning prose across the board, with almost no HIGH chunks at all.

What's interesting is the "yield score" metric — the ratio of HIGH+MEDIUM chunks — and the finding that it can be used to predict corpus quality before RAG generation. Intercom scored 31%, HubSpot 32%, and KPMG 8%; the lower the number, the more cautious phrasing and robust fallbacks a brand needs.

RAG benchmarks tend to assume homogeneous source quality, but this is a practical study demonstrating just how heterogeneous real-world content is.