Links: October 4-17

I promised to post every week, but personal and work commitments kept me from posting last week. I’m making up for it with a huge tranche of links I came across over the last couple of weeks. I’ve grouped them by topic so they’re easier to browse.

Agents learning from experience

• Agent Plasticity - Tests whether an agent carries useful lessons into fresh runs, and what those lessons cost.
• RLTL;DR (Kirchhof) - Turns failed attempts into short lessons, then trains the model on those lessons.
• Harness Learning - Trains a model to improve the code around another model; adaptation happens in the harness at test time.
• Memento 3 - An agent revises a rulebook and turns it into executable code while its language model stays frozen.
• SPADE - One model builds executable training environments and learns to solve them.
• AutoEnvScaling (Yu) - A proposer studies the solver’s attempts and builds new training environments around what it struggles with.
• Terminal-RSI - Combines rewriting past attempts with further policy training for terminal agents.

Context, search, and coordination

• Fork-and-Flush (Cai) - Keeps the workspace, clears the conversation, and branches the search into fresh attempts.
• Scaling Discovery Requires Forgetting - Restarts from the best candidate while dropping search history that may be holding the agents back.
• An Empirical Study of Harness Design for Coding Agents - Tests how planning, tools, and context management affect coding agents across different models and budgets.
• HERMES and its modular Dev-Primitives - Gives repository components their own agents, with execution failures guiding which components need work.
• Alex Zhang’s compositional-generalization essay - Explains how a harness can break an unfamiliar task into smaller tasks the model already handles.
• Gemini-agent Minecraft run - Four agents beat Minecraft after more than a day of coding and recovery in one Easy-difficulty run.
• HERMES and Dev-Primitives - A discussion page for the same HERMES paper, with its modular approach to repository work.
• Harness-design study PDF - The full harness-design paper, including the experiments behind the planning and context-management comparisons.
• Forgetting paper commentary - Independent commentary on the forgetting paper; the original paper is linked above.
• DAIR.AI on Harness^2 (Harness^2 paper) - Harness^2 revises the current harness, then carries lessons from those edits into future tasks.

Self-report and interpretability

• Can LLMs Introspect? A Reality Check - Asks whether models report their internal state or simply pick up clues in the prompt.
• Identifying Introspection From the Inside (Atkinson) (Bau) - Studies whether learned preferences and reports of those preferences share internal mechanisms in a controlled experiment.
• Verbalized and Internal Probabilities Are Coupled (Williamson) - Tests how a model’s stated confidence relates to the probabilities it assigns to answers.
• Training Agents to Self-Report Misbehavior (Lee) - Trains agents to report their own misbehavior and studies how that works alongside external monitoring.
• Grafting model beliefs across checkpoints (Feng) - Learns a weight update on a base model, then transfers it to a post-trained checkpoint.
• Neural Networks Are Already Clean (Cammarata) - Argues that native neurons and circuits may be easier to interpret than we assumed; presented as a working hypothesis.
• Complete or Sparse - Works out conditions under which a learned representation can recover the variables that generated the data.

Models and training

• Base Models Can Reason By Taking a Cue From Training Data - Studies how the first few response tokens can trigger reasoning patterns learned during pretraining.
• The Implicit Curriculum Hypothesis - Tracks when models learn individual skills and when they learn to combine them on designed tasks.
• LLMs Learn Better In-Context from Rules than from Examples (Fu) - Compares rules with examples across five learning tasks; rules generally work better in those tests.
• Queen (Bhaskar) - Connects a chess expert to a language model so it can explain moves as well as play them.
• Loop, Think, & Generalize (Yao) - Studies whether repeatedly using the same transformer layers helps models combine knowledge across more reasoning steps.
• LOOM (Liu) - Addresses unstable hidden states and repeated expert choices when looping mixture-of-experts models.
• Dust (Samip) - Trains through activation perturbations instead of backpropagation, though the compute cost still needs substantial work.
• Lilian Weng’s Scaling Laws, Carefully - Shows how training choices and measurement details can change what a scaling curve seems to tell us.
• Kolibri (Aleph Alpha) - An English-German model released with open weights under Apache 2.0.
• Mistral Large 4 (Mistral) - Mistral’s new large model, available as an API preview with weights promised later in October.
• Sasha Malysheva on latent communication (Full-bandwidth transformer, Latent Planning) - Two papers on latent computation: feeding hidden states back into a model and finding plans before they appear in text.

