Daily Digest

simonw uses his SVG-pelican test to visually track Spark → Spark 1.1 → Spark 1.2 improvements across one model family over four months (36 likes)

four years to the day after prompting GPT-3 + DALL·E for imaginary game concepts, he had Fable build the actual game from those old images as the spec (156 likes)

Neon/Castform's benchmark argues small purpose-built models win on retrieval price-and-efficiency; HN's takeaway: harnesses should offload narrow tasks to targeted cheap models (HN 233 pts, 57 comments)

OrangeAI and collaborators launch a hand-picked, human-tested skills marketplace, arguing skills are "better refined than numerous" (one user with 10,000 skills made their agent *worse*) (87 + 59 likes)

pushes agents toward tasks that span a full "day" of work rather than bounded prompts — part of a visible surge in long-horizon harness papers (12↑)

finds models invent user traits that aren't supported by the conversation, and that asking them to self-monitor makes it *worse*, not better — a caution for every "it learns you" product (18↑)

proposes a metric for how well a model composes reusable skills over long-horizon reasoning; connects to the week's running "skills" debate (8↑)

the argument that a hobby is the *process*, not the output, so tools that skip to the result miss the point; spawns a thread on "Slow Code" meetups and craft-for-its-own-sake (HN 153 pts, 152 comments)

a memory scheme that avoids spending context tokens on read/write bookkeeping (HN 93 pts, 12 comments)

OpenAI reported two additional accidental-cyberattack incidents (from AISI and Irregular); simonw notes he's now had to make an "accidental-cyberattacks" tag with four entries (91 likes)

argues any long task inherently requires taste, judgment, and creativity, which capable agents now exercise; separately flags a "weird contradiction": models follow complex instructions better yet increasingly treat instructions as *suggestions* — big implications for skills (374 + 194 likes)

an MIT/Stanford paper finds most people would be financially better off following LLM financial advice (GPT-5.2 & Gemini 3 Flash); another finds AI conversations can *increase* loneliness, contradicting earlier work — effect depends on the chatbot's approach (114 + 187 likes)

AI coding has years of engineering QA and real-user feedback to catch slop; AI knowledge-work has no such quality bar yet, so it risks being slop end-to-end (31 likes)

on the reported AMD/Anthropic compute relationship and the widening GPU-supply picture (Aug 5)

a reconstruction of how ChatGPT Work (the Codex harness extended to cloud/knowledge work) does Memory, Proactivity, and more — it crossed 10M users three weeks after its July 9 launch (170 + 48 likes)

studies how knowledge flows between modalities, when they synergize, and how early "unification" happens during multimodal pretraining, with practical recipes (11↑)

lets a unified model call tools mid-generation to produce images agentically rather than in one shot (13↑)

another OPD variant, this one hardened against teachers that pass on spurious correlations — the week's on-policy-distillation streak continues (9↑)