AI Briefing — 2026-07-06
🔬 Research
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
91 upvotes · Wentao Zhang et al. (Harvard) · Paper
PAW introduces a new paradigm where "fuzzy" tasks (log alerting, JSON repair, intent-based ranking) are compiled from a natural-language spec into small neural adapters rather than delegated to a cloud LLM at every call. A 4B compiler model trained on 10M examples emits parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter — which then matches Qwen3-32B performance at 1/50th the memory footprint and runs at 30 tokens/s on a MacBook M3. This reframes the foundation model as a tool builder invoked once per function, making subsequent calls cheap, local, and reproducible — a compelling alternative to per-query API calls.
AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents
53 upvotes · Xiangchen Cheng et al. · Paper
The paper tackles a core problem with long-horizon agents: unbounded context accumulation makes it hard to isolate what memory components actually contribute to decisions. The authors propose a "bounded contract" where each decision receives a fresh, typed-retrieval prompt with no raw cross-decision transcript — tested inside Slay the Spire 2, a demanding stochastic deck-builder. Adding a strategic skill layer raised win rates from 3/10 to 6/10 at the lowest difficulty (vs. 0 wins reported on a public frontier-LLM benchmark). The released testbed — 298 tagged trajectories, frozen snapshots, and analysis scripts — is the main practical contribution for researchers studying agent memory design.
The Mirage of Optimizing Training Policies: MIPI for LLM RL
52 upvotes · Jing Liang et al. · Paper
This paper identifies a subtle but critical flaw in LLM RL training: even with synchronized weights, the training engine and inference engine assign inconsistent probabilities to the same trajectories, creating a persistent off-policy gap that poisons updates. Crucially, improvements to the training policy don't guarantee improvements to the deployed inference policy — a misalignment prior work ignored. The proposed MIPU framework adds a two-step process that constructs candidate updates and filters them through an inference-side gap proxy, improving reasoning performance and training stability at two model scales.
🛠 Tools
openai/codex-plugin-cc — Use Codex from inside Claude Code
1532 stars/day · Repo
This official OpenAI plugin bridges the two dominant AI coding agents: it lets Claude Code users invoke OpenAI Codex for code reviews (/codex:review, /codex:adversarial-review) or delegate tasks to it as a background subagent (/codex:rescue, /codex:transfer). It requires only a ChatGPT Free subscription or an OpenAI API key, installs in three commands, and already has ~26K total stars. For teams using Claude Code as their primary driver, it's a practical way to cross-check work with a second agentic system without switching context.
Zackriya-Solutions/meetily — Local-first AI meeting assistant
1409 stars/day · Repo
Meetily is a fully local, open-source meeting transcription and summarization tool for macOS and Windows, built in Rust with Parakeet/Whisper for live transcription (advertised as 4× faster), speaker diarization, and Ollama for summarization. No audio ever leaves the machine — all processing happens on-device with zero cloud dependency. For teams with privacy constraints or cloud-averse policies, this is a self-hostable drop-in alternative to services like Otter.ai or Fireflies.
addyosmani/agent-skills — Production engineering workflows for AI coding agents
1114 stars/day · Repo
With 70K+ total stars, this repo packages senior-engineer workflows as 8 slash commands (/spec, /plan, /build, /test, /review, /ship, etc.) that AI coding agents can follow consistently across the full development lifecycle. The key insight is encoding quality gates and best practices as reusable "skills" so agents don't drift from team standards mid-task. A /build auto mode chains spec → plan → implementation in a single approved pass for a more autonomous flow.
📡 Analysis
sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)
2026-07-05 · Post · tags: sqlite-utils, claude-code, agentic-engineering, llm-pricing
Simon Willison used Claude Fable (Anthropic's latest model, available on Max plans) as a coding agent to drive sqlite-utils toward a 4.0 stable release — 37 prompts, 34 commits, +1,321/-190 lines across 30 files, for $149.25 total. Fable caught a critical silent data-loss bug in delete_where() that left database connections perpetually in-transaction, something Willison hadn't spotted himself. The experiment is a concrete, costed data point on what agentic coding with a frontier model looks like on a real open-source project.
sqlite-utils 4.0rc3
2026-07-06 · Post · tags: sqlite-utils, annotated-release-notes, gpt, claude-mythos-fable
A day later, Willison pushed rc3 after continuing to work through the issue backlog with both Claude Fable 5 and GPT-5.5 in tandem. The biggest additions are support for compound foreign keys (with a minor breaking change to table.foreign_keys) and case-insensitive column name handling following SQLite's own conventions. The stable 4.0 release is imminent; the growing changelog since rc2 is the only thing holding it back.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net