← Home

AI Briefing — May 9, 2026

🔬 Research

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

58 upvotes · Li et al. · Paper

Proposes Direct Corpus Interaction (DCI), where agents search raw text with terminal tools (grep, file reads) instead of vector databases or retrieval APIs. The approach outperforms conventional sparse, dense, and reranking methods on BRIGHT and BEIR benchmarks while requiring no offline indexing. This matters because it removes the bottleneck of fixed similarity interfaces that filter out evidence too early, enabling multi-step reasoning and hypothesis refinement that agentic tasks demand.

Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

58 upvotes · Shi et al. · Paper

Introduces a framework where a single policy learns to select skills from a library, use them during execution, and distill new skills from experience—all optimized toward one task-outcome signal. Unlike existing methods that optimize these capabilities separately, Skill1 shows superior performance on ALFWorld and WebShop benchmarks. This addresses the persistent challenge of maintaining reusable skill libraries across tasks without conflicting evolution between skill components.

MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction

55 upvotes · Cui et al. · Paper

Delivers real-time simultaneous seeing, listening, and speaking through "Omni-Flow," a streaming framework that aligns all modalities on a shared temporal axis. With 9B parameters, it approaches Gemini 2.5 Flash performance while running on edge devices with under 12GB RAM. This represents a shift from turn-based to continuous interaction, enabling proactive behaviors like issuing reminders based on live scene understanding rather than just reactive responses.


🛠 Tools

Hmbown/DeepSeek-TUI — Terminal coding agent for DeepSeek V4

3731/day · Repo

A Rust-built terminal interface that streams DeepSeek V4's reasoning blocks, edits local workspaces with approval gates, and includes auto mode for model/thinking level selection. Distributed as prebuilt binaries via npm, Cargo, Homebrew, or Docker with support for Linux, macOS, and Windows. The 22.5K stars reflect strong adoption for developers wanting DeepSeek's coding capabilities without leaving the command line, especially valuable given DeepSeek V4's competitive performance at lower cost.

anthropics/financial-services — Claude agents for FSI workflows

3660/day · Repo

Production-ready agents for investment banking, equity research, private equity, and wealth management that work both as Claude Cowork plugins and via the Managed Agents API. Includes named workflow agents (Pitch Agent, Market Researcher, GL Reconciler) plus vertical-specific skills and data connectors. This matters because it provides compliant, review-gated financial automation—the agents draft work product but require human sign-off for all investment decisions and transactions.

addyosmani/agent-skills — Production engineering workflows for AI agents

1893/day · Repo

Seven slash commands (/spec, /plan, /build, /test, /review, /code-simplify, /ship) that encode senior engineer workflows and automatically activate appropriate skills based on development phase. With 36.4K stars, it represents a systematic approach to making AI agents follow production-quality engineering practices. This addresses the gap between AI coding capability and real-world software development discipline, ensuring agents consistently apply quality gates and best practices.


📡 Analysis

Quoting Luke Curley

May 9th · Post · tags: webrtc, openai

Luke Curley (formerly Discord) explains why WebRTC is problematic for AI voice applications: it aggressively drops audio packets to maintain low latency, which works for conference calls but destroys expensive AI prompts. Users would prefer waiting 200ms for accuracy over getting garbage responses from degraded audio. This highlights a fundamental mismatch between real-time communication protocols and AI interaction patterns where prompt fidelity matters more than instant response.

Using Claude Code: The Unreasonable Effectiveness of HTML

May 8th · Post · tags: html, security, markdown, ai, prompt-engineering, generative-ai, llms, llm, claude-code

Anthropic's Thariq Shihipar argues for requesting HTML output from Claude instead of Markdown, enabling SVG diagrams, interactive widgets, and rich navigation that make explanations more effective. Simon demonstrates this with a GPT-5.5 analysis of a Linux security exploit, producing an interactive HTML explanation. This challenges the Markdown default many adopted during token-limited GPT-4 days, suggesting HTML's expressiveness now outweighs its token overhead for complex explanations.


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