AI Briefing — May 14, 2026

🔬 Research

MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image

79 upvotes · Alan Arazi et al. · Paper

Researchers introduce MulTaBench, a benchmark of 40 datasets for multimodal tabular learning that combines structured data with text and images. Unlike existing benchmarks that focus on simple co-occurrence, this benchmark emphasizes tasks where modalities provide complementary predictive signals, requiring target-aware representations rather than generic embeddings. The work addresses a critical gap in tabular foundation models, which currently struggle with unstructured data, and spans high-impact domains like healthcare and e-commerce to enable development of next-generation multimodal tabular models.

MinT: Managed Infrastructure for Training and Serving Millions of LLMs

66 upvotes · Mind Lab et al. · Paper

MindLab presents MinT, a system for managing millions of LoRA-adapted policies over shared base models up to 1T parameters. The key innovation is keeping base models resident while moving only LoRA adapters (under 1% of base model size), achieving 18.3x speedup on 4B models and supporting 100K+ policy catalogs. MinT enables massive-scale personalized AI deployment without the prohibitive cost of materializing full checkpoints for each adaptation, making it practical to serve diverse specialized models at enterprise scale.

AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

61 upvotes · Yuchao Gu et al. · Paper

AnyFlow tackles a key limitation in video diffusion: consistency-distilled models that perform worse as you give them more sampling steps. Instead of distilling for fixed step counts, AnyFlow optimizes the full ODE trajectory using flow-map transitions over arbitrary time intervals. The approach maintains performance across 1.3B to 14B parameter models while scaling effectively with sampling budgets, addressing the trade-off between speed and quality that has limited practical video generation applications.

🛠 Tools

mattpocock/skills — AI agent skills for real engineering

3392 stars/day · Repo

Matt Pocock's collection of practical AI agent skills designed for actual software development rather than "vibe coding." The 80K+ star repo provides composable, model-agnostic skills covering issue triage, testing workflows, and development processes based on decades of engineering experience. With a simple npx skills@latest add mattpocock/skills installer and 60K+ developers following updates, this represents a shift toward structured, production-ready AI-assisted development workflows.

tinyhumansai/openhuman — Personal AI assistant in Rust

1696 stars/day · Repo

OpenHuman is an open-source agentic assistant built in Rust that integrates with daily workflows while maintaining privacy. Despite being in early beta, it's gained nearly 7K stars and offers cross-platform installation via curl/PowerShell scripts. The project positions itself as a privacy-first alternative to cloud-based AI assistants, running entirely locally while providing "super intelligence" capabilities for personal productivity.

obra/superpowers — Complete agentic development methodology

1401 stars/day · Repo

Superpowers is Jesse Vincent's comprehensive software development methodology for coding agents, reaching 190K+ stars. Rather than jumping straight to code, it makes agents step back to understand requirements, create digestible specs, and execute through "subagent-driven-development" with true TDD practices. The framework emphasizes structured planning and autonomous execution, with Claude reportedly able to work independently for hours while following the established plan.

📡 Analysis

Welcome to the Datasette blog

May 13, 2026 · Post · tags: datasette, ai-assisted-programming, codex

Simon Willison announces the new Datasette project blog, built using OpenAI Codex desktop. He highlights Codex's session transcript export feature as particularly useful, sharing the complete development session that created the blog. This signals growing maturity in AI-assisted development workflows, where the development process itself becomes a shareable artifact for learning and documentation.

Quoting Boris Mann

May 13, 2026 · Post · tags: ai-agents, agent-definitions

Boris Mann's critique that "11 AI agents is meaningless as a phrase" highlights the current confusion around AI agent terminology. Comparing agents to spreadsheets or browser tabs, he points out that the number alone doesn't indicate capability or purpose. This observation reflects the need for better standardization and clearer definitions as AI agents become more commonplace in workflows, moving beyond marketing buzzwords toward practical categorization.

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