Morning brief

Friday, July 17

4 people · 0 companies · 5 topics

TL;DR Moonshot AI’s Kimi K3 (2.8T-parameter open MoE) landed July 16 and is the biggest HN story in months (951 pts, 1,600 comments) while Bloomberg labels it a “Kimi moment” as AI stocks fell; Gemini 3.5 Pro missed its third launch deadline.

Models & research

  • model-releases Kimi K3 launched July 16: 2.8T-parameter MoE with 16 of 896 experts active, 1M-token context, two variants (K3 Max for chat/agent, K3 Swarm Max for parallel processing). Benchmarks: 93.5% GPQA Diamond, beats Claude Opus 4.8 and GPT-5.5, trails only Claude Fable 5 and GPT-5.6 Sol. Priced $3/Mtok input, $15/Mtok output; full open weights drop July 27. Bloomberg specifically called it a “Kimi moment” with AI and semiconductor stocks falling on the news. kimi.com · Bloomberg · Fortune · HN

  • Gemini 3.5 Pro missed a third consecutive launch deadline as of July 16. Delays are specifically tied to coding performance gaps vs GPT-5.6; Google is reportedly weighing an interim stopgap release. No model card, pricing page, or API listing has been confirmed. TechTimes · 9to5Google

  • “Detecting LLM-Generated Texts with Classical ML” argues SVM and logistic regression on stylometric features outperform neural detectors for AI-text detection. 153 pts / 103 comments on HN. blog.lyc8503.net · HN

People

  • simonw ran his pelican benchmark on Kimi K3 July 16: cost $0.25, consumed 13,241 reasoning tokens, and found K3 Max showed “genuine understanding of the task.” Also shipped a new Mermaid-to-Unicode-box-art tool on the same day using the Rust renderer from the now-open-sourced grok-build. simonwillison.net · grok-mermaid tool

  • sama posted July 16: “we did not have our best last 12 months ever, which is mostly my fault, but we are about to have our best 12 months to date.” Rare candid admission from Altman amid executive turnover and IPO timing uncertainty. X

  • emollick updated his “harbor towns through history” benchmark July 16-17 to add Kimi K3, Inkling, and GPT-5.6 Pro, then made the interactive simulations publicly playable. X

  • swyx published AINews July 16 naming Inkling and Kimi K3 the dominant stories of the day, calling Inkling “new best American Apache 2.0 open model.” latent.space

Tools & dev

  • claude-code v2.1.212 shipped July 17: /fork now copies the active conversation into a new background session (the previous in-session subagent renamed to /subtask); claude auto-mode reset added; session-wide WebSearch capped at 200 calls (tunable via CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION); subagent spawns also capped at 200 per session to stop runaway delegation chains; MCP tool calls over 2 minutes auto-move to background. Plus a symlink worktree vulnerability fix and Windows Group Policy PowerShell 5.1 fallback. CHANGELOG

  • LM Studio Bionic launched July 17: a built-in agentic layer for open models inside LM Studio, positioning it as a local-first agentic runtime. 151 pts / 62 comments on HN. lmstudio.ai · HN

  • EU ruled it will force Google to share search data and open AI on Android (July 17), continuing antitrust pressure on Google’s AI distribution advantages. ArsTechnica

OSS

  • Nutlope/hallmark is the day’s biggest GitHub gainer (+1,486 stars, 11.6k total): an anti-AI-slop design skill for Claude Code, Cursor, and Codex. Timely given the community’s ongoing debate about agent output quality. GitHub

  • Graphify-Labs/graphify holds Python trending #1 again (+1,476 stars, 89k total): converts code folders, databases, and docs into queryable knowledge graphs for AI coding assistants. GitHub

Other

  • WAIC 2026 opened July 17 in Shanghai (runs July 17-20). Key hardware debut: Huawei Atlas 950 physical AI super node alongside near-memory 3D chips. Key software: MiniMax M3 multimodal model, ZTE Nubia’s “world’s first AI Agent smartphone” (StepFun Agent OS + ByteDance Doubao). On July 16, 29 nations signed the founding agreement for the World Artificial Intelligence Cooperation Organization (WAICO), a rival governance body pitched at the Global South. Globaltimes · China Embassy

  • (Polymarket: 97% yes, ~$6.8M volume) “Which company has best AI model end of July 2026” now puts Anthropic at 97%, up from 96% prior. Kimi K3’s launch did not move the needle against Claude Fable 5 in market confidence. Polymarket

  • MIT released a working paper “Speculative Growth and the AI ‘Bubble’” probing whether current AI investment trajectories are sustainable. 48 pts / 54 comments on HN. economics.mit.edu · HN