Models & Releases 4
first open model at 2.8T params: Kimi Delta Attention + Attention Residuals, 1M context, max-thinking-effort default; API at $3/$15 per 1M tokens (~$0.94/task — half of Opus 4.8, on par with GPT-5.6 Sol); Artificial Analysis long-horizon Elo +732 over K2.6, behind only Fable 5; weights due July 27 with tech report
#1 Frontend Code Arena (1679 pts, beat Fable 5 in 6 of 7 domains) and #1 on an editorial-writing Elo board at ~$0.25/script; oran_ge: "the gap between Chinese open models and the world's frontier may be down to under 6 months"
emollick finds it "pretty rough" in tests, failing the Lem Test that every frontier model passes since DeepSeek r1; meanwhile it's already runnable in Claude Code via `hf claude`
Google folds the beloved research tool into the Gemini brand; HN mostly groans about Google's rename-then-kill pattern (251 pts, 129 comments)
Tools & Demos 4
coding/research/document agent that runs fully local, via LM Link, or on LM Studio Secure Cloud; ships offline voice input (Mistral's Voxtral) and a Zero Data Retention pledge; HN debates "just another harness?" vs local-first ergonomics (166 pts, 64 comments)
after rewriting SQLite, Turso declares itself "the LLVM of databases": one modern core, many database frontends; Postgres compatibility is next, embedded-in-browser and self-updating materialized views on the roadmap (415 combined engagement, simonw approves)
give Claude access to credentials without pasting secrets into context; agent-credential management going mainstream (HN 25 pts)
upload/create/edit Skills, local Skill folders; Google chasing the Claude Skills pattern; dotey meanwhile blasts Gemini's shipping pace vs Codex
KOL Takes 5
Claude Code's creator: every org has one engineer 10x-ing their output while the rest haven't caught up; publishes a stage-by-stage adoption playbook (1.5k likes)
the easiest way to make a model loved online is pretty websites and 3js worlds, far more shareable than backend or analysis quality (449 likes)
misapplied statistics in complex academic-audit tasks, with GPT-5.6 Pro catching the errors; benchmark hype ≠ reliability (259 likes)
emollick: K3 has no model card yet, open models are easy to jailbreak, and nobody knows if US/UK-style frontier vetting can apply to weights you can't recall (95 likes)
a calm, measured take on where LLM patches fit in kernel work; HN notes it's "cogent and relaxed," a contrast to the usual outburst genre (39 pts)
HN Threads 4
the critics' arguments hold AND the productivity is real; thread converges on skill-atrophy: "you cannot have only higher-level thoughts without doing lower-level work" (193 pts, 202 comments)
end-to-end AI-directed music videos; HN's verdict: hilariously literal ("Michelle Pfeiffer, that white gold" → blonde woman in gold top), the absurd literalism epitomizes AI slop (146 pts, 90 comments)
the 1996 IRC-client-as-comic-strip cult classic (of jerkcity fame) gets an official open-source release (556 pts, 122 comments)
an AI-assisted proof of a longstanding optimization conjecture, written up by the researcher (23 pts)
Lab Updates 2
Our approach to bioresilience — joint framework with Isomorphic Labs for deploying frontier AI as proactive defense against future outbreaks (405 likes) | blog | @GoogleDeepMind
Industry 2
NVIDIA's open-model stack (incl. Nemotron) joins Sakana Fugu, the multi-agent orchestrator that dynamically combines multiple models behind one API; collective intelligence over monolithic scaling (1.5k likes)
The Information says Google refreshed Gemini's training data late last month to fix lagging coding skills, but results were "disappointing"; Bloomberg confirms the launch slip
Podcasts 2
AWS and the Future of AI Investment — AWS's $1B AI investment, plus takes on Inkling and GPT-Red (Jul 17) | episode
The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez-Bombarelli of Lila Sciences on AI-run wet labs (Jul 16) | episode
Research & Papers 7
Base/Turbo/Edit/Edit-Turbo family, competitive text-to-image with fast inference (HF 110↑, the day's runaway paper)
frames deployed self-improving agents around "controllable evolution" — adapting from experience with minimal human input (HF 14↑)
register tokens fix high-norm patch-token outliers as diffusion goes pixel-space (HF 16↑)
attacks the fragmented, tightly-coupled agent-eval pipeline problem (HF 9↑)
multi-step tool-calling agents running on the phone itself, not a server (HF 7↑)
shortening CoT via RL length penalties hides the influences steering the answer: an alignment-relevant trade-off (HF 4↑)
distillation alternative to outcome-based RL for long-horizon multi-turn agents (HF 3↑)