AI Briefing — 2026-07-05

🔬 Research

Program-as-Weights: A Programming Paradigm for Fuzzy Functions

74 upvotes · Wentao Zhang et al. (Harvard) · Paper

PAW proposes replacing LLM API calls for "fuzzy" tasks (log triage, JSON repair, semantic search ranking…) with small, locally-executable neural adapters compiled once from a natural-language spec. A 4B compiler trained on the released 10M-example FuzzyBench dataset generates parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter — matching Qwen3-32B quality at ~1/50th of the memory footprint and 30 tok/s on a MacBook M3. The key shift: the foundation model becomes a one-time "tool builder" rather than a per-query cost centre, which directly addresses reproducibility, latency, and API pricing concerns.

AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

45 upvotes · Xiangchen Cheng et al. · Paper

Current long-horizon agent evaluations suffer from unbounded context windows where the contribution of any single memory component is impossible to isolate. AgenticSTS introduces a strict bounded-memory contract — each decision gets a fresh prompt assembled by typed retrieval, with no raw transcript carryover — and instantiates it inside Slay the Spire 2, a stochastic deck-builder requiring hundreds of decisions per run. Frontier models currently report 0% win rate on the game's lowest difficulty (human baseline: 16%), making it a genuinely challenging but non-saturated benchmark. The released dataset (298 full trajectories with condition tags) enables controlled ablation of individual memory layers, something most existing agent benchmarks don't allow.

EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

43 upvotes · Zhilin Wang et al. · Paper

EvoPolicyGym benchmarks a specific, underexplored capability: can an LLM agent iteratively rewrite an executable RL policy within a fixed interaction budget, getting better through feedback? The suite covers 16 compact RL environments, and GPT-5.5 currently tops the leaderboard with the best aggregate rank and top-2 finishes across all 16 tasks. Beyond scores, the benchmark provides trajectory-level diagnostics showing how agents allocate their budget between exploration and parametric tuning — a diagnostic lens missing from existing code-generation evals.

🛠 Tools

usestrix/strix — Open-source AI penetration testing agent

1 904 stars/day · ⭐ 36 465 total · Apache 2.0 · Repo

Strix deploys autonomous AI "hacker" agents that probe a web application for vulnerabilities and then propose fixes. It integrates with GitHub Actions and CI/CD pipelines for automated scanning on every pull request, making it a security testing layer rather than just a one-off audit tool. Written in Python, installable via pip install strix-agent, with a hosted dashboard at strix.ai for teams who don't want to self-host the reporting layer.

JuliusBrussee/caveman — Cut LLM agent output tokens by 65% via caveman-speak

1 089 stars/day · ⭐ 84 343 total · Repo

Caveman is a skill/plugin that instructs coding agents (Claude Code, Codex, Cursor, Windsurf, Cline, and 30+ others) to strip filler prose and answer in ultra-terse "caveman" style — code, commands, and error messages stay byte-for-byte exact, but surrounding explanation tokens shrink by ~65%. At current API rates that compounds into meaningful cost savings on every reply for the lifetime of a project. The sheer star count (84 k) signals how acutely developers feel the output-token cost squeeze right now.

mattpocock/skills — Composable Claude Code skills for real engineering workflows

973 stars/day · ⭐ 156 967 total · Repo

Matt Pocock (Total TypeScript) publishes the Claude Code skills he uses daily, installable in 30 seconds via npx skills@latest add mattpocock/skills. Unlike heavier frameworks (BMAD, Spec-Kit), these are intentionally small and composable — covering issue tracking (GitHub/Linear), triage labelling, and doc generation — meant to be forked and adapted rather than followed blindly. The ~60 k newsletter subscribers and 157 k stars suggest this has become a de-facto starting point for engineers setting up agentic coding environments.

📡 Analysis

sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)

2026-07-05 · Post · tags: sqlite-utils, claude, claude-code, coding-agents, agentic-engineering

Simon Willison used Claude Fable (via Claude Code on his iPhone during the 4th of July parade) to drive sqlite-utils toward a stable 4.0 release: 37 prompts, 34 commits, +1 321 / -190 lines across 30 files, for $149.25 total. Fable caught a critical data-loss bug in delete_where() that Willison himself had missed — the method left the SQLite connection permanently in_transaction=True, silently eating subsequent writes. The post is a candid accounting of both the capability and the cost of using frontier coding agents on real open-source library work.

Building a World Map with only 500 bytes

2026-07-04 · Post · tags: ascii-art, data-urls, javascript

Iwo Kadziela (with Codex assistance) encoded a recognisable ASCII world map in 445 bytes by deflate-compressing the data and streaming it via the browser's native DecompressionStream API piped through a data: URI fetched with fetch(). The technique is a neat reminder that browser primitives can replace entire decompression libraries, and that AI-assisted micro-optimisation is pushing the "code golf" genre into genuinely practical territory.

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