AI Briefing — 2026-08-13

🔬 Research

On-Policy Self-Distillation without Any Supervision (U-OPSD)

189 upvotes · Yijiang Li et al. · Paper

U-OPSD trains LLMs to improve themselves using only their own outputs — no ground truth labels, no reward models, no teacher models. It samples multiple rollouts, builds a pseudo-solution via majority vote, then distills the model on disagreeing completions, forcing self-correction precisely where the model is "confidently wrong." On five math benchmarks (AIME24/25, HMMT25, MATH500, AMC23), it beats the base Qwen3 by up to +10.7% at 8B scale and outperforms supervised OPSD by up to +3.2% — a meaningful result for post-training without any labelled data.

ComBodied Agents: a New Paradigm of Human-Centric Agentic AI

176 upvotes · Qianggang Ding et al. · Paper

The paper argues that both digital and embodied AI agents share a blind spot: they transform states (software or physical) but never model the person as the primary object. "Combodied Agents" propose a closed loop — multimodal event perception, longitudinal correctable memory, Personal World Models for counterfactual reasoning, and an admissible intervention policy — using robots, sensors, and services as channels rather than endpoints. The framework is positioned as a unifying architecture for health agents, AI companions, and adaptive assistants, with explicit attention to consent, reversibility, and user control. Relevant for anyone building long-horizon, human-in-the-loop agent systems.

Beyond Pixels: From Video Priors to 4D Worlds (Latent-to-4D)

133 upvotes · Zihao Liu et al. · Paper

Latent-to-4D skips the RGB frame entirely: it feeds the final denoised latents of a video diffusion model directly into a 4D decoder, avoiding the distribution mismatch that plagues RGB-then-reconstruct pipelines. A single checkpoint trained on ~1K clips transfers across multiple video diffusion transformers that share the same VAE family, without retraining. On Text4D-200 and I4D-200 benchmarks, it outperforms same-latent Wan+4RC cascades by +2.88–5.81 points in DINO-F1 and wins human preference evaluations on geometry and temporal stability.

🛠 Tools

cathrynlavery/diagram-design — Editorial diagram templates for coding agents

2855 stars/day · Repo

A collection of 27+ self-contained HTML+SVG diagram types (architecture diagrams, flowcharts, flywheels, etc.) packaged as a Claude Code / Codex / Pi agent skill. No Figma, no Mermaid, no JavaScript build step — just open the HTML in a browser. It can also redraw existing draw.io or Mermaid sources. With 12K total stars and an active push today, it's gaining traction fast as a drop-in visual layer for AI-assisted writing and documentation.

msitarzewski/agency-agents — Curated library of specialist AI agent personalities

1873 stars/day · Repo

A MIT-licensed collection of opinionated, personality-driven agent prompts (frontend developers, community managers, "whimsy injectors", etc.) installable into Claude Code, Cursor, Codex, Gemini, and others via a native desktop app (macOS/Linux/Windows). With 144K total stars, this is one of the most starred prompt-engineering repositories on GitHub. Useful as a reference for how to structure deliverable-focused, domain-specific system prompts.

stablyai/orca — Parallel coding agent orchestrator (ADE)

1235 stars/day · Repo

Orca is an "Agent Development Environment" that runs Codex, Claude Code, OpenCode, and Pi side-by-side in isolated git worktrees, tracked from a single UI available on desktop, mobile, and VPS. The mobile companion lets you monitor and steer agents from your phone. At 44K stars and MIT licensed, it's the most mature open-source option for developers who want to parallelize multi-agent coding workflows without vendor lock-in.

📡 Analysis

DeepSeek V4 Pro 0813 (on OpenRouter)

2026-08-12 · Post · tags: ai, llms, deepseek, llm-release, ai-in-china

DeepSeek quietly shipped a new frontier model (V4 Pro 0813) with no official announcement page — benchmarks leaked via WeChat, then Reddit, then Hacker News in ASCII-art form. Simon notes the model produces noticeably different outputs at low/medium/high reasoning levels, a behaviour he hasn't seen in other models. Open weights seem likely given that V4-Pro (April) and V4-Flash-0731 (July) are already on HuggingFace, but unconfirmed.

alchemy-utils 0.1a0 — A database-agnostic sqlite-utils

2026-08-12 · Post · tags: python, sqlite, postgresql, duckdb, sqlalchemy, coding-agents

Simon used Codex + GPT-5.6 Sol Ultra in a single morning to build an alpha of alchemy-utils, a SQLAlchemy-backed port of his popular sqlite-utils CLI/library that works across SQLite, PostgreSQL, and DuckDB. The project reached a releasable alpha state with "very few follow-up prompts" — a concrete data point on the ceiling of agentic code generation for well-scoped library tasks. The optimised DuckDB CSV import dropped from ~60 minutes to ~35 seconds after a Codex pass.

Florian Herrengt: "AI is removing the middle class of software engineering"

2026-08-12 · Post · tags: ai-misuse, cognitive-debt, ai-assisted-programming

Herrengt describes a team that has delegated so much to AI that nobody can answer "where does the data come from?" — they just ask Claude and watch a confident wall of text appear. The post frames this as cognitive debt: AI accelerates feature shipping while eroding the team's capacity to understand its own system, making the next bug progressively harder to fix. A useful counter-weight to pure productivity narratives around AI-assisted development.

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