AI Briefing — June 9, 2026
🔬 Research
SWE-Explore: Benchmarking How Coding Agents Explore Repositories
90 upvotes · Zhang et al. · Paper
SWE-Explore introduces a granular benchmark for evaluating how coding agents explore codebases, moving beyond binary success/failure to measure specific capabilities like code localization and context retrieval. The benchmark covers 848 issues across 10 languages and 203 repositories, with ground truth derived from successful agent solution paths. Early results show agentic explorers clearly outperform classical retrieval methods, though line-level coverage and efficient ranking remain key differentiators among state-of-the-art systems.
ResearchClawBench: A Benchmark for End-to-End Autonomous Scientific Research
79 upvotes · Xu et al. · Paper
ResearchClawBench tests AI agents on complete scientific research workflows across 40 tasks spanning 10 domains, with each task grounded in real published papers and requiring full paper-level rediscovery. Current systems perform poorly: the strongest autonomous agent (Claude Code) achieves only 21.5% average performance, with failures concentrated in experimental protocol mismatches and missing scientific reasoning. The benchmark reveals a significant gap between today's coding agents and true autonomous research capability.
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
66 upvotes · Bigverdi et al. · Paper
Imaginative Perception Tokens (IPT) help vision-language models reason about unobserved spatial information by generating intermediate perceptual representations of alternative viewpoints or hidden spaces. Tested on 20K examples across perspective taking, path tracing, and multiview counting tasks, IPT supervision consistently outperforms textual chain-of-thought training and improves accuracy by 3.4% on multiview counting. The approach suggests that forcing spatial reasoning through language creates a modality mismatch, while visual intermediate representations provide more natural supervision.
🛠 Tools
mvanhorn/last30days-skill — AI agent skill that researches topics across social platforms
3558 stars/day · Repo
This tool aggregates real-time sentiment from Reddit upvotes, X likes, YouTube transcripts, Polymarket odds, and other social platforms to research any topic, then synthesizes findings into grounded summaries. Unlike traditional search engines that aggregate editor opinions, it searches based on what real people engage with and bet money on. The skill works across 50+ AI agent platforms including Claude, Cursor, and Gemini CLI with zero configuration required for most sources.
RyanCodrai/turbovec — Vector index built on Google's TurboQuant algorithm
1729 stars/day · Repo
TurboVec compresses 10 million document corpora from 31GB to 4GB while searching faster than FAISS, using Google's TurboQuant algorithm that matches Shannon's theoretical distortion limits. The Rust implementation with Python bindings supports online vector ingestion without rebuilding, filter-aware search kernels, and hand-optimized NEON/AVX-512 instructions. It targets privacy-sensitive RAG applications where keeping embeddings local is critical.
roboflow/supervision — Reusable computer vision tools
1288 stars/day · Repo
Supervision provides a comprehensive toolkit for computer vision workflows, offering utilities for dataset manipulation, model evaluation, and annotation management. The library aims to reduce boilerplate code across computer vision projects by providing standardized, reusable components. While the brief description suggests broad utility, the high daily star count indicates strong community adoption in the CV space.
📡 Analysis
Siri AI at WWDC 2026
June 8, 2026 · Post · tags: vision-llms, apple, generative-ai
Simon Willison takes a skeptical stance on Apple's new Siri AI announcements, given their 2024 Apple Intelligence overpromises. The new features leverage vision LLMs to extract screen information without requiring custom app integration, running on Google Cloud with NVIDIA hardware through Apple's Private Cloud Compute. While the technology seems more feasible than previous promises, Willison maintains a "believe it when I see it" policy until real user reports emerge from the iOS 27 Developer Beta waitlist.
datasette-agent-edit 0.1a0
June 7, 2026 · Post · tags: datasette-agent, llm-tool-use
Willison released a foundational plugin for Datasette Agent that implements storage-agnostic text editing tools inspired by Claude's editor design. The plugin provides view, str_replace, and insert operations that can be adapted across different editing contexts like Markdown collaboration, SQL query updates, and SVG file modification. This base toolset aims to standardize agentic text editing patterns across multiple Datasette Agent plugins.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net