AI Briefing — 2026-08-12
🔬 Research
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
306 upvotes · Engdahl, Kosowski, Chorowski et al. · Paper
BDH-CQ is a 150M-parameter reasoning model that updates a recurrent memory from in-context examples and solves queries by iterating entirely in latent space — no chain-of-thought verbalization. On ARC-AGI-1, it reaches 29.5% pass@2 at just $0.0007 per task, breaking the previously known cost-accuracy Pareto frontier. For teams exploring efficient reasoning under tight compute budgets, this is a meaningful datapoint: comparable accuracy at a fraction of the inference cost of larger models.
On-Policy Self-Distillation without Any Supervision
167 upvotes · Li, Wang, Liang et al. · Paper
U-OPSD improves LLMs using only the model's own generations: it samples multiple rollouts, builds a pseudo-solution via majority vote, and fine-tunes the model specifically on disagreeing completions. No ground truth, no external reward signal, no teacher model. On five math benchmarks (AIME24/25, HMMT25, MATH500, AMC23), it beats GRPO and matches or exceeds supervised OPSD — +8.5% / +10.7% over base Qwen3 at 4B and 8B. Code is available at github.com/williamium3000/u-opsd.
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
73 upvotes · Razzhigaev, Gritsaev, Kaznacheev et al. · Paper
Ouroboros is an agent framework that rewrites its own tools, prompts, and core code via reviewed commits — essentially eating its own tail to improve. It scores 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified, both new state-of-the-art results. A 161-day live deployment ("Hope") demonstrates the framework in continuous free evolution, raising real questions about operational safety when an agent can self-modify and select its own model APIs.
🛠 Tools
cathrynlavery/diagram-design — Editorial SVG diagrams for Claude Code
1,616 stars/day · Repo
A collection of 27 diagram types (architecture, flowcharts, pyramids, flywheels, etc.) delivered as self-contained HTML+SVG files — no Figma, no Mermaid, no build step. Ships as a Claude Code / Codex / Pi agent skill that reads your site to auto-map your brand's colors and fonts in ~60 seconds. Version 2.0 added the "Loop" flywheel type; the repo has 7,600+ total stars and was last updated today.
PrimeIntellect-ai/prime-agent — Self-improving RLM agent for coding & research
1,138 stars/day · Repo
Prime Agent is an open-source TypeScript agent built around two concepts: the Recursive Language Model (RLM), which treats prompt context as mutable variables and sub-agents as function calls, and the Continual Harness, which stores memories and skills as durable state the agent can refine over time. Designed for long-running autonomous coding and research tasks, it has 14,400+ total stars and is actively maintained (last push today). It connects to PrimeIntellect's broader RL training infrastructure (prime-rl, verifiers).
msitarzewski/agency-agents — Collection of specialized AI agent personas
958 stars/day · Repo
A library of meticulously crafted AI agent "personalities" (frontend wizards, community managers, whimsy injectors, reality checkers…) installable into Claude Code, Cursor, Codex, and Gemini. Each agent has a defined domain, communication style, and deliverables. With 143,960 total stars and a companion native app for macOS/Linux/Windows, this is clearly very popular — though engineers should note it's mostly prompt engineering at scale rather than novel architecture.
📡 Analysis
There are no lossless transformations of natural-language text
2026-08-11 · Post · tags: writing, ai, generative-ai, llms, ai-misuse
Simon links to Sophie Alpert's AI writing policy for engineers, built around a key insight: every rewrite by an LLM loses information, because the model doesn't share the author's precise mental model. The practical rule: you must be able to defend every sentence as representing your own thinking — "AI wrote that, ignore it" is not acceptable. Relevant to any team setting AI writing guidelines for docs, design specs, or PRDs.
Stealing Reasoning Traces from Proprietary LLM APIs
2026-08-11 · Post · tags: jailbreaking, openai, prompt-injection, llm-reasoning, anthropic, gemini
Researchers discovered that Anthropic, OpenAI, and Google share the same encryption key across all models in a family for their encrypted chain-of-thought blocks — meaning a trace from a strong model could be replayed into a weaker sibling and jailbroken to reveal raw reasoning in plaintext. Claude Haiku 4.5 was the easiest target. All providers have since patched the issue, but the paper's appendix provides a rare glimpse into unfiltered proprietary reasoning traces, and also documents a prompt-injection variant. A significant finding for anyone building on or auditing LLM APIs.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net