George Cameron breaks down an agent benchmark's $21 task cost, its over-$200 uncached estimate, and how to choose models against your own workload.

Two migration campaigns opened 2,333 requests and landed 729. Stronger checks helped, but owner review and deployment remained real work.

A complaint about an apparently automated support reply gets a warmer apology. Whether anyone fixed the bug remains unanswered.

AI-assisted discovery challenges a profession built around individual achievement. Three mathematicians argue for the joy of understanding; readers ask who sustains the people doing it.

Reading raw network parameters and writing Python to improve a model use different capabilities. Honnibal separates the experiments behind the headline.

Simile's Joon Sung Park explains interview-grounded behavioral simulation, its validation problem, and the gap between customer panels and predicting society.

Source control for parallel coding OpenAI explains how it rebuilt Git hosting for parallel coding — Its engineering account shows how shared downloads and independent branch updates can reduce contention when many coding assistants work …

IBM's Jeff Crume explains data visibility, classification, jurisdiction, access controls, and continuous audit evidence for AI systems.

Neo4j's Zachary Blumenfeld turns skills and decision traces into executable graphs, with explicit routing and inspectable loan-approval runs.

Factory's Nikhil Harithas explains customer-owned agent compute, outbound relays, and the lifecycle machinery behind portable coding agents.

Aamir Shakir explains multimodal retrieval, demonstrates search over Intel reports, and reports a customer accuracy gain with important limits.

A review finds faster single-request generation with four-bit weights, little gain in one large-batch test, and sparse matched accuracy evidence.

Two grading models can agree on the same mistake. Cross-provider agreement filtered one unreliable grader while discarding most comparisons.

Cristóbal Valenzuela demonstrates Continuum, a research prototype for interactive visual explanations and shared visual work with coding agents.

Editing a model to stop reasoning loops Researchers use one AI model to edit another’s weights — The experiment shows a possible way to fix a specific failure in a small model without retraining it. MIT researchers asked GPT-6 Astra to …

MIT researchers report targeted weight edits that let a small Qwen model finish more answers, plus experiments reading and writing game networks.

A dashboard for human decisions, agent run histories, workflow changes and regression tests, backed by Archon and Postgres.

Will Brown makes the case for turning agent traces and human decisions into tests and training data for organization-specific models.

Rewriting earlier reasoning often shortens later reasoning, but the compressor can erase time savings and task quality remains uncertain.

Across 120 historical distributed-system bugs, coding assistants find relevant files more reliably than the precise code that caused the failure.

As AI makes producing proofs easier, a mathematician argues for rewarding understanding, good questions and the work that connects results.

Figma's Ezra Mechaber shows a shared product-management skills repo, agent-assisted testing and an experiment workflow.

Finding and fixing software bugs A PostgreSQL maintainer says AI-found bugs are slowing the next release — Finding defects faster can move the bottleneck to human review and repair. Major contributor Robert Haas expects PostgreSQL 19 to …

IBM's Legare Kerrison walks through six inputs that help a coding agent work within an existing repository.

Speakeasy's Sagar Batchu lays out identity, policy and audit controls for agents that can act on production systems.

Sourcegraph's Andrew Norrish connects review pressure with incident response and outlines three code-intelligence workflows.

Adam Nahari argues that AI projects need a measurable business problem, workflow redesign and committed users to produce financial impact.

Zack Proser shows how he turned writing experience into a tool for colleagues, and applies the same idea to bedtime stories for his children.

A packaged local AI stack A new local AI stack switches between chat and image models — A small container-based project offers a way to try both on one machine without keeping both models in graphics memory. Solène Rapenne’s project …

Nick Nisi turns code changes into simulated smells and proposes a risk-based review pipeline.

Separate working copies keep code intact, but task-specific history and progress records help coding assistants resume interrupted work.

Maddy Zhang argues that verification, judgment, agent workflows and communication matter as coding agents change engineering work.

Jaclyn Konzelmann discusses short strategy documents, user co-creation, experiments, and designing for an agent-filled world.

Subscription limits OpenAI plans to combine ChatGPT and Codex usage allowances — People using both products will need to budget their work against one allowance. Starting in December, the shared allowance applies to Plus, Pro and Business …

Avery Blanchard explains the three compilation stages, graph breaks, and how to measure whether they help.

Compound's Yijia Liang explains shared agent inboxes, Office-compatible engines, and checking loops for financial deal teams.

Ramakrishnan Lokanathan on testing complete AI workflows, retaining regression checks, and turning real failures into evaluations.

Oracle's Jeff Smith explains vetted reports, schema context and database-level permissions as ways to limit the risks of connecting agents to company data.

A coding assistant finishes a peaceful factory game A coding assistant completes a peaceful Factorio run — A recorded game run shows both how a coding assistant plans a long task and where its automation breaks down. In Ryan Madden’s …

Barrow explains how Perplexity Decider scores classification choices in one model pass, and why fast scores still need careful validation.

Oracle's Richmond Alake proposes an intelligence plane for agent knowledge, memory, data access and learning.

Cole Medin examines a study of AI-influenced expert decisions and suggests ways to keep judgment in the planning process.

A web measurement study finds that successful status codes can hide empty shells, bot checks and substitute pages, distorting citation checks.

Site owners get a proposed shared training preference AI companies plan one shared way to decline training use — Website operators may eventually express one preference to participating providers instead of maintaining separate instructions …

Bumble's Supriya Subudhi shares five lessons from testing AI photo feedback: adoption, bad outputs, costs, model choice and regional rollout.

imec benchmarks coding agents across models, harnesses and hardware, finding that utilization and concurrency can change the economics of self-hosting.

Upstart and Coval describe human review, AI judges and voice simulations for a regulated loan-verification agent.

Oracle's Wojciech Pluta demonstrates a memory-promotion gate that retains an operator's preferences while rejecting an unsafe blanket lending rule.

Matt Wolfe reviews Anthropic's proposed IPO, model releases, a hands-on Genie 4 demo, and an AI-generated complexity-theory claim.

Cursor expands the list of malicious extension versions Cursor adds an earlier Claude Code version to its warning — Review installation history for both affected versions, because updating the extension can leave the attacker’s separate …