Models & Releases 4
first from-scratch open-weights family from Mira Murati's lab: 975B MoE / 41B active, 1M context, text+image+audio reasoning, controllable thinking effort; deliberately positioned as a customization base on Tinker, not a frontier-benchmark chaser
"RoPE is dead, long live GRAPE": query-gated attention that drops RoPE, attention sinks, and gated attention in one move
full agent harness, fullscreen mouse-interactive TUI, extensible; simonw found gems like a self-contained Unicode box-art Mermaid renderer
keeps a voice conversation going while checking flights, weather, and building an itinerary in parallel (1.1k likes)
Safety & Security 2
OpenAI says common robustness evals are saturated, so it trained an attacker model whose prompt injections crack its own previous models, then used it to harden GPT-5.6
Anthropic ran high-stakes simulations across models from six labs (incl. GPT-5.5, Gemini 3.1 Pro, Grok 4.3, DeepSeek V4, Kimi K2.6); cites the real-world MJ Rathbun/OpenClaw hit-piece incident as a warning sign; all transcripts public
Tools & Demos 5
GPT-5.6 & Ultra, parallel work with /goal, faster computer use, AppShots, inline edits, Sites, mobile/SSH workflows, full PR flow (2.3k likes)
OpenAI × Work Louder hardware controller with mappable buttons/joystick for agent workflows; OpenAI's merch-drop era continues (7.3k likes)
dotey's app packages transcription, speaker ID, polish, translation and simple cutting behind a CLI + Skill so Claude Code/Codex can drive it, with a GUI for human preview/edit
Markdown files now open in Notion as read-only previews and import as pages (2.5k likes)
the agent-CLI aesthetic is now a UI kit (HN 99 pts)
KOL Takes 4
Claude Code's creator on automation as the highest-leverage engineering activity (6.1k likes)
plus "software engineering is the profession of automation" (1.6k likes)
flips the common anti-agent argument on its head (1.4k likes)
you need taste (or hire it) and build your own skills/instructions (572 likes)
HN Threads 2
maps agent harness concepts onto Unix primitives; sparks debate on generic vs domain-specific harnesses (173 pts, 87 comments)
extreme budget local inference (240 pts, 158 comments)
Lab Updates 3
"Conjecture machines": AI agents and the new validation bottleneck in science — essay arguing real-world testing, not hypothesis generation, is now the constraint; four priorities for policymakers (606 likes) | essay | @GoogleDeepMind
First-of-its-kind research partnership with film studio A24 | blog
Industry 3
open-source agentic engineering platform went 0→3M developers in 16 months; swyx calls the growth "one of the greatest mysteries" (783 likes)
the standalone AI enterprise-services firm from Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs; no official Anthropic announcement yet (513 likes)
per Reuters sources; HN's biggest fintech thread of the day (366 pts, 214 comments)
Podcasts 2
Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden (Jul 14) | episode
Adam Brown — a deep but accessible introduction to general relativity (Jul 10) | episode
Research & Papers 6
Jim Fan's team scales a robot policy to 8,000 timesteps (~5 min) of context at constant inference cost — 3 orders of magnitude past SOTA — by running a gradient step on a tiny inner model per sensor reading (807 likes)
RL with verifiable rewards, no human-annotated data, at trillion-param scale for emergent reasoning (HF 13↑)
real documents mix length, layout, modality, difficulty; this controls each factor (HF 56↑)
SpectraReward is training-free: measure how well an MLLM can read the prompt back from the generated image (HF 44↑)
evolving the knowledge boundary in agentic visual generation: generators confidently fabricate what they don't know; agents fetch the missing world knowledge (HF 32↑)
no RLHF/DPO preference tuning at all: RLVR + multi-teacher on-policy distillation (176 likes)
arxiv: AI Agents 2
user-permission systems for AI agents, from interfaces to enforcement, against prompt-injection and hallucination-driven leaks (Jul 15)
making evolving agent harnesses readable, navigable, and editable; the harness, not just the model, defines agent capability (Jul 14)