Daily Digest

Ethan Mollick's shader tests place it below Kimi K3; OrangeAI's take: "a high-value K3" that powers his memory-heavy Cola assistant well (274 + 67 likes)

community framing (per OrangeAI): Hailuo's H3 is "~80% of Seedance 2.0" and SOTA among open video-generation models — a bigger deal than yesterday's bare release suggested (61 likes)

Anthropic's Thariq points out that connecting Gmail/Calendar/Slack to Claude also gives Claude Code access to them — including from inside Artifacts (814 likes)

Mollick told Codex to build an otter-in-animal-mech-suits game; it took over the machine and produced game, assets and all — his point: agents now span entire creative toolchains (51 likes)

write the spec as a document, /compact, hand off to a cheaper executor, then have the planner verify — his recipe for reliable-yet-affordable complex tasks (367 likes)

Nate Silver notices Claude "gets more stubborn as the context window grows"; Mollick confirms the effect is real but warns models have poor self-knowledge — his fix: compact to a markdown summary and start fresh (390 + 199 likes)

a local agent doing authorized network scans from a phone; HN debates on-device agents for security work (HN 104 pts, 30 comments)

don't paste, retype: the thread splits between "a new form of prayer" mockery and veterans saying it's just the old "type out the examples" study habit (HN 398 pts, 337 comments)

layer-by-layer loading trades speed for the ability to run big models on tiny hardware (HN 194 pts, 75 comments)

the engineering post behind the day's headline: a rebuilt client-to-model audio architecture that keeps sound flowing during reasoning and tool use

his math: lab revenue 10×es yearly while compute only 3×es, so margins, compute prices, and the inference share must all rise; cites Anthropic inference margins going from ~40% to ~80% in a year (Aug 3)

fresh off a $13B round, Baseten on the "inference inflection", plus Ali's viral 48-hour teardown of the Kimi K3 modeling code (Aug 3)

quick news roundup: OpenAI's influencer retreat, Alibaba's new model, Visa acquiring BioCatch, Reddit's AI concerns, and Apple's rumored Siri subscription (Aug 3)

Sequoia hosts OpenAI's Core Automation leads on automating AI research itself (Jul 29)

extends world models beyond physics to hidden mental state: what a person believes, wants, intends, and feels — the substrate human behavior actually runs on (55↑)

breaks OPD's core assumption that the teacher must be stronger: shows how weaker teachers can still improve frontier students (41↑)

the first large-scale world model trained to predict future *touch* as well as future vision, for contact-rich robot manipulation (28↑)

fuses RLVR's sparse rewards with OPD's dense teacher signal to get past the teacher-quality ceiling (24↑)

benchmarks whether coding assistants learn *your* recurring ambiguities across sessions instead of re-asking every time (21↑)

finds that simply having constitution-like content present during midtraining drives alignment improvements (3↑)