AI Briefing — May 15, 2026

🔬 Research

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

107 upvotes · Yafu Li et al. · Paper

Researchers present SU-01, a 30B parameter model that achieves gold-medal performance on International Mathematical and Physics Olympiad problems through a three-stage training recipe: reverse-perplexity curriculum SFT, two-stage RL (verifiable rewards → proof-level RL), and test-time scaling. The model can handle reasoning trajectories exceeding 100K tokens and demonstrates strong generalization beyond mathematics to scientific domains, marking a significant advance in automated theorem proving and competition-level problem solving.

Self-Distilled Agentic Reinforcement Learning

54 upvotes · Zhengxi Lu et al. · Paper

SDAR addresses the instability issues that arise when applying self-distillation to multi-turn AI agents by treating it as a gated auxiliary objective alongside RL. The method uses sigmoid gates to strengthen distillation on teacher-endorsed tokens while softly attenuating negative rejections, achieving substantial improvements over GRPO baselines (+9.4% on ALFWorld, +10.2% on WebShop). This work is particularly relevant as multi-turn agents become more prevalent in production environments.

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

49 upvotes · Xiyu Ren et al. · Paper

MemLens introduces the first systematic benchmark comparing long-context LVLMs versus memory-augmented agents across 789 questions requiring genuine multimodal evidence. The study reveals that long-context models excel at short interactions but degrade with conversation length, while memory agents maintain stability but lose visual fidelity during compression. With most systems capping below 30% on multi-session reasoning, this benchmark highlights the need for hybrid architectures combining long-context attention with structured multimodal retrieval.

🛠 Tools

tinyhumansai/openhuman — Personal AI super intelligence

3329 stars/day · Repo

OpenHuman is a Rust-based personal AI assistant positioning itself as "private, simple and extremely powerful" with 8,269 total stars. Currently in early beta, it offers local AI deployment with active development and strong community engagement across Discord, Reddit, and social platforms. The project emphasizes privacy and user control, making it appealing for developers wanting self-hosted AI without cloud dependencies.

mattpocock/skills — Engineering skills for AI agents

2987 stars/day · Repo

Matt Pocock's collection of agent skills designed for "real engineering" rather than "vibe coding" has exploded to 83,641 total stars. The skills work with any model and focus on being small, composable, and based on decades of engineering experience, covering issue tracking (GitHub/Linear), triaging, and documentation. With 60,000 developers on his newsletter, this represents a significant shift toward practical, production-ready agent tooling.

rohitg00/agentmemory — Persistent memory for AI coding agents

1879 stars/day · Repo

AgentMemory provides persistent memory capabilities for AI coding agents across platforms like Claude Code, Cursor, and Gemini CLI, reaching 9,319 total stars. Built on the iii engine, it implements confidence scoring, lifecycle management, and hybrid search to extend Karpathy's LLM Wiki pattern. The project addresses a critical gap in agent workflows where context is lost between sessions, making agents more practical for ongoing development work.

📡 Analysis

Not so locked in any more

May 14, 2026 · Post · tags: react, coding-agents, ai-assisted-programming, generative-ai, ai, llms

Willison shares an anecdote about a company that rewrote legacy iPhone and Android apps to React Native using coding agents, with the reasoning that if it proves wrong, they can "just port back to native." This exemplifies Mitchell Hashimoto's observation that programming languages are no longer "LOCK IN" - AI coding capabilities are making technology choices more reversible and reducing the traditional switching costs between platforms and languages.

Quoting Mitchell Hashimoto

May 14, 2026 · Post · tags: zig, ai, mitchell-hashimoto, llms, rust, generative-ai, agentic-engineering, bun

Hashimoto's quote on Bun's Zig-to-Rust migration captures a fundamental shift: "Programming languages used to be LOCK IN, and they're increasingly not so... Bun has shown they can be in probably any language they want in roughly a week or two. Rust is expendable." This observation highlights how AI-assisted development is making language migrations dramatically faster, fundamentally changing how we think about technical debt and architectural decisions.

datasette-ip-rate-limit 0.1a0

May 14, 2026 · Post · tags: datasette, rate-limiting, codex

Willison had Codex (GPT-5.5 xhigh) build a rate-limiting plugin for Datasette to combat crawler abuse on datasette.io. The plugin offers configurable IP-based rate limiting with exemptions, time windows, and blocking periods - a practical example of using AI to solve immediate operational problems rather than just development tasks.

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