Daily Digest

reasoning traces, OpenAI Responses API support, server-side tools, and smarter logging for the tool that talks to hundreds of models (174 likes)

simonw generated video locally on an M5 Pro: ~115GB download, ~45 minutes per clip; open video generation is now genuinely local (185 likes)

the agentic market-analysis system for wealth management moves from pilot to real deployment — a Japanese enterprise-AI milestone (136 + 25 likes)

driver weirdness, game incompatibilities, startup programs that won't start — "hours saved" by pointing Codex/Claude Code at the OS itself (1.1k likes)

LinearUncle's list of production open-source agent harnesses to study; Baoyu's pick: "go deep on pi-mono rather than skimming all of them" — plus a wink at the leaked Claude Code source (355 + 468 likes)

last year it was an AI video mockup; this year the model built the real game, inventing the core mechanic itself: painting the landscape in brushstrokes as seasons evolve (113 likes)

poisoned experiences don't just sit in memory: agents distill them into durable "skills" that keep executing long after the poison is gone — uncomfortably timely alongside today's AISI report (11↑)

benchmarks whether retained preferences, histories, and skills translate into improved future behavior — the foundations of recursive self-improvement, measured (18↑)

Tikal's open-source skill set imposes team conventions on Claude Code and Codex; HN debates whether skills are the new lint config (HN 74 pts, 39 comments)

"Soup" squeezes fine-tuning (not just inference) onto consumer hardware — yesterday it was AirLLM for inference, today training follows (HN 126 pts, 24 comments)

sells simulated markets as training gyms; HN asks whether "teaching LLMs to trade" ends anywhere good (HN 33 pts, 26 comments)

the company's account of two incidents at external evaluators: what happened, containment, and changes to third-party testing (HN 46 pts)

Alibaba merges three product lines into "Qianwen Office", ByteDance's WorkBuddy hits ~1M DAU (ahead of all domestic rivals), Doubao is "too entertainment-coded" to convert, and Feishu is squeezed from every side (123 likes)

OrangeAI documents the markup between DeepSeek's own API and cloud-hosted resellers (85 likes)

Microsoft shipped GPT-4 (in Bing) before OpenAI did and didn't retreat after Sydney; Google had deep research first — a counter to the "incumbents always hesitate" narrative, one day after the LMChat story (597 likes)

knowledge graphs failed three years ago on cost; "good enough" intelligence flipped the economics (21 likes)

Mollick's side-note on the incident: open benchmarks, fast testing, and communication "neither hyped up nor hidden by technical language" (97 likes)

news roundup on the reported Anthropic cloud commitment and Musk's Tesla AI comments (Aug 4)

Sequoia hosts the Chai team on applying the scaling playbook to molecular design (Aug 4)

a 16B autoregressive-diffusion model that edits video as it streams, no future frames needed — live video manipulation, not clip rendering (56↑)

Tencent tackles 3D's data scarcity with one multimodal model across the whole 3D workflow (40↑)

text is generative modeling's last discrete holdout; AURORA autoencodes it into continuous latents and diffuses like images (40↑)

drops agents into a persistent simulated shop where success means staying purposeful over weeks, not finishing a task (40↑)