🤖 AI HOT Daily · Sep 15, 2026

RedNote's AllSpark team open-sources Search Agent model Iris — 35B and 397B versions leading their weight class, weights and eval code public; SiliconFlow hosts open-source Hy4 preview (770B parameters, 49B active, 1M context, Apache 2.0) aimed at coding, analysis, and research, pluggable into Claude Code, Codex, and Cursor; Apple launches next-gen Apple Intelligence as a rebuilt Siri AI debuts in English beta with the 2027 system update — personal context, on-screen awareness, system-level app control, cross-device conversation — expanding to French, Japanese, Korean, Portuguese, and Spanish next month; Anthropic eyes a Nasdaq listing, pitching a second straight profitable quarter toward a ~$2T valuation, though figures use adjusted metrics excluding stock comp, over-80% gross margin, and omit Amazon rev-share and training costs; DeepSeek-V4.1-Flash (Max) rises to No.3 on Agent Arena's open models (+4.87%, median $0.07/task, 68% cheaper than No.2 Hy4 preview while trailing it by just 0.09 points) — reshaping the Pareto frontier; Altman, Amodei, Hassabis, and Musk loosely agree over the weekend to slow AI — third-party audits, domestic-lab regulation, a global slowdown pact — while critics call it a cartel to suppress rivals and open source; a benchmark pits GPT-5.6 Luna against GPT-6 Astra on code review — Luna finds 69 verified bugs for $0.20 at 74% precision, Astra 92 for $5.66 at 96%; VC blogger Tomer Tunguz unpacks five factions (interpretability, labor, economics, geopolitics, regulatory capture) behind Amodei's slowdown call and notes none offers a concrete speed; and Anthropic rebuilt its test-impact-analysis service as agentic coding strains CI — engineers now ship 8x the 2021-25 code baseline, 80% written by Claude, driving a 25x CI-job surge in six months.

1. 1. RedNote Open-Sources Its Iris Search Agent

   AllSpark ships Iris — both the 35B and 397B sizes lead their class; weights and eval code are out, with data and recipes to follow.

2. 2. Hy4 Preview Hits SiliconFlow: 770B Params, 1M Context

   770B total / 49B active per token under Apache 2.0, built for coding, analysis, and research — wire it into Claude Code, Codex, or Cursor.

3. 3. Next-Gen Apple Intelligence: Rebuilt Siri AI Arrives

   Personal context, on-screen awareness, system-level app actions, cross-device conversation — the English beta rides the 2027 update, expanding to five more languages next month.

4. 4. Anthropic Aims for Nasdaq, Targets ~$2T Valuation

   Pitched on a second straight profitable quarter — though on adjusted metrics excluding stock comp; gross margin tops 80% excluding Amazon rev-share and training costs.

5. 5. DeepSeek-V4.1-Flash Is No.3 Open Model on Agent Arena

   Up +4.87% at a median $0.07/task — 68% cheaper than Hy4 preview yet just 0.09 points behind; No.12 overall and reshaping the Pareto front.

6. 6. Big Tech's AI Slowdown: Safety Pact or Cartel?

   Four CEOs loosely agree to press pause with third-party audits, domestic-lab oversight, and a global pact — critics dismiss it as a cartel squeezing rivals and open source.

7. 7. Code Review on a Budget: Luna ($0.20) vs Astra ($5.66)

   Across 50 public PRs with identical prompts, Luna nails 69 verified bugs at 74% precision for $0.20; Astra finds 92 at 96% for $5.66.

8. 8. Tunguz Maps Five Factions Behind the Slowdown Call

   He lines up interpretability, labor, economic, geopolitical, and regulatory-capture camps — and notes none of them names a concrete pace.

9. 9. Anthropic Scales TIA as CI Workloads Explode 25x

   Quarterly code output is 8x the 2021-25 baseline with 80% written by Claude; CI jobs balloon 25x in six months, forcing a TIA service rebuild.