Systems and tools

• KV-streams (Penaloza) - Carries the KV cache through context compaction instead of rebuilding it each time.
• RaReCache - Reuses a smaller model’s KV cache in a larger model, recomputing selected positions to correct transfer errors.
• Processing in Memory primer - An older primer on moving computation closer to memory to reduce the cost of moving data.
• Modal’s serving case study - A serving case study covering parallelism, speculative decoding, cache tiers, and routing for coding-agent workloads.
• ScaleDown - Tools for finding AI calls that could move to task-specific small models and testing the migration.
• Simon Willison’s Decisions plugin - A command-line plugin for asking models constrained questions: yes/no, choices, and scores.
• AgentRun (Rios Berrios) - Mixes code, typed decisions, and agent calls in workflows you can inspect and rerun.
• REA - Gives agents tools to inspect application behavior and reverse-engineer native binaries.
• Danielle Morrill’s credential setup - A practical account of scoped credential access for agents, including where that protection stops.
• The people holding up the internet - Shows how much widely used software depends on the work of a small number of maintainers.
• Processing in Memory PDF - The full processing-in-memory survey, covering the hardware and software needed to make it useful.
• ScaleDown homepage - The product site for the task-specific small models used by the migration tools.
• Internet maintainers - The original shared URL for the same visualization of open-source maintenance work.

Research and evaluation

• Priced Guidance (Wen) - Measures how much help a model needs to reconstruct a research idea by charging for each hint.
• Epiplexity (Fein) - Defines useful information in terms of what an observer with limited compute can learn.
• Kernaut (Suwandi) - Lets coding agents search for new kernel programs while construction rules enforce mathematical validity.
• LLM-as-a-Verifier - Uses finer scores and repeated evaluation to give agents feedback; these scores are judgments, not proofs.
• Crosby’s evaluation account - Combines deterministic document checks with lawyer comparisons to evaluate work on actual contracts.
• John Langford’s Research at AI speed - John Langford on the pipelines, experiment capacity, and review needed to do research at AI speed.
• Jo Jiao’s MATS and Redwood lessons - Practical research advice: make predictions, run experiments that test them, and update your understanding.
• Learning what to investigate next (MIRA paper) - Separates choosing the next research question from executing it, so learning can target those decisions.

Learning and working

• Oguzhan Olguncu’s learning essay - Keep doing the work that teaches you; let the model handle the surrounding setup.
• Felix Rieseberg’s ambition essay - Push beyond doing the same tasks faster, with better tools and checks outside the model.
• Seeing Theory - An interactive introduction to probability and statistics, now archived for reference.
• Scott Aaronson’s AI Alignment Theory course - A course on AI alignment theory that leaves room for disagreement about its foundations.
• Sonia Baschez on household assistance - Asks for an assistant that remembers and coordinates household work without waiting for every instruction.
• Baschez interview - An interview expanding on Baschez’s request for help with the household mental load.
• Ethan Mollick on class attendance (study) - A 2023 post on a 2010 meta-analysis linking class attendance with grades; correlation is not causation.

AI and mathematics

• Raghu Meka on barriers of perception - Raghu Meka on how beliefs about what is possible can stop us from attempting worthwhile problems.
• The collected mathematicians’ reactions - Collects mathematicians’ reactions to announced AI results, including questions about proofs, credit, and training.
• Bill Gasarch’s reflections - Bill Gasarch discusses surprising announced results and the gap between formalization and human understanding.
• Derek Thompson’s essay - On mathematical work, identity, and the grief that can accompany automation; the full interview is paywalled.
• Gautam Kamath on 3SUM and shortest paths (paper) - Points to a paper claiming faster algorithms for 3SUM and all-pairs shortest paths.

Economics, institutions, and risk

• Some Simple Economics of AGI - Models an economy where execution gets cheap but verification and responsibility remain scarce.
• The Acemoglu-Gans exchange - A debate about how AI affects growth, with worker complementarity and new tasks central to the disagreement.
• Michael Jordan’s 2018 essay - Michael Jordan’s 2018 argument for building AI infrastructure around people and institutions.
• Michael Jordan interview and preface - Jordan on AI and society; read his preface explaining what the headline misses.
• Timothy Gowers’s reflection on extinction risk - Gowers explains his growing concern about AI extinction risk without treating catastrophe as inevitable.
• Acemoglu-Gans exchange - The original shared URL for the same exchange on AI, workers, and economic growth.
• AGI economics PDF - The PDF of the same economics paper, including its model of verification and apprenticeship.
• The Mathematics of the State - Frames government decisions as optimization problems, while making the political choices in their objectives explicit.
• Anna Stansbury on truck safety - Asks whether electronic logging requirements actually made trucking safer, rather than just easier to monitor.
• An Anatomy of Asset Returns (paper) - A paper separating gradual price movements from jumps to study patterns hidden in aggregate returns.

Biology and neural circuits

• Symmetries and continuous attractors in disordered neural circuits - Models built from head-direction cell data show how apparently disordered circuits can still have structured dynamics.
• How biological synapses self-assemble gradient learning (paper abstract) - A model of how local synaptic rules could assemble circuits that learn.
• A re-look at The Selfish Gene after 50 years - Vidyanand Nanjundiah revisits The Selfish Gene, questioning its account of cooperation and its gene-centered framing.