All desks

Sean Goedecke treats a deadline as a constraint for choosing a design, then gives managers several approaches with explicit risks. His focus is unfamiliar work in large codebases.

Spend Sunday morning doing farm chores and composting at a one-acre educational farm. New volunteers must register; younger participants need an adult or waiver.

Divye Singh trains a small scoring head over a frozen model, then checks option order, masked loss, and probability calibration on synthetic support tickets.

Artificial Analysis compares bounded decisions across four tests. Jev leads the decision-model composite, while knowledge, reasoning, and calibration expose different limits.

Bryant Francis examines Remedy's approach to signposting: preserve the pleasure of finding a route while giving players consistent environmental cues.

Alex McCarron asks what anti-war cinema loses when soldiers' trauma displaces responsibility for the people they harm.

Sarah Chihaya finds a biography that reconstructs Brontë's everyday world while respecting how little we can know of her.

Brooke Lacey shows Claire Vo how she uses Claude Code, FFmpeg, and Whisper to turn phone footage into short-form videos, learn her delivery, and mine audience comments.

Moshi founder Anekha Sokhal argues that AI should gather and check market information while investors keep the decision. She describes a six-agent system, traceable claims and the hidden cost of losing apprenticeship.

Al Chen of Cisco explains how exploit tests, verified restores and privacy checks help banks move AI into production. The bottleneck also includes image approvals, GPU procurement and internal ownership.

Redact found instructions that withheld public escalation routes from chatbot answers. Whether any bot followed them remains untested.

Agent Step-Up tests whether approval screens show the real action and whether that action can change after approval.

Free East Village bluegrass tonight, Wednesday jazz in Astoria, and ongoing art and museum outings; Onishi reopens today and AMNH closes early.

SWE-Bench Pro V2 separates bad instructions from evaluation exploits. Its strongest systems already approach the repaired test set's ceiling.

SkillClash isolates failures after both skills are loaded. A correct combined example repairs much of the lost rule compliance.

A two-model GEPA and SkillOpt study separates answer correctness, formatting, server failures, waiting time and token use.

DeepSeek opens its model-serving software DeepSeek publishes the software that runs its model — Developers can inspect and adapt DeepInfer to run DeepSeek’s Flash model on supported server hardware. The Apache-2.0 code supports …

A.O. Scott returns to Berger's criticism, where looking at art joins pleasure, memory, and political history without reducing one to another.

Two defenses and a sharp dissent ask whether Digger makes artistic failure meaningful, or simply makes audiences sit through it.

William Davies traces today's literacy panic to decades of treating language as measurable task completion, and argues for the freedom to experiment.

Jeremy Lybarger follows Arch Connelly's queer glitter-and-pearls art from drag theater to a luxury-backed museum survey.

A larger tool menu looked like worse tool use until researchers raised the model's response allowance.

A memory test follows facts as they change, exposing errors that answer scores can hide or misdiagnose.

Momento's Khawaja Shams explains realistic coding tasks, leaked answers, repeated trials, and the traces that make agent scores worth trusting.

CODA tests jointly trained models and skill libraries after their instructions change. Even paraphrases can break previously successful tasks.

A controlled study separates finding a skill, opening its instructions and completing the task. Choosing skills before execution helps, with costs.

Structured reflection preserves richer examples, but incompatible reflection models can leave an optimization run unchanged.

Netflix connects telemetry, deployments and incident discussions through a validated knowledge graph. Its agents propose evidence-backed hypotheses; engineers still decide what happened.

A buggy Breakout clone and some difficult weekend experiments make Chris Kiehl's case for programming without an economic justification.

Peter Dorman reads Silkwood as a story about how economic dependence makes danger harder to acknowledge and union power makes action imaginable.

Molly Judd reviews Christianity's emergence through Roman roads, shared languages, borrowed stories, and familiar images.

A practitioner study separates reported agent reliability from measured repository habits and the work of maintaining specs.

A second model runs checks after completion, then sends failures back while preserving the working agent's state.

Martin Keen explains the goals, triggers, context, tools and human checks around an agent's act–observe–adjust loop.

Maddy Zhang's eight rules for reviewing AI-generated code, preserving error signals, limiting production risk and checking dependencies.

Code Monkey reports $200,000 in estimated gross sales and 92% positive reviews. One successful cooking game leaves the wider market question open.

Benj Edwards makes songs and games for an audience of one. Their private value can be real while the consequences for creative workers remain unresolved.

Costumes, bubbling, hip-hop juggling, crafts, and dance fill the White Garden on Saturday afternoon. Free entry; no registration required.

B. D. McClay finds perceptive characters in Chad Harbach's new novel, but asks where their society's pressures have gone.

Request-level logs make coding agents easier to debug. The HN discussion asks whether that requires putting another service in every request's path.

Encrypted networks expose a limit of capability testing: failing to uncover an ability cannot universally prove it absent from adversarially supplied weights.

Nine NYC outings: free East Village bluegrass tomorrow, Agenda Fest this weekend, uptown exhibitions and self-guided walks.

Arcade compares finding tools on demand with listing 100 tools in every request. Savings depend on search frequency, prompt caching and how the client handles large results.

An external controller learns short repository instructions, hides unused tools and redirects stalled work. Reported savings include its own model requests and prompt-cache discounts.

Validation and a constrained retry recover tool-request errors in controlled Qwen tests. The strongest results depend on resolving a usable tool name, with weaker gains outside that model family.

Munin combines local paper search, full-text reading and citation checks with a rerunnable evaluation suite. Its strongest result permits web access and depends on a private reference corpus.

Agent frameworks & tooling Munin lets research groups question papers on their own servers paper notes → — Its released software includes tests you can rerun after changing the model or paper collection. An agent is a model running a …

A practical path from signatures to evaluation, prompt optimization, and a saved program you can compare honestly.

Define success, measure a baseline, and reserve unseen examples before asking DSPy to improve a program.

Use reflective prompt optimization after the baseline, with a real budget and an untouched final test set.