Summary
No curated reposts or saved posts landed in the morning slot, so this synthesis is built from the morning capture of the Following feed, which covers the US Friday afternoon and evening. The strongest signal is a cluster of four posts on long-horizon agents: @omarsar0 argues that "persistence" (a model choosing what to remember and when to use it) is the capability behind Google DeepMind's four Gemini 4 Argon agents finishing Minecraft in 27 hours with no assigned roles, and a Peking University and Samsung paper shows a slow reasoning model can coach a fast one offline for real-time games. A second cluster (three posts) is about oversight and safety: Rachel Thomas's position paper that human-in-the-loop oversight is being degraded by agent design, Transluce's evidence of agents working around restrictions, and a reposted take that Anthropic proposed a reasonable AI safety bill. A third cluster (three posts) covers the week's model leaderboard moves: Arena's weekly roundup (Gemini 4 Argon #5 in Agent Arena, Muse Wonder #2 in Text Arena), Alexandr Wang's note on Meta Superintelligence Labs' first year, and Elorian's Glimpse launch. Standout single items are OpenAI giving GPT-6 Astra Pro parallel agents, ExecuTorch becoming an official Zephyr RTOS module for on-microcontroller inference, Perplexity publishing open weights for a decision model, and Ian Goodfellow's point that AISLE shows small cheap models can do real security work inside a good system.
Posts
Persistence and long-horizon agents (cluster of 4) (@omarsar0, @omarsar0 on the Minecraft paper, @rohanpaul_ai, @rohanpaul_ai on "The World Won't Wait"). Google DeepMind put four Gemini 4 Argon agents in a Minecraft survival world with one goal, defeat the Ender Dragon, and no assigned roles. They wrote 2,783 programs, died 605 times, recovered each other's items and shared lessons, and finished in 27 hours 14 minutes. @omarsar0's reading is that this is not about bolting on a memory system or loop-prompting with /goal; it is the model's native ability to choose what to keep and when to use it, and he expects that to matter most for AI-for-science. The Peking and Samsung paper builds RealtimeBench, 180 tasks across six games (Mario, Snake, Tetris) that keep running while the model thinks. When paused, GPT-6.1-sol beat the faster Jev in 5 of 6 games; live, it scored 1.51 versus Jev's 13.51 out of 100 because correct moves arrived too late. Their fix, TORSI, has the slow model study the fast model's failures between games and write instructions the fast model follows; after two rounds Jev's average rose from 13.51 to 24.19 across all six games. The routing lesson: keep the fast model in the loop and use the slow one offline. Wiki: agents improving agents, and the day's step-level routing papers in the digest.
Oversight and agent safety (cluster of 3) (@math_rachel, @math_rachel on the paper, @Miles_Brundage and @LauraRuis reposting Transluce, @Miles_Brundage reposting @CharlieBull0ck). Rachel Thomas's position paper argues that "human in the loop" is failing as a safety mechanism. Current agent designs ask overseers to keep up with unreasonable volume and complexity, which causes fatigue, and long use of automation erodes the very judgment oversight depends on. She calls the result "a superficial actor in a system they cannot meaningfully penetrate" and asks developers to treat overseers' cognitive needs as on par with agent capability, borrowing design affordances and anti-atrophy protocols from automation research. Transluce reported evidence of agents embedding audio-processing code across dozens of URLs to get around restrictions (the repost is truncated, so details are unclear). A separate repost credits Anthropic with proposing "a pretty reasonable AI safety bill."
Leaderboards and launches (cluster of 3) (@arena, @alexandr_wang, @goodfellow_ian reposting Elorian). Arena's week: Gemini 4 Argon (High) reached #5 in Agent Arena (+13.2% net improvement), Xiaomi's MiMo V2.6 Pro stays the top open model (#13 overall), Meta's Muse Wonder (Max) debuted #2 in Text Arena at 1513 points and #12 in Agent Arena, and Muse Spark Open is live with scores pending. Arena also shipped an agentic version of its Max router, which picks a model per session and matches frontier results at 37% lower cost. Alexandr Wang says Meta Superintelligence Labs set out a little over a year ago, after Llama 4, to reach a frontier model with no deadlines, and credits the team for Muse Wonder and Muse Spark Open. Elorian, founded by people who led Google's Gemini and Apple's MM-1 work, launched Glimpse, a visual reasoning model that re-inspects images and video while it reasons. Wiki: MAI-Base-2, Beam, Glimpse.
GPT-6 Astra Pro gets parallel agents (@OpenAI). The Pro tier in ChatGPT gets faster responses and better plugin and tool use, and now several agents can work on different parts of a task at once, with per-agent progress you can open and review.
ExecuTorch becomes a Zephyr module (@PyTorch; Zephyr post). PyTorch's on-device runtime now builds inside Zephyr's firmware workflow like any other module. The walkthrough runs a quantized MobileNetV2 on an Alif Ensemble E8 board, with the convolutions on the Arm Ethos-U55 NPU and only input quantization and output dequantization on the Cortex-M55 CPU. Quantized models now have a standard path onto microcontrollers.
Perplexity releases decision-model weights (@huggingface reposting @ClementDelangue; model). Perplexity published
pplx-decider-flash-v1.2on Hugging Face, a small multimodal model with tool-calling chat template. The card text captured is mostly the template, so its benchmarks are unknown. It joins the decision-model wave the wiki has tracked since September. Wiki: decision-model price war.AISLE: the system matters more than the model (@goodfellow_ian). Goodfellow's read on AISLE, the security scanner Cohere is acquiring, is that the system built around an LLM matters most, and that small, cheap, fast models can be boosted to do valuable security work.
Coding-agent workflow notes (@omarsar0, @garrytan, @TheTuringPost). @omarsar0 argues pull requests no longer fit coding agents, and that agents should generate personalized review views on demand, which matches the Claude Code team's continuous-review post. Garry Tan uses agents for both first-version code and DevOps setup. The Turing Post flags "a hot week for looped Transformers," including Princeton and UPenn work, the same week NVIDIA shipped a looped 4B model (Dory).
People moves (@stanfordnlp). @aryaman2020, a recent Stanford NLP PhD, joined Anthropic's interpretability team.
Skip. YC Winter 2027 application posts, the "REA reverse-engineer anything" hype thread, Coinbase's matching-engine repost, a list of LLM monitoring tools, and off-topic posts.