SATURDAY · SEPTEMBER 26, 2026 · ISSUE 19 · YESTERDAY No. 19
Daily News — 2026-09-26
~6,500 packages released on PyPI·22 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
A cognitively grounded benchmark that exposes two distinct failure modes in video models and shows fine-tuning on synthetic physics data meaningfully closes one of them.
Video generation models fail at something human infants master by three and a half months: knowing that an object behind an occluder still exists. WROP (World Reasoning with Object Permanence) is a systematic attempt to…
A thorough, well-sequenced reference that treats production AI infrastructure as a discipline rather than a collection of API calls.
AI System Design is the discipline of wrapping an LLM in everything else a production system needs: inference servers, caches, load balancers, retrieval pipelines, agent loops, guardrails, and cost controls. This…
Papers
A cognitively grounded benchmark that exposes two distinct failure modes in video models and shows fine-tuning on synthetic physics data meaningfully closes one of them.
WROP pairs a 1.5M-sample object-permanence corpus with a 300-question blind Elo exam; their 16B PWM-WROP ranks first among continuation video world models.
Editing the agent's reasoning before execution—not simulating tool responses—is what actually breaks the compounding-error cycle in long-horizon tasks.
AEWM edits noisy reasoning and actions in task history rather than predicting tool responses, improving agent scores by 3.2–6.7 points across six benchmarks.
PoS categories live in compact latent groups, not single features — and the group size tracks how lexically variable the category is.
A study finding that part-of-speech distinctions are recoverable from SAE activations via compact latent groups, not one-to-one latent-to-category mappings.
A rigorous account of building mobile agents where the harness and the model improve each other across iterations, with honest cost and memory numbers to back it up.
Qwen-Planner-Agent uses an agentic data flywheel and CARE reward shaping to build mobile planning agents in a closed AI-for-AI loop.
Role-decoupled prompts and per-role advantage normalization let a single 8B model beat most 30B deep-search agents by fixing context collapse at the architectural level.
IterSynth splits deep search into a Planner and Synthesizer to reduce role coupling and context noise, training both with role-specific advantages via RDPO.
Exploration works, but it's brittle: the same model under the same budget can end 72 points apart, and stating a rule correctly still leaves a 30-point execution gap.
A benchmark that tests AI exploration in rule-based "Alien Worlds" whose logic conflicts with training data, so recall alone cannot solve the tasks.
7 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Freeze-at-deployment coding agents beat hand-engineered TAMP planners without receiving any of the symbolic scaffolding those planners depend on.
Coding agents hit 56–95% success vs 47% for hand-engineered planners on generalized TAMP, using an order of magnitude less compute per instance.
A closed-form architectural scalar that turns resource-constrained Transformer design into an exactly solvable knapsack, with no model weights or data required.
NSC ranks architectures from spec alone—no data or training—and its DP solver finds the globally optimal layout under resource constraints in seconds.
Cleaning SUN RGB-D's labels alone shifts a frozen model's mIoU by over two points — RGBD20K is built on that uncomfortable finding.
RGBD20K pairs 20,000 RGB-D images with 160 fine-grained categories, quadrupling the semantic coverage of NYUv2 and SUN RGB-D.
Tools & packages
A thorough, well-sequenced reference that treats production AI infrastructure as a discipline rather than a collection of API calls.
Step-by-step design guide covering LLM, RAG, and agent architectures for production AI systems.
A serious MCP-native reverse engineering workbench whose value is in the routing logic and knowledge base, not just the tool wrappers.
A multi-platform reverse engineering and CWE-602 client-side authorization audit workbench designed for agent-native workflows.
A local-first, MIT-licensed clip tool that routes AI calls through your own OpenRouter key and is unusually candid about what data goes where.
BridgeMind releases an open-source desktop app for AI-powered video clipping.
A disciplined template for agent-directed generative film: the scaffolding that makes a coding agent a reliable creative collaborator is as carefully built as the eight films it produced.
A collection of procedural risograph films as single HTML files, with docs and tools for making your own.
A semantically differentiated set of agent loading indicators, backed by unusually rigorous cross-engine parity testing and a render budget that leaves plenty of headroom.
A SwiftUI library of nine 3D-styled loading indicators in two sizes, designed for AI and agent interfaces.
A native GPUI client that renders Herdr daemon surfaces directly, with documented frame-time benchmarks to back the architectural tradeoff.
A native macOS client built with Rust and GPUI for viewing terminal sessions, workspaces, Git worktrees, and agent activity via a local Herdr daemon.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A logically sequenced AI engineering curriculum that earns its depth on inference and agents, but every link leads back to the author's own paid school.
Structured AI engineer learning path with a dedicated blog post for each step, from fundamentals to deployment.