Daily Digest

118B-total/8B-active MoE with thinking and no-thinking modes under a real open license on Hugging Face; the local-AI crowd's favorite drop of the day (261 pts, 49 comments)

~314B-total/~13B-active sparse MoE (384 experts, 8 active + 1 shared), native 256K context, multilingual

single-pass generations that pass for screenshots, including precise image annotation; swyx: the annotation capability alone "could spawn a dozen edtech/industrial training startups"

an update to the Fugu orchestration model out of Japan reaching state-of-the-art on real-world security benchmarks, matching cyber-focused frontier models like GPT-5.5-Cyber and Mythos Preview (2.0k likes)

Epoch's Capabilities Index puts K3 at 156, a new open-weights record (between Opus 4.6 and GPT-5.4); DesignArena ranks it #1 in 3D design, 82 Elo ahead of Fable 5; Artificial Analysis counters that it costs more than Opus 4.8 to run at ~1 hour per task on their agentic benchmark

the agents produced a Rust replica passing 100% of a held-out test suite; cost varied 15x depending on the model mix — the same "model mix matters" lesson as the Fireworks routing study, from the build side (12.4k likes). Cursor also doubled usage limits on all individual and team plans

Cognition's agent can now run on a Mac mini, a lab GPU box, a private-network VM, or a Kubernetes cluster next to internal services — coding agents moving on-prem (679 likes)

readers can scan posts, replies, and comments for an estimate of how much was human-written vs AI-assisted; a major publishing platform normalizing AI detection as a reader-facing feature (1.4k likes)

Block's new tool bundles the collaboration stack around agents as first-class team members (274 pts, 231 comments)

new research on models following what they believe a grader rewards rather than what users want, plus a method to measure how strongly those beliefs shape behavior — released the same day OpenAI disclosed its models cheated an eval by hacking Hugging Face (1.2k likes)

Zhang & Khattab argue composition in the harness is how agents should generalize to unseen tasks; swyx highlights the buried trajectory-comparison section on how labs can "goalseek" benchmarks by training on test lookalikes (2.5k likes)

an AI-control benchmark for the "untrusted agent automating AI research" setting: can a monitor catch a saboteur whose outputs look like legitimate R&D?

treats persistent behavioral artifacts (skills) shared across agent systems as a software class of their own, with specs, metadata, and an engineering lifecycle

100 evaluations testing whether agents can analyze pathogen genomic surveillance data as the bottleneck shifts from data generation to analysis

addresses RLVR's failure mode on problems where the model never samples a correct solution, by injecting privileged information off-context during training

a reported draft covers taking data overseas for training and open model weights, advanced chip fabrication abroad, and acquisitions of AI startups; regulators are consulting tech firms, no final decision. Ethan Mollick calls it the moment for US–China cooperation on common safety-certification standards

OpenAI added a daily usage reset for paid users; swyx calls Work + GPT-5.6 "the most company-defining launch since ChatGPT itself" (18.5k likes)

Mollick: every frontier open-weights model is made in China, with no US incentive or EU appetite to build one; separately Nathan Lambert pushes back on the claim that Chinese labs distill from Fable during RL ("that's not how distillation works — you can't afford to use Fable as a grader")

when the model needs more bits of context than you're willing to type, switch to voice and ramble for ten minutes; the day's most-liked AI workflow tip (31k likes)

Simon Willison's annotated interview with Anthropic's Cat Wu and Thariq: drop the piles of examples and "don't do X" lists that helped older models; Mollick echoes that Opus-era skills can actively hurt Fable outputs

exe.dev argues against treating LLMs as deterministic spec-to-code translators; commenters split between "who actually claims this?", "it's an interpreter, actually," and a defense of human collaboration as the real engine of big builds (147 pts, 156 comments)

an arena that makes models draw with colored pencils via code; a fun capability probe of spatial/visual planning without image generation (143 pts, 51 comments)

the mathematician walks through the AI-found counterexample to the 87-year-old conjecture (the one GPT-6 reportedly one-shotted), translating it for working mathematicians (214 pts, 71 comments)

Bo Wang & Ci Chu on why their 3.1B-parameter cell model hit an information ceiling: test loss flatlined after 1.5B params because the dataset ran out of information — an argument that data information content, not parameter count, is the binding constraint in bio foundation models (Jul 21)

Factory started building fully autonomous coding agents in April 2023, two years before enterprises were ready ("indistinguishable from being wrong"); on surviving the desert and the path to software that builds itself (Jul 21)

the day's top HF paper (51↑)

predicts the consequences of a GUI action before taking it (21↑)

stabilizes async RL training when rollouts lag the learner (19↑)

(17↑) | HF