AI Briefing — June 3, 2026

🔬 Research

Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding

125 upvotes · Jianuo Huang et al. · Paper

Domino tackles the core trade-off in speculative decoding: autoregressive drafters model dependencies well but are slow, while parallel drafters are fast but miss token relationships. The framework uses a parallel backbone to generate preliminary drafts, then applies a lightweight "Domino head" to add causal corrections. Experiments on Qwen3 show up to 5.49× speedup with Transformers and 5.8× with SGLang serving.

Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

34 upvotes · Pengcheng Jiang et al. · Paper

This 20B search agent separates semantic decisions (what to search, what to keep) from routine bookkeeping by moving state management to the environment. The "harness" maintains working memory, candidate pools, and verification records while the policy focuses on search strategy. Achieves 0.730 average curated recall across eight retrieval benchmarks, outperforming the next best open agent by +11.4 points.

Trust Region On-Policy Distillation

28 upvotes · Xingrun Xing et al. · Paper

TrOPD addresses the instability in on-policy distillation when teacher and student distributions diverge significantly. The approach restricts distillation to "trust regions" where teacher supervision is reliable, uses outlier estimation to handle problematic areas, and adds off-policy guidance to encourage exploration toward reliable regions. Shows consistent improvements over existing OPD methods across mathematical reasoning, code generation, and general benchmarks.

🛠 Tools

nesquena/hermes-webui — Web interface for Hermes Agent

1,722 stars/day · Repo

A lightweight web UI for Nous Research's Hermes Agent, offering full parity with the CLI experience in a three-panel layout. Features session management, workspace file browsing, and a circular context ring for token usage visualization. Built with vanilla JS and Python with no build step required, making it accessible for quick deployment and mobile use.

affaan-m/ECC — Agent harness performance optimization system

1,533 stars/day · Repo

A comprehensive agent harness system targeting Claude Code, Codex, and Cursor with 204K+ total stars. Provides skills management, instincts, memory systems, and security features for agent development. The project emphasizes research-first development and has attracted significant community attention with 170+ contributors and broad multi-language support.

chopratejas/headroom — Token compression for LLM inputs

1,265 stars/day · Repo

Compresses tool outputs, logs, files, and RAG chunks before they reach the LLM, achieving 60-95% token reduction while maintaining answer quality. Available as a library, proxy, and MCP server, addressing the growing cost of token consumption in production LLM applications. Particularly valuable for applications with large context windows or frequent tool interactions.

📡 Analysis

Microsoft's new MAI models

June 2, 2026 · Post · tags: llm-release, generative-ai, ai, microsoft

Microsoft released MAI-Thinking-1 (1T parameters, 35B active) for reasoning and MAI-Code-1-Flash (137B total, 5B active) for GitHub Copilot. Both claim training on "clean and appropriately licensed data" without third-party distillation. However, the technical paper reveals standard web crawling practices, including Common Crawl and proprietary web scraping of 1.2 trillion pages, undermining the licensing claims.

datasette-agent-micropython 0.1a0

June 2, 2026 · Post · tags: python, sandboxing, datasette, webassembly

Alpha release enabling Datasette Agent to execute Python code safely through WebAssembly sandboxing. Part of Simon's broader effort to add secure code execution capabilities to Datasette Agent, leveraging MicroPython's WebAssembly compilation for isolated execution environments.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net