Summary
No curated reposts or saved posts landed in the morning slot. This synthesis comes from the morning capture of the Following feed, which covers the US Sunday evening. The strongest signal is a paper on training data agents: Data World builds verifiable data-analysis environments from a written-down generating mechanism, and RL on them adds 20.9 points on DABstep-Hard. The second theme is a cluster of four posts on harnesses and who controls them. One is a self-revising harness claim, one an AI Engineer NYC guide asking whether harnesses are dead, plus DAIR's harness-heavy papers of the week and Memorilla's 16-token memory, recirculating. A smaller cluster of three posts covers human oversight of long-running agents. On compute, Elon Musk says SpaceX cannot rent inference capacity from AWS or GCP, which fits the tight-supply story in today's digest. GitHub Security Lab's 24 Android vulnerabilities found by an open-source agent is the most concrete deployment result. Anthropic's Public Benefit Fund and staff-survey numbers drew the most commentary, much of it skeptical.
Posts
Data World: build the world before the files (@omarsar0; paper). Data agents write code over messy files to find insights, but training them needs environments that are costly, often private, and hard to grade when the answer is an open-ended finding. Data World writes down the mechanisms that generate a dataset first, then produces the files, tasks and grading rubrics from them, and audits that each answer can be recovered from the files. It produced 2,135 training tasks from 1,037 separate worlds. RL from scratch on Qwen3.6-35B-A3B, with no distilled trajectories, adds 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard and 12.13 on JobBench. Ten frontier models compute correct numbers more often than they draw correct conclusions. See agent shorts.
Harnesses: who writes them, and are they dying? (cluster of 4) (@virgilxbt, @TheTuringPost, @dair_ai, @omarsar0). A hype-styled thread describes a "The End of Harness Engineering" paper: a frozen solver, an editor proposes 8 edits to the harness file per round, each is scored on held-out questions the editor never sees, and the best is kept. The held-out score went from 0.31 to 0.64 over 41 rounds, and the human writes only the scoring set. Treat the framing with care, but the idea matches Harness² and ReHarness from the week. Turing Post's AI Engineer NYC guide reports engineers removing harness parts that newer models no longer need, while others build harnesses for multi-day runs. DAIR's top papers for Oct 12-18 lead with Harness², ReHarness and LongHarness. Memorilla, which writes a whole document set into 16 memory tokens for a frozen LLM, kept recirculating (summary). See agent harness engineering.
Human oversight of long-running agents (cluster of 3) (@junjiechou1122, @math_rachel, @stanfordnlp). A researcher argues that after hours of autonomous work, humans can still approve actions but no longer know what is happening. Real oversight needs three things: enough information about state and evidence, a translation into the human's mental model, and a dialogue instead of an "Approve" button. Rachel Thomas started a thread of papers on instruction-following failures, opening with "cognitive inertia," where models struggle with instructions that break their training conventions (paper), and the PrefEval benchmark of 3,000 user-preference pairs. Noah Goodman shared a draft conceptual map for an AI safety lecture.
SpaceX cannot rent inference compute (@elonmusk). Musk says SpaceX tried to rent inference capacity from AWS, GCP and others to move load off its training clusters and could not get it. One more anecdote for the sold-out-compute story.
GitHub's security agent finds 24 Android bugs (@github; blog). GitHub Security Lab ran its open-source AI security agent with reusable "taskflows" over Android apps and found 24 vulnerabilities, including a location leak in a navigation app with 10M downloads. The post explains how to run the taskflows on your own project.
An agentic factory for consulting work (@sermakarevich; repo). A consultant describes a Temporal-based workflow system that sends hard and easy tasks to different models. It automates daily change summaries across Jira and Git, paper distillation into a knowledge wiki, tutorials and spec-driven coding. A practitioner version of routing by task difficulty.
Pretraining is saturating (@sarahookr). Sara Hooker says most pretraining gains are now data-recipe tweaks, which pushes innovation to post-training and test-time compute. She calls that good for competition, since pretraining concentrated resources in a few labs.
Anthropic's benefit fund and staff survey (@rohanpaul_ai, @GaryMarcus, @elonmusk). Anthropic's negotiation-norms push and Public Benefit Fund (5% of shares, growing with valuation) dominated commentary. Marcus agreed with David Sacks that the fund is a play for public favor, and Musk said the shares should go directly to the public. See the daily digest.
Generative-model research notes (@sedielem, @TheTuringPost). Sander Dieleman reposted LLM-to-Flow Distillation, which adds natural-language control to scientific generators, and multimodal bitstream co-generation. Stanford and Together's Next-Concept Modeling, coarse-to-fine text generation through compressed codes, resurfaced (covered 10-16).
Skip. The "REA reverse-engineers 100 things" thread, an "Anthropic published 400+ agents" post (a third-party repo, misattributed), a "$40,000/month solo agency" Grok thread, a "Karpathy: delete everything, keep Graph" clip summary, Google-ecosystem listicles, election posts and off-topic posts